Category: Industry Analysis

Fifty-five percent of surveyed executives say a lack of business context is the main reason their analytics agents produce substandard insights. That finding comes from MIT Technology Review Insights, in custom research sponsored by Snowflake and published November 2, 2026. A day later, at Expedition 2026, Snowflake answered with product: Business Ontology to ground agents in approved definitions, and Cortex AI Gateway as a front door for policies, spend attribution, and model routing. The pairing is deliberate. Context without guardrails still burns money. Guardrails without context still produce answers nobody trusts.

The same tension shows up in the customer quotes Snowflake put next to the launch. Andy Owens, VP of Data and AI at Kargo, said the company “couldn’t answer was which teams and which agents were driving the spend.” After the gateway, Kargo got a breakdown, hard limits, and routing that keeps routine work off the most expensive models. Emrecan Dogan, chief product officer at Glean, put the other half of the problem in one line: as agents grow more sophisticated, context has to be matched with equally strong governance. This post reads the announcement against the MIT Technology Review Insights numbers, then asks what AI agent governance still requires once a single-platform gateway is in place.

What Snowflake announced at Expedition 2026

Snowflake’s November 3 press release frames two Horizon Catalog capabilities as the foundation for trusted agents. Business Ontology, described as public preview soon, is the place to record what business terms officially mean and to serve only approved definitions to agents. Cortex AI Gateway applies one set of policies to agent connections across clouds, frameworks, and providers. Christian Kleinerman, Snowflake’s EVP of Product, said the enterprises that win “won’t be the ones with the most agents” but the ones whose agents know what the business means, abide by guardrails, and can account for the actions they take.

On the gateway side, Snowflake highlights cost governance that attributes spend to a user, team, department, agent, or app and enforces hard limits; observability that records agent activity in the customer’s own Snowflake account; Dynamic Model Routing that matches requests to organization-approved models by cost, accuracy, and latency; Agent Abuse Detection that analyzes full trajectories; and MCP governance for tool calls. Availability is not uniform. Cost governance and observability are generally available in the footnotes Snowflake published with the release. Dynamic Model Routing, Agent Abuse Detection, and Agent Trace Intelligence are listed as private preview. MCP governance through Cortex AI Gateway Tools is public preview. Business Ontology itself is public preview soon. Leaders evaluating the stack should treat those footnotes as part of the product story, not fine print.

Taken together, the announcement is a bet that AI agent governance starts with shared meaning and a single policy edge for agents that touch Snowflake data. That bet maps cleanly to the problems the sponsored research called out the day before. It also leaves open a measurement question that most enterprises already live with: plenty of agents never route through Snowflake at all.

What the MIT Technology Review Insights numbers add

The MIT Technology Review Insights report “Raising the quality of insights in agentic analytics,” published November 2, 2026, is custom content from Insights, not MIT Technology Review’s editorial staff, and was sponsored by Snowflake. Methodology on the press release: July and August 2026, 400 CDAO, CIO, CTO, CAIO, and other senior data and technology executives at organizations with at least $500 million in annual revenue. The numbers that matter for governance are blunt.

On average, 48% of analytics-agent insights lead to actions that improve the business, up from 37% a year earlier according to respondents. “Insight leaders,” a smaller cohort, say more than 70% of agent insights lead to improving actions. Fifty-five percent of executives cite lack of business context as the main reason for substandard insight quality; 45% cite fragmented or inaccessible data. Fifty-two percent say agents generate too many insights to properly evaluate. Sixty percent are prioritizing stronger AI and data governance, and 45% are adding a semantic layer.

Those priorities explain why Snowflake led with ontology and a gateway rather than another model announcement. A semantic layer and governed definitions attack the 55% context problem. Cost attribution, hard limits, and routing attack the spend side of the same agent sprawl. What the survey does not claim is that platform-native controls alone close the loop on outcomes. Nearly half of agent insights still fail to improve the business on average, and more than half of organizations cannot evaluate the volume of insights their agents produce. That is a measurement gap as much as a model gap, and it is where an AI system of record starts to matter beyond any one data cloud.

The measurement gap a platform gateway does not close

A Snowflake-native gateway can govern agents that live on Snowflake data: attribute their spend, enforce hard limits where those controls are available, route routine work to cheaper models, and keep traces where the customer can query them. That is necessary work for any team running agents against a warehouse at scale. It is not the same as enterprise-wide AI agent governance. Most large organizations also run ChatGPT Enterprise, Microsoft Copilot, Claude, Cursor, n8n workflows, and custom agents on other clouds. Those agents have their own keys, owners, and invoices. A hard limit inside one platform does not see a coding agent that jumps overnight, or a stolen key that burns a Bedrock bill in thirteen hours, the pattern we covered in our LLMjacking analysis.

This is the same unit-of-measure problem finance teams hit when they watch a total AI invoice instead of spend by owner, project, and outcome. In our write-up of Deloitte’s AI cost findings, only a small share of CFOs said their organizations respond to failed AI cost controls in near real time. FinOps for AI fails the same way when attribution stops at the provider boundary. Cost governance that stops at Snowflake is real progress for Snowflake workloads. It still leaves the CFO and CISO stitching together screenshots from every other stack.

Olakai was built for that wider surface. As a vendor-neutral Enterprise AI Intelligence Platform, Olakai sits as a measurement layer across coding, assistive, and agentic tools, attributing spend and usage to people, projects, and providers rather than to a single warehouse edge. Olakai Agentic and the broader complete AI monitoring view exist so leaders can see agents that never touch Snowflake next to the ones that do. Platform gateways and a cross-stack system of record are complementary: one enforces policy where the data lives; the other answers which agents exist, who owns them, what they cost, and whether they return anything.

What leaders should do now

Snowflake’s launch and the Insights survey point to the same operating checklist, whether or not your agents run on Snowflake tomorrow:

• Inventory every production agent and agent-like workflow across stacks, including Copilot, ChatGPT, Claude, Cursor, n8n, custom services, and warehouse-native agents, and assign a named owner to each one.
• Require spend attribution to a user, team, or cost center before an agent goes to production, and turn on hard caps where the platform supports them, the way Cortex AI Gateway’s generally available cost governance is designed to do.
• Treat business definitions as governed assets: if two dashboards disagree on net revenue, an agent will invent a third. A semantic layer or ontology with accountable owners is the fix the 55% context finding is asking for.
• Measure outcomes, not only calls. Track which agent insights lead to actions that improve the business, and cut or retrain agents that generate volume nobody evaluates.
• Separate provider-side enforcement from cross-stack visibility. Quotas and hard limits stop a runaway process. A measurement layer tells you which keys and agents exist when the bill arrives from somewhere else.

For a CFO, the immediate ask is attribution and caps tied to owners, plus an AI ROI view that connects spend to outcomes rather than to call volume. For a CISO, the same inventory is a risk register: unowned agents are shadow AI with a budget. Inside AI spend governance and n8n agent monitoring show how that inventory looks when workflow tools and shared service keys are in scope, not only chatbots.

Where to start

Start with the agents that already touch revenue or customer data and answer four questions for each one: who owns it, which platform it runs on, how spend is attributed today, and what share of its insights led to an improving action in the last quarter. If the answers require a tour of five consoles, you have the problem Snowflake is selling into and the gap a single gateway cannot cover alone. Organizations that put Business Ontology-style definitions and gateway-style limits where their data lives, and keep a vendor-neutral system of record across every other agent stack, will treat the next spend spike as an incident with an owner. Those that only buy the gateway will govern one door and leave the rest of the building unlocked.

Talk to an expert about inventorying agents across stacks, attributing spend to owners, and measuring which insights actually improve the business. Olakai gives finance, security, and AI leaders one Enterprise AI Intelligence Platform view of what agents cost, who runs them, and what they return, with AI governance that extends past any single data cloud.

Thirteen hours. That is how long it took for one stolen enterprise AI key to make about 1.7 million calls to Amazon Bedrock, more than a million of them successful, before someone shut it off. A second stolen key made roughly 800,000 calls over thirty hours. Wiz, which traced both keys back to a breach at an AI developer platform, estimates that the two cases alone likely ran up more than $100,000 in usage at list price. The company that owned the keys paid for every one of those calls, and none of them did any work for that company.

Security researchers call this LLMjacking: stealing the credentials that grant access to AI models and then reselling or consuming that access on someone else’s bill. It used to be a niche threat. In October 2026 it became a pattern, and it changes how finance and security leaders should think about an AI key. A key is no longer just a credential to protect. It is a line of credit, and anyone who holds it can spend from it.

What happened in “Black October”

In a report published October 30, Wiz listed at least nine AI developer platforms that disclosed security incidents during the month, starting with Fireworks AI on October 6 and ending with Northflank on October 30. Wiz found no single vulnerability or dependency linking the incidents, and the root causes vary from credential access to a third-party vendor dependency. What the public reports share is the target. According to Wiz, they all describe attackers “explicitly hunting stored environment variables, with a focus on those granting AI access.” Three of the platforms reported indications of an “agent swarm” attack, and two attributed their incidents to attempts at distillation, which means training another model on a stronger model’s answers.

Railway’s preliminary report is one of the more detailed first-hand accounts. It says the attack ran from October 20 at 13:58 UTC to October 22 at 03:32 UTC, and that early investigation pointed to an attempt “to obtain AI provider API keys to perform distillation attacks.” Railway reported that logs from the attacker’s payloads showed the specific API key names they were searching for, and it recommended that every customer who signed up before October 21 rotate the keys stored on the platform.

How a stolen key turns into a bill

The most useful part of the Wiz report for a finance or AI leader is the timeline of what happened after the theft. In both cases Wiz traced, the first unauthorized use came within 30 minutes of the reported credential access at the vendor: a single model call from a residential IP address, part of a one-minute sweep that tested stolen keys across several organizations. For days afterward, each key saw only a trickle of traffic, between 14 and 120 calls a day, which Wiz reads as a job checking that the keys still worked. Then the volume arrived.

In the first case, four days after that initial test, the attacker probed which models were available, enumerated quotas, and started bulk traffic. The key’s policy allowed global cross-region inference, so the load spread across regions and only about 6% of the roughly 800,000 calls were throttled. About three quarters of the successful calls went to Anthropic’s Opus models, the most expensive tier. In the second case, the key was restricted to US regional inference profiles. The reseller pool first failed thousands of times because it requested global profiles, then found working model IDs and opened the flood. Nearly all of the calls that did not succeed were throttled by the account’s regional quotas.

Wiz describes the usage pattern as resembling resold coding-agent traffic, with many different end users behind a reseller, and it links the keys to two overt LLM resale operations offering discounted access to frontier models. The victims were quietly subsidizing someone else’s AI program.

Why this is a spend problem, not only a security problem

Most organizations still treat AI keys the way they treat any other secret: store it, rotate it on a schedule, and hand it to the security team when something leaks. Black October shows why that is incomplete. A stolen AI key exposes the budget, and the damage compounds by the hour. In our post on Deloitte’s AI cost findings, we noted that only 7% of surveyed CFOs said their organizations respond to failed AI cost controls automatically or in near real time. Put that next to a key that can generate a million successful calls in thirteen hours, and the gap between detection and response becomes the cost of the incident.

This is also why a total AI bill is the wrong unit to watch. A spike on an invoice tells you that something happened weeks after it happened. What you need to know is which key, which project and which owner the spend belongs to, and whether that pattern matches anything the business actually does. This is the signal Olakai was built for. As a vendor-neutral Enterprise AI Intelligence Platform, Olakai attributes AI spend by developer, provider and project, and its Projects feature groups shared service keys into named cost centers with owners and monthly limits. Spend that cannot be tied to a specific person still counts toward program and provider totals but drops out of the person-level views, so a key that suddenly spends on behalf of nobody you can name stands out instead of disappearing into an aggregate.

What actually contained the damage

The most practical lesson in the Wiz report is about what slowed the attackers down, and it was not a dashboard. In the second case, the key’s policy only allowed US regional inference, and the account’s regional quotas throttled nearly all of the calls that did not succeed. In the first case, the key allowed global routing, and only about 6% of calls were throttled. Key scope and provider quotas acted as the hard stop. Both were set up long before the attack.

That points to a split worth making explicit. Provider-side controls, such as key scope, regional restrictions, spend caps and quotas, are what stop a flood in progress. A measurement layer is what tells you which keys exist, who owns them, what normal looks like and when the trajectory is wrong. Olakai budgets are evaluated once a day, right after the cost-import job pulls fresh provider spend, and they fire threshold alerts at 50%, 80% and 100% along with a forecast alert when the run-rate projects past the limit by month end. Those budgets cover per-token, cost-bearing providers such as Anthropic, Cursor and OpenAI. A daily alert will not stop a thirteen-hour flood. It will make sure the morning after is a phone call instead of an invoice three weeks later, and it shows which project and owner the key belongs to so the right person can act. We wrote about the same gap in Uber’s AI budget blowout, where the guardrail arrived only after the money was gone.

What leaders should do this week

If your teams build on any of the nine platforms Wiz lists, the first step is the one every affected vendor has already recommended: rotate every AI provider key stored there and treat the old ones as exposed. Beyond that, a short list covers most of the exposure:

• Inventory every AI provider key in use, including service keys stored on third-party build, hosting and agent platforms, and give each one a named owner.
• Scope each key to the models and regions it actually needs, and set provider-side quotas or caps so a stolen key hits a ceiling.
• Attach every shared service key to a project or cost center so its spend has a baseline and an accountable person.
• Review spend per key at least daily, and decide in advance who is authorized to revoke a key without waiting for a meeting.
• Ask each AI platform vendor how it stores your keys, and how quickly it will tell you when they may be exposed.

For a CISO, this belongs in the same AI risk program that already covers shadow AI and data exposure, because an unknown key is a form of shadow AI with a direct cost attached. For a CFO, it belongs in the monthly AI review, alongside the unit-cost metrics that show whether AI spend is creating value at all. Olakai brings those two views together. The same AI system of record that tells finance what each project returns also tells security which keys exist, who owns them and what they spent yesterday.

Where to start

Pick your three highest-spending AI provider accounts and answer four questions for each one: how many keys are active, who owns each key, what each key is allowed to call, and how long it would take to revoke one at 2 a.m. If any answer requires a search through environment variables on platforms you do not control, Black October is a preview of your exposure. Organizations that can see every key, every owner and every day’s spend in one place will treat the next breach as an incident. Those that cannot will meet it on an invoice.

Talk to an expert about attributing AI spend to every key, project and owner, and setting budgets that alert before a bad month gets worse. Olakai gives finance, security and AI leaders one view of what AI costs, who is using it and what it returns, with AI governance that extends to the keys and vendors your AI depends on.

Deloitte asked 3,500 leaders what they track about AI cost. 71% said the total bill. 44% said what one unit of it costs. Eight days earlier, a separate Deloitte survey of 200 CFOs had found that when an AI cost control fails, only 7% of their organizations respond automatically or in near real time.

Read side by side, the two surveys describe a situation most finance teams will recognize. The AI bill is a number that keeps growing. Nobody can say with confidence which workflow, team or model is driving it. And when it crosses a line someone approved, it takes days or weeks to do anything about it. This post walks through what the surveys found, why the gap between the total and the unit matters more than the total itself, and what a finance leader should ask to see before the next budget cycle.

What Deloitte found, in two surveys

The State of AI in the Enterprise report, released October 29, surveyed 3,500 director to C-suite leaders across 25 countries and six industries between May and June 2026. Most of the headline is positive. 81% of respondents report faster decision-making and execution, and 28% now report revenue gains from AI, up from 20% in Deloitte’s January report, which we covered in The Enterprise AI Revenue Gap. The cost finding sits further down the page and deserves a slower read. 71% track total AI inference and API spend, 51% track token volume, and just 44% track per-unit inference cost.

Costi Perricos, Deloitte’s Global GenAI Business leader, framed it this way: “Many enterprises are getting better at watching their total AI bill, but far fewer understand the unit cost underneath it.” The next line was short: “That gap is where ROI can get lost.”

The Q3 CFO Signals Survey, published October 21, polled 200 CFOs at North American organizations with at least $1 billion in revenue. 53% said their primary practice is to review AI usage costs retrospectively, through vendor bills, periodic reports, dashboards or manual analysis, while 47% monitor in or close to real time. Only 20% have advanced AI-specific controls such as token or workload-level attribution, AI gateways or caching. 56% rely on basic AI-specific controls and 24% on general financial controls alone. The two surveys sample different people, so the numbers should not be stacked into a single statistic, but they point in the same direction.

Why a total bill is not a number you can manage

A total tells you what happened. It does not tell you why, and it does not tell you what to change. That worked well enough in a seat-based software world, where cost stayed predictable and higher usage usually meant the rollout was succeeding. Deloitte’s CFO survey notes that commercial AI providers now more commonly charge per token than per seat, and that bills can spike quickly because of large context windows or long conversation loops. Under that pricing, usage and cost stop being separate questions.

The CFOs were asked what limits their ability to oversee AI spending. 40% named a lack of AI cost governance, processes or policies. 38% named limited timely insight into token usage and unit costs. 36% named limited standards for model selection, prompts, data retrieval or workflow design. Notice that the second and third answers describe visibility and standards below the level of the invoice. The same pattern shows up in our writing on FinOps for AI, where the cloud cost playbook starts to break once the unit of billing becomes a token, and in why an AI bill can grow even as per-token prices fall.

What unit cost means for a copilot, a coding tool and an agent

“Unit cost” is not one thing, which may be part of why fewer than half of Deloitte’s respondents track it. For a copilot, a useful unit is cost per active user per month, set against how often that person actually uses the tool. For a coding assistant, it might be cost per developer or per merged pull request. For an agent, it is cost per completed workflow or per resolved case. Each one requires the same groundwork: attribute spend to a person, team, provider and project, then join it to a work outcome. Deloitte’s own advice points the same way, saying leaders need visibility into workflow-level costs and the usage that drives them.

This is the layer Olakai was built for. Olakai is a vendor-neutral Enterprise AI Intelligence Platform, and it attributes spend across tools such as Anthropic, Cursor and OpenAI by developer, provider, project and department. It also reconciles provider cost data against what was actually billed, including the day or two of lag in Google Vertex AI cost data, so the unit costs a finance team sees trace back to a real invoice. Custom KPIs then tie that usage to an outcome the business already measures.

The days between an overrun and a shutoff

The most uncomfortable finding in the CFO survey is about speed. When AI cost controls fail, 74% of respondents say their organizations typically pause or shut down the related usage within several business days to several weeks. 19% address it the same business day. 7% respond automatically or in near real time. Compare that to how an AI bill moves. A long conversation loop or a runaway agent does not wait for the next monthly review.

Closing that gap has two halves, and most organizations have neither. The first is detection early enough to matter, meaning an alert that arrives while the month can still be saved rather than an invoice that arrives after. Olakai budgets are evaluated daily against imported cost data, with threshold alerts at 50%, 80% and 100% of a budget and a forecast alert when the current run-rate projects past the limit by month end. We describe the mechanics in Inside AI Spend Governance. That is a day, not a week, and we would not call it real time. The second half is a decision about who is authorized to pause what, written down before the overrun instead of negotiated during it. No tool replaces that policy. It can only make sure the alert shows up while the policy can still be applied.

Autonomy is rising while accountability stays unclear

The State of AI report adds a reason why response speed matters beyond cost. 40% of enterprises now allow AI systems to execute decisions or take actions without human approval, at least to a moderate degree, while 23% say accountability is unclear when an AI system produces a negative outcome. Deloitte’s guidance is that governance and safety controls should scale with the level of autonomy and the potential consequence of an action, with systems that execute needing more oversight than systems that only recommend. We cover that side in our AI governance work.

An agent that can act is also an agent that can spend. It is a reasonable inference, though not a Deloitte finding, that organizations unable to name an owner for a bad outcome often cannot name one for a bad bill either. Agent IQ in Olakai addresses this at the agent level, with executions, cost, outcomes and an audit trail for each agent, so ownership and spend sit on the same record.

What a CFO scorecard should show

If the goal is to move from watching the total to managing the unit, the monthly AI review needs a different set of numbers. A workable starting scorecard has five lines:

• Unit cost for each major workflow, copilot seat or agent, not just spend by vendor.
• Token volume by tool, team and model, so a spike has a place to be traced to.
• Time to pause, meaning the median hours between a budget breach and an actual decision.
• Share of AI spend attributed to a named owner.
• Number of AI systems allowed to act without human approval, each with an accountable person.

None of these require a new finance process. They require the same attribution and ownership data that an AI system of record holds, which is why that record matters to finance, and why the CFO view in the CFO use case is built around spend, usage and return together. If you want a shorter list, our post on three token cost metrics for CFOs is a good place to begin.

Where to start

Pick the single most expensive AI tool or agent in your portfolio and answer three questions for the last month: what did it cost per unit of work, who owns it, and how long would it take to pause it. If any of those takes more than an afternoon to answer, you have found your version of the 44% and the 7%. The case for AI is getting stronger, with 28% of Deloitte’s respondents now reporting revenue gains. But the same question we raised about JPMorgan’s $700 million AI bet applies at every scale: the return is only believable when the cost side is measured as carefully as the benefit side.

Talk to an expert about measuring AI cost per unit, setting budgets that alert before the month is lost, and putting an owner on every AI system. Olakai gives finance, IT and AI leaders one view of what AI costs and what it returns.

From Enterprise AI Weekly, recorded 23 October 2026. Barclays is cutting about 5,000 jobs this year, and the headlines say AI did it. On the same call, the bank’s own CEO called AI “the icing on the cake, not the cake.” So which one is it? That is the topic of the week, and around it we do the full read of the enterprise AI room: the JPMorgan story we broke on Tuesday and what the big US banks just did about outside AI agents, a line Meta’s CFO put on an earnings call that I have never heard a big-tech CFO say out loud, Satya Nadella’s annual letter and a scoreboard of the five vendors fighting for your AI budget, and the number of the week, which belongs to the CFO.

Same format as always: one round, five stops. The math is directional, so check it and tell me where you read the room differently.

Stop one: did AI really cut 5,000 jobs at Barclays?

Barclays reported its third quarter on Thursday, and the line everyone picked up was the headcount: about 5,000 roles going this year, roughly 5% of the whole workforce, most of them already gone and the rest in the second half. Let me give credit first, because Barclays is running a disciplined plan. Pre-tax profit was up 12% on the quarter. The bank is targeting more than £2 billion of cost-efficiency savings out to 2028. The cost-to-income ratio is 57% year to date, with a target in the low 50s by 2028, and finance director Anna Cross called that ratio the bank’s North Star. That is a CFO who knows exactly what she is measuring.

They also shared real operating numbers, which I always appreciate. Mortgage applications went from 45 minutes to under 15. Call volumes are down 10%. Whether that second one is AI, I am not sure, but it is the kind of evidence I look for. Then it gets interesting. Cross said about 60% of the cuts are in support functions, “including fewer developers as AI makes our people more productive.” And the CEO said AI is the icing on the cake, not the cake. So the press wrote “Barclays cuts 5,000 jobs because of AI,” and the CEO is telling you most of it is simplification: standard data, modern infrastructure, fewer platforms. AI sits on top.

From my seat, both can be true. We have covered several AI-related layoff stories on this show this year, and only one side of each is ever measured. The headcount is measured to the person. The AI contribution is a sentence on a call. If I were a board member at Barclays, the question is simple: of the savings, how much came from AI, how much came from simplification, and how would we know? The developer line is the one I would watch hardest. “Fewer developers because AI makes the rest more productive” is a bet on coding tools, and that bet is measurable. Baseline each developer before and after, track cycle time and the mix of work that ships, and put a cost on each finished change. If you cannot show that, you cut the team and hope the tools carry it. That is the question Olakai Agentic was built to answer before the headcount decision, not after.

Stop two: JPMorgan, and the banks closing the door on outside agents

If you watched Tuesday’s episode, you know the story. JPMorgan spends about $700 million a year on AI, and its CFO Jeremy Barnum told analysts the benefits “aren’t observable in our current quarterly earnings.” A very different line from the Barclays CEO. On Tuesday we broke that bill across about 150,000 weekly users: each one needs to save about an hour a week for the spend to pay off, JPMorgan’s people say they save four, and the CFO still cannot book it. Put the two banks side by side. Barclays can show you the cost it took out and cannot yet split how much of it was AI. JPMorgan can show you the usage and cannot yet show the return. Same gap, opposite ends of the ledger.

And this week the big US banks did something new. Axios reported that JPMorgan, Morgan Stanley, Bank of America, Citi and Wells Fargo are restricting outside AI agents after some of them reached customer accounts and did things customers had not asked for. Jamie Dimon put it plainly: “We’re not going to allow just any outside agents to come in inside the company.” Here is why that matters to you. The banks’ problem is attribution: which agent did this, and on whose instruction? That is the exact same question inside your company. If you cannot say which agent took which action for whom, you cannot price it and you cannot claim a return on it.

Stop three: Meta has a token bill, and said so

This one honestly surprised me, and I double-checked it. On Wednesday’s call, Meta CFO Susan Li explained what drove expense growth. Her words: legal-related costs, infrastructure costs, and third-party AI token costs. Think about that. Meta built its own frontier models, the same ones we priced against Claude and GPT a couple of weeks ago. It is guiding $135 to $145 billion of capex this year. And it still buys enough tokens from somebody else to name the line to Wall Street. She also said total serving costs are expected to rise, and Mark Zuckerberg said Meta expects to make its models more than ten times more cost-efficient over the next year. That is a target, not a result, and I will be watching for it.

The read-through is simple. If the company with more owned compute than almost anyone carries a token bill worth disclosing, your company carries one too. Most of you just do not have a line for it yet. That is the FinOps for AI problem in one sentence: the cloud bill has a line, the token bill is scattered across five vendors and a dozen expense reports.

Stop four: Nadella’s cost-to-outcome frontier, and the five-vendor scoreboard

Satya Nadella published his Microsoft annual letter on Wednesday. Big numbers in it: Azure passed $100 billion, up 41%, and there are over 30 million paid Copilot seats. But one line matters more to a CFO than any of those. “Customers will increasingly measure the cost of tokens against the outcomes they produce.” He calls it the cost-to-outcome frontier. I agree with him, and if you know this show you know I have been standing on that thesis for a while.

So I lined up the five vendors competing for your AI budget. Microsoft, with the installed base. Google, whose cloud grew 82% last quarter. OpenAI, where enterprise revenue passed consumer this summer. Anthropic, the enterprise and coding story, which we covered when investors got nervous about a $16.8 billion quarter. And Meta, which still makes its money from ads. Every one of them reports its own consumption. Microsoft even built financial controls for agents into Agent 365, for the agents running inside Microsoft. None of them reports the return on your AI spend across all five, and most enterprises I speak with buy from several. The vendor sees its line item. Your CFO owns the whole P&L.

Stop	Number	Source
Barclays job cuts	About 5,000 roles in 2026, roughly 5% of the workforce, 60% in support functions	Barclays Q3 2026 call
Barclays efficiency plan	Pre-tax profit up 12%; more than £2B of savings to 2028; cost-to-income 57%, targeting the low 50s	Barclays Q3 2026 results
Barclays operating evidence	Mortgage applications 45 minutes to under 15; call volumes down 10%	Barclays Q3 2026 call
JPMorgan AI spend	About $700M a year; benefits “not observable” in quarterly earnings	JPMorgan Q3 2026 call
Banks restricting outside agents	JPMorgan, Morgan Stanley, Bank of America, Citi, Wells Fargo	Axios
Meta	Third-party AI token costs named as an expense driver; $135B to $145B capex guide	Meta Q3 2026 call
Microsoft	Azure past $100B, up 41%; 30M+ paid Copilot seats; “cost-to-outcome frontier”	Nadella annual letter
Google Cloud	Grew 82% last quarter	Alphabet
CFOs over AI budget	83% over budget on consumption-based AI; 14% within; more than a third over by 25%+; confidence 8.2/10	Pigment CFO Index, about 2,000 finance leaders

Stop five: the number of the week is 83%

Who is feeling all of this? The CFO, and the FP&A team under them. The number of the week is 83%. That is the share of finance leaders in Pigment‘s latest CFO survey, about 2,000 of them across the US, the UK, France and Germany, who said consumption-based AI costs pushed spending above budget. Only 14% stayed inside it. More than a third went over by more than 25%. And here is the part I love: their confidence that AI delivers value is 8.2 out of 10. They believe in it and they cannot forecast it at the same time. That is the measurement problem, that is the gap, and it is fixable. It is a vendor survey, so treat it as directional, but it matches what I hear in every CFO conversation and every engagement we run.

Being over budget is not the failure. Being over budget with no forecast and no alert that fired before the overrun is the failure, because it means the first person to see the number was the one paying the invoice. An 8.2 confidence score with an 83% overrun rate is a room full of people who are sure the cake is good and cannot tell you what it cost to bake.

Why we built the Enterprise AI System of Record

Every stop this week is the same gap seen from a different seat. Barclays has the cost and not the AI share. JPMorgan has the usage and not the return. The banks have agents and not the attribution. Meta has a token line and no outcome next to it. The vendors each see their own meter, and the CFO owns the whole bill. This is why we built Olakai, the Enterprise AI System of Record. It captures every AI interaction across your copilots, coding tools and agents, from every vendor on that scoreboard and beyond. It structures them by person, team, model and workflow, and it puts the cost of each finished piece of work next to the outcome it produced. So when the board asks Barclays’ question, “how much of this was AI?”, you have an answer that is a number, not a sentence.

And because every bank asks this first: it runs in your own cloud and you hold the keys, which is the posture our financial services customers need before anything else.

The one line for the board

Here is the line I would take to the board this week: we know what our AI costs across every vendor and what it returned, before we cut a single role because of it. And the one question for the comments. If your CFO asked how much of last quarter’s savings came from AI, could you answer with a number? Your AI is an investment, so let’s measure it like one. If you read Barclays differently, challenge me. I will see you on Tuesday, where we break the math.

From the AI ROI Series, Show Me the Math, recorded 20 October 2026. $700 million. That is what JPMorgan says it spends on AI every year. Last Tuesday on the bank’s third-quarter call, an analyst asked what that money actually returns, and CFO Jeremy Barnum said the benefits “aren’t observable in our current quarterly earnings.” He said he is really trying to move away from quantifying them, and that some of the benefit is going to be offset by token costs. $700 million a year from the number one AI bank in the world, and the return does not show up in the numbers.

So let’s do what this show does and break the math. On Friday’s episode I promised a deep dive on JPMorgan, and here is the plan: first what $700 million actually is, then what they built with it, then the math on both sides of the ledger, and the one thing that is missing, which honestly most of us are missing too. The math is directional, so check it and tell me where you read it differently.

What $700 million a year actually is

$700 million a year is about $1.9 million a day, every day, weekends included. And it used to be bigger. A year ago Jamie Dimon said, “For $2 billion of expense, we have about $2 billion of benefit.” I always wondered what he meant by $2 billion of benefit and how it was measured. The missing piece arrived last week, when the CFO said that $2 billion is not exactly right. The real AI number is something like $700 million, and the old benefit figure included classic machine learning like fraud models. That is a very different conversation, and in a way a more honest one.

Now put the $700 million against the people using it. About 150,000 JPMorgan employees use the bank’s in-house LLM Suite every week. Divide it out and that is about $4,700 per user per year, roughly $390 a month. So how much time does each person have to save to pay that back? At $100 an hour fully loaded, $700 million buys 7 million hours. Spread across 150,000 people, that is about 47 hours a year, or about one hour a week. One hour a week and the AI pays for itself. JPMorgan’s own people say they save four.

So on paper this is the best AI investment in banking: four times break-even, a very decent ROI, and the CFO says he cannot see it in the earnings. That gap is what this whole episode is about, and it is the same gap I see in almost every enterprise AI program. The one-hour-a-week break-even is easy to compute and impossible to defend, because nobody can point at the hour. That is the problem a system of record for AI exists to solve, and we will get there.

Credit where it is earned

Let me be fair, because JPMorgan earned its reputation. The Evident AI Index just ranked it number one among the world’s banks for the fifth year running. LLM Suite went from zero to 200,000 people in about eight months. The bank runs close to a thousand AI use cases, and Dimon himself said the really important ones number about 50. I love that line. It is the most honest portfolio view I have heard from a CEO. And $700 million is only about 3.5% of a $19.8 billion technology budget, so they can afford to get this right and make a few mistakes along the way. When I first covered what JPMorgan measured a year ago, the question was whether anyone could run AI at that scale. They answered it. The question now is what it is worth.

What JPMorgan built with it

A lot, and almost all of it in house. LLM Suite routes work to OpenAI and Anthropic models. There are coding tools for 65,000 technologists, with AI use now part of their performance goals, a rollout I looked at when the bank pushed agentic AI to 250,000 employees. And there is Claude Code for engineers, with a reported cap of $2,000 a month per person. That is another large company assigning a specific allowance to an employee, in this case a developer. On Friday we talked about Accenture centrally funding $29 of Claude a month per person. If your company does something similar, I would love to hear the number.

Then there is what I found in the job postings. JPMorgan is hiring principal engineers at $204,250 to $285,000 base salary to build a GenAI gateway with routing, quotas and, quote, “per-tenant cost attribution.” Translation: they are building their own AI meter. They have probably built some of it already. Here are the numbers from the episode in one place.

Line item	Number	Source
AI expense	About $700M a year, about $1.9M a day	CFO, Q3 2026 call
Previous figure	$2B expense, $2B benefit (included classic ML)	Dimon, October 2025
Weekly LLM Suite users	About 150,000	JPMorgan
Cost per user	About $4,700 a year, $390 a month	Our math
Break-even	About 47 hours a year per user, one hour a week at $100/hour	Our math
Self-reported savings	Four hours a week	JPMorgan employees
Value if the four hours are real	27.6M hours, about $2.8B a year	Our math
Technology budget	$19.8B, AI is about 3.5%	JPMorgan
Claude Code ceiling	8,000 licenses x $2,000 a month, up to $192M a year	Our math, ceiling not actual
Bank of America	About $400M spend, about $800M benefit	Bank of America

The value side: four hours nobody will book

Back to the math. If the four hours are real, what is the full value? 150,000 people times four hours a week is 600,000 hours a week. Over 46 working weeks, that is about 27.6 million hours a year, and at $100 an hour, roughly $2.8 billion. Four times the $700 million. So why won’t the CFO book it? Because Dimon told investors in February that those hours are, quote, “not in an NPV.” They are self-reported. Nobody can tie them to a finished piece of work.

This shows you how hard true AI ROI is. Either JPMorgan is sitting on a $2 billion return it cannot prove, or the hours are not real, and the CFO can book neither. Many of us are in exactly this position, especially at a public company where the right instinct is to be conservative. A survey is not a ledger. The hour saved only becomes value when you can say which task it came out of, what that task cost in tokens and licences, and what the finished work was worth. That is the FinOps for AI problem in one sentence: the cost side has an invoice and the value side has a questionnaire.

The cost side: measured to the penny

I promised the cost side as well. In July the CFO called token expense “a trivial number” for the first half, with meaningful acceleration expected in the second. By October it is a whole other conversation. A JPMorgan payments data executive said some employees are spending more on tokens than their salary. I have heard those stories before, and the pattern is familiar: a burst of tokenmaxxing in the first half, then the caps arrive. And those Claude Code caps are real money. 8,000 reported licences times $2,000 a month is up to $192 million a year. That is the ceiling, and the real bill is probably far lower, but the ceiling alone is more than a quarter of the whole AI budget, on one tool, for one job family.

Gartner says inference cost per agentic workflow rises more than fivefold through 2028. So here is the asymmetry: the cost side is measured to the penny, and the value side is a self-reported survey. A cap per developer is the right reactive move, the same one Uber made after its budget blowout, but a cap tells you who spent more. It does not tell you whose work came out better. For that you need budgets that alert before they bite and a cost per finished task sitting next to each one.

Build versus buy

Everything they built measures activity

I respect a bank that builds. JPMorgan can afford it, and its CISO told SaaS vendors last year that most integration models were not good enough for the bank. Fair. But look at what got built: a gateway, cost attribution, and a dashboard that tags engineers as light, heavy or non-users. Every one of those measures activity. None of them tells you whether the work that came out was worth what went in. And every internal tool is a product. It has a team, a roadmap and an on-call rotation. Fifty engineers at those salaries, fully loaded, is roughly $17 million a year before a single dollar of value gets booked.

MIT NANDA studied around 300 AI deployments in “The GenAI Divide.” Tools bought from specialist vendors succeeded about two-thirds of the time; internal builds succeeded about a third as often. Statistically speaking, building your own measurement layer is a risky game even for the best engineering bank in the world. And to be fair to the industry, one bank did put a number out. Bank of America says about $400 million of AI spend generates about $800 million of benefit, two times the cost. You can argue with the method, but it is a number the board can hold them to.

So did JPMorgan live up to the hype?

On adoption, absolutely yes. On engineering, best in the world. On proof, not yet, and the CFO said it himself: the benefits are not observable in current quarterly earnings. A meter without a value ledger is the current state, at JPMorgan and at most of the companies I talk to. This is exactly why we built Olakai, the Enterprise AI System of Record. It captures every AI interaction in your business across copilots, coding tools and agents, structures it by person, team, model and department, and ties the cost of each finished task to the outcome it produced. The four hours become evidence, and the token bill sits right next to the return. The cost side is the easier half of this industry; the finished-task side is the half that is still mostly missing, and it is the half a CFO can actually defend.

And because banks care about this one: Olakai runs in your own cloud. No standing access, no inbound path. You hold the keys and it sits behind your firewall, which is the posture every financial services customer asks us for first. You do not need 50 engineers to build it, and you do not carry the maintenance. You just need the record.

Four questions and one line for the board

Here is how I would close this quarter. Every AI leader should be able to answer four questions. What did we spend on AI, by tool and by team? That one is easy. How many finished tasks did that buy? What did each one cost, and what did it return in dollars? And which of those returns would we defend to the board? The first is a bill. The second needs visibility into every coding tool and agent. The third and fourth need cost per finished task by team and by model next to the outcome, which is where most programs, JPMorgan included, still have a blank.

And the one line I would take to the board: we know what each finished piece of AI work cost, and we book only the value we can prove. One question for the comments. If a bank analyst asked your CFO what AI did for your company last quarter, could they answer in one number? Your AI is an investment, so let’s measure it like one. If you read JPMorgan differently, challenge me. I work it all day long.

From Enterprise AI Weekly, recorded 16 October 2026. Anthropic made $16.8 billion last quarter. It grew 45%, and it turned a profit. And this week, some of its investors started talking about paying less for it. That is the question I want to work through today, and I think the answer is sitting in your AI budget.

Same format as always: one round, five stops. Anthropic’s big quarter as announced. The company that just put Claude on a budget. The price war getting an end date. Meta and Apple being friends again. And Wall Street, where one bank had a great quarter and the other stopped counting. If you own the AI budget, or defend it to a CFO, this one is for you. The math is directional, so check it and tell me where you read the room differently.

Stop one: the car stopped accelerating

On Wednesday Anthropic held its investor day in San Francisco, ahead of what could be the biggest IPO ever. First quarter, $4.7 billion of revenue. Second quarter, $11.6 billion. Third quarter, $16.8 billion, and profitable on an adjusted basis. Credit first, because it is earned: last year Anthropic lost more than $8 billion on an operating basis, and now it is profitable at this scale. Very few companies in history have done that.

So why are some investors now talking about $1.5 trillion instead of $2 trillion? Think of it like a car. From the first quarter to the second, foot to the floor; revenue more than doubled. In July, Anthropic was running at $65 billion a year, a number we used on this show a couple of weeks ago. Now take the third quarter and multiply it by four: about $67 billion. The car is still going very fast, but over the summer it stopped accelerating. My math says the run rate grew about 3% from July to September, and when you pay $2 trillion, you are paying for acceleration. $2 trillion is about 30 times this revenue; $1.5 trillion is closer to 22 times. That is the gap investors are arguing about.

Why did it slow? Here is my read, and challenge me on it. First, the buyers got smarter. Two weeks ago we covered Meta cutting its Claude Code users in half and Microsoft cutting engineers’ AI budgets, both moving work onto their own tools. They are eating their own cooking now, and you will see another big buyer do something similar in stop two. Second, the field caught up. OpenAI keeps shipping, and on Tuesday’s episode GPT-6 Astra came out cheapest per task of the frontier models we compared. Google’s new Gemini 4 Argon topped Claude on a couple of independent benchmarks this week. Microsoft is building its own models, and Meta joined the frontier on Monday. Two years ago Anthropic had the enterprise lane mostly to itself. Today it is a crowded road.

Let me be fair: one quarter is not a trend. Analysts who met the team this week still talk about a run rate as high as $110 billion by year end, and Claude is still the model a lot of us, me included, use every single day. But from my seat, the hyper-growth phase was every company trying everything. This phase is every company asking what it is actually getting for the money. Anthropic’s revenue is our spend, and you can see our behaviour in the vendor’s numbers.

Stop two: Accenture puts Claude on a budget

According to an internal memo reported by 404 Media, from October 1 Accenture centrally funds $29 of Claude per person per month. Above that, the cost has to be charged to a specific project, what Accenture calls a work breakdown structure. The same reporting describes heavy users spending $500 to $1,000 a month, and Accenture’s generative AI strategy lead told TechCrunch that AI was becoming “material to the cost structure.” One memo, one outlet, so it is directional. But honestly, I love this. This is AI FinOps arriving at one of the largest consulting firms in the world: chargeback per project.

The next question writes itself. Which projects earn the average? A cap tells you who spent more. To know whose work came out better, you need cost per task by project, by department, by team if you want to go there. We saw the same shape when Uber capped AI spend per employee after a four-month budget blowout: the cap is the reactive move, and the measurement is what makes it a management tool. The detail I find most interesting is that the $29 is an average, and the super users are allowed to spend more. That is also what measurability is for: enablement, coaching, and finding the people whose $1,000 a month is the best money the firm spends. Hold that thought, because it comes back in stop five.

Stop three: the price war gets an end date

The buyers are capping the bill, so what are the sellers doing? On Tuesday Google made Gemini 4 Argon generally available at $2 per million tokens in and $10 out, and in the same announcement it published the date that changes. On December 1 it moves to $4 in and $20 out. That is exactly Meta’s price for Muse Wonder and exactly Claude Opus 5.5’s price from Tuesday’s episode. From December, three labs sit at the same frontier price, and the cheap rate is a 46-day window.

Model	Input / output per million tokens	From
Gemini 4 Argon (launch rate)	$2 / $10	Until November 30
Gemini 4 Argon	$4 / $20	December 1
Meta Muse Wonder	$4 / $20	Now
Claude Opus 5.5	$4 / $20	Now

Here is why that matters to a CFO, to FP&A, and to every FinOps professional watching. If your 2027 AI plan was built on this month’s rate, your cost per task on Argon doubles on December 1. Same work, double the price. And one detail most people will miss: cached input on Argon is 95% cheaper. If you know your cache hit rate, you can protect a lot of that bill. If you don’t, that is your first homework, because the cache line is where your money is. It all comes back to knowing your own cost per task, cost per turn, cost per merged pull request. There are so many tasks, subtasks and projects to measure before you understand the value of your AI, and every one of them starts with collecting the data.

Stop four: Meta and Apple are friends again

This is the most fun one this week. On Thursday Meta and Apple announced they have been working together to run Meta’s open-weight model, Muse Spark Open, locally on a Mac. It is an 816-billion-parameter model running across two or four Mac Studios with Apple’s new M5 Ultra chips. Meta’s AI chief Alexandr Wang put it best: “We are excited for mommy and daddy to be playing nice again.” After the last few years between those two companies, that is quite a line.

Why do I care for the enterprise? Because it runs locally. The vendor never sees your prompts, and there is no token bill; you pay for hardware instead. That is a new column on the cost comparison we built on Tuesday: price per token, cost per task, and now cost per task on hardware you own. To be clear, there is no Apple Intelligence integration announced; this is about running a model on a Mac. But Meta is serious, and I would keep watching it. It is the same reason customers ask us to run Olakai inside their own walls: once the model is on your hardware, the measurement layer has to be there too.

Stop five: two banks, two very different quarters

Goldman Sachs reported $18.7 billion of revenue against about $16.7 billion expected, with equities revenue up 72%. A blowout. JPMorgan missed on both revenue and profit for the first time in nearly three years. But the part I cannot stop thinking about came on JPMorgan’s call, when an analyst asked about AI. CFO Jeremy Barnum said the AI expense is something like $700 million a year. Then he said the firm is “really trying to move away from quantifying the benefit,” that the benefits “are not observable in our current quarterly earnings,” and that some of them “are going to be offset with token costs.”

So here is the largest bank in the US spending hundreds of millions a year on AI, and the CFO says the return is not visible in the numbers yet, and the token bill is eating part of it. To be fair to him, it is the most honest AI answer I have heard on an earnings call this year. It is also exactly why I want to go deep on JPMorgan on Tuesday. My very first video that went viral, about a year ago, was about what JPMorgan measured, so I cannot wait to see what has changed since last November. What are they actually building, what does it cost, and what would you need to see to call it a return? Those are questions every enterprise AI leader should be asking of their own program.

The one question behind all five stops

Let’s put the room together. Anthropic’s growth slowed, Accenture capped Claude, Google dated its discount, Meta offered a way around the token bill, and JPMorgan’s CFO said the bill eats part of the benefit. Every one of those comes back to the same question: what does each finished piece of AI work cost you, and what did it return?

That is what we build at Olakai, the Enterprise AI System of Record: every AI interaction, every vendor, every model captured and structured, so you can answer that question before your CFO asks it. Accenture’s $29 cap becomes a budget with alerts rather than a memo. Argon’s December 1 price change becomes a forecast you can run against last quarter’s actual tokens, cache hits included. And JPMorgan’s “not observable” becomes cost per finished task by team and by model, side by side with the outcome, which is the page a CFO can defend. The same record sits underneath whether you run coding agents or employee assistants, and whether you read it through our platform, through Kai, or through your own AI over our MCP connector.

One question for the weekend, for the comments: does your company cap AI spend per person yet? And if it does, what is the number? Your AI is an investment, so let’s measure it like one. That was this week’s round. Tell me if you read the room differently, and I’ll see you on Tuesday for JPMorgan.

From the AI ROI Series, recorded 13 October 2026. On Monday Meta launched Muse Wonder, and on the independent tests it sits within a few points of Claude Opus 5.5. Meta priced it at exactly the same list price as Opus, $4 per million input tokens and $20 per million output, to the cent. Mark Zuckerberg says Meta has reached the frontier. And yet in all my conversations with enterprise clients, I have yet to meet one running a Meta model in production. So who is this for, and is it worth a dollar of your AI budget?

If you own the AI budget, or you pick the models your company runs on, this one is for you. If you lead FP&A and have to defend that budget, share it. The plan is simple: where Meta sits in the enterprise league table, who Meta actually built this for, the tokenomics against Claude and OpenAI, and whether this is a buy today or a direction to watch.

The room Meta just walked into

Menlo Ventures surveyed about 500 enterprise AI buyers at the end of last year, which is a long time ago in this market, but it is the best league table we have. Anthropic held about 40% of enterprise model spend, OpenAI about 27%, Google about 21%, and everyone else, Meta, Mistral, Cohere and the rest, shared about 12%. Chinese models were around 1% of enterprise usage, and open-source share fell from 19% to 11%, partly because Llama stalled after Llama 4. So that is the room Meta just walked into: three vendors with almost 90% of the spend, and xAI pushing for a seat as well.

Let me be fair, because this is a real model. On Artificial Analysis, Muse Wonder Full scores 56 on the Intelligence Index, tied with Claude Sonnet 5.5 and two points behind Opus 5.5 at 58. On the test of real professional work it is close to a tie with Opus. That is 14 months after Zuckerberg rebuilt the whole lab and six weeks after the previous model, which is a team shipping fast, and I would expect the next one to be better. Credit to them.

Who Meta actually built this for

Zuckerberg was direct about it. He told the Financial Times that everyone else is focused on knowledge work and automation, while Meta’s goal is what it calls personal superintelligence for consumers. Follow the money and it makes sense. Meta made about $60.8 billion in revenue last quarter, almost all of it from ads, and that one quarter is roughly what Anthropic makes in a year. Meta does not need to sell you an API to pay for its AI, because AI pays for itself inside Meta. Two weeks ago on this show we covered Meta cutting its internal Claude Code users from about 60,000 to 30,000 and moving them onto its own model, and its Muse app is now the most downloaded app in the US. The first customer is Meta itself.

Enterprise is on the list. On the July earnings call Zuckerberg said they see a large enterprise opportunity to sell to businesses through APIs, business agents, and maybe compute. But the launch announcement names no enterprise customers and no cloud partners. So today, Meta’s AI is first for Meta.

The price sheet tells you nothing

Say you want to try it anyway. What does it cost per token? Identical to Claude Opus 5.5, so the price sheet tells you nothing at all. Artificial Analysis runs the same work on every model and adds up the bill, and that is where the models separate. Muse Wonder comes in at about $4.52 a task and Opus at about $5.98, so roughly a quarter cheaper, because it uses fewer tokens to get there. The one I find most interesting is GPT-6 Astra. Its list price is two and a half times higher, yet per task it is cheaper than both of them at about $3.26, and GPT-6.1 Sol is around $0.72.

Model	List price per million tokens (in / out)	Cost per task (Artificial Analysis)	Intelligence Index
Claude Opus 5.5	$4 / $20	$5.98	58
Meta Muse Wonder	$4 / $20	$4.52	52 (Full: 56)
GPT-6 Astra	$10 / $50	$3.26	53
GPT-6.1 Sol	n/a	$0.72	n/a

One quick lesson inside that number. Muse Wonder reads about 7.7 million tokens for every 54,000 it writes, and cache reads are about 20 times cheaper than fresh ones. Without the cache, I get about $32 a task instead of $4.52. My math, so check it. The cache is doing most of the work, and almost nobody reports it. It is a little geeky, I know, but the cache line is the biggest number on your AI bill, and it is the reason you have to measure cost per task, per execution, per unit of value delivered, rather than reading a rate card.

Did the work come out right?

Then the number that really matters. Nathan Romeo D. runs his own private benchmark, MSP Bench, where a model has to find and fix hidden problems in a real IT provider’s systems with a million dollars at risk. Great work, go follow him. Muse Wonder protected about $657,000 of that million. Claude Opus 5.5 protected about $893,556. Muse was much faster, four to nine minutes a run against 17 to 30, but it protected about a quarter less, and Sol protected more than Muse at a sixth of the cost. One tester, one benchmark, so it is directional. But it is the outcome, and the outcome is what goes to the board.

Think about your own enterprise AI program for a second. Do you measure it like that? Forget which model was faster; that is not my point. I am giving you the detail to make the case that you should absolutely be measuring, benchmarking, and working this out on your own workloads. Maybe that is your next step.

What to check before this goes near production

Before this goes anywhere in production, here is what I would check from Meta’s own documentation and the launch coverage. The API version is listed as a preview. The stronger model, Muse Wonder Full, is not on the API at all, only inside Meta’s Muse Code subscriptions. Those subscriptions default to a contributor mode in which Meta may train on your prompts and code, and zero data retention is available subject to approval. The launch had bugs that Meta’s own AI chief owned publicly, and there is no research paper yet. All of that is pretty normal for a week-one launch, and it is exactly what security and procurement will ask on day one. One point in Meta’s favour: the API speaks the same language as OpenAI’s and Anthropic’s tools, so trying it is close to changing a model name. A test is cheap.

Buy or watch?

So is it justifiable today? For most enterprises, honestly, I would call it a direction to watch and a cheap test on non-sensitive work through the API. It is not a platform decision yet, but I will watch it closely. Meta is spending $130 to $145 billion on capex this year, it ships every few weeks, and it clearly wants to sell to businesses the way Anthropic does. They want a piece of the pie.

For your budget this year, I think the biggest effect of Meta at the frontier may be on the price of the models you already use. Every new contender gives buyers leverage at renewal, but only if you can show what each model cost per finished task on your own work. Without that, you pay list. That is the same lesson as the cost-to-serve story from last month, seen from the buyer’s chair: the vendors measure every token to the cent, and the buyer who cannot has nothing to negotiate with.

The page you bring to a renewal

What you see in the video comes from our demo environment, built with a handful of prompts. My customers do this on their own data. Every AI dollar broken down by provider and by model, side by side. Cost per finished task by model on this company’s own work, rather than on a benchmark we quote here. And for each workload, which model would have done the same work for less. That is the page you bring to a renewal, and it is the page a CFO can defend. The same record sits underneath whether you read it through our platform, through Kai, or through your own AI with our MCP connector. This is what Olakai, the Enterprise AI System of Record, was built to do: capture every interaction and every outcome across every coding tool and agent you run, so the model question becomes a measurement question rather than a guess.

So let me ask you four things, and they are table stakes for every AI leader all the way up to the CFO.

1. Do you know which models your teams used last quarter?
2. Do you know what each one cost per finished task on your own work?
3. Do you know whether that data went through a door you approved?
4. If a new model showed up tomorrow, could you test it in a week?

The one line I would take to the board: we know what each model costs per finished task on our own work, so we can test any new one in a week and negotiate with the numbers.

Directional as always. The data is public, so check my math, and tell me in the comments whether anyone in your company has run a Meta model in production yet, and on which workload. I’m Paul, co-founder of Olakai. Your AI is an investment, so let’s measure it like one.

On October 1, IBM announced that Bob, its agentic software-development platform, is now generally available as a self-hosted deployment. The release lists the environments it now runs in, on-premises, private cloud, sovereign cloud and air-gapped, and names what stays inside the customer’s own infrastructure: “sensitive source code, application context, regulated customer information, and intellectual property.” Models can be ones the customer has licensed, or external model services reached through what IBM calls hybrid configurations. Neel Sundaresan, IBM’s General Manager of AI and Automation, put the reasoning in one sentence: “Organizations need AI that operates inside environments they have control over, especially when working with sensitive code and regulated data.”

Read that again with the vendor’s name removed and it is a statement about architecture, not about IBM. The largest enterprise software company in the world has decided that an AI coding agent belongs where the code lives, and has built the product to match. For anyone who read last week’s piece on why we built an on-prem version of Olakai, this is the same argument arriving one layer down the stack, from a company whose customers are the banks, insurers and governments that make the argument for a living.

Why IBM did it

IBM did not have to guess what its buyers wanted. In June its Institute for Business Value published a study of 1,000 senior executives across 16 countries, surveyed between February and April, and the numbers describe an enterprise that feels it has lost control of its own AI estate. Sixty-eight percent said meeting data-residency and sovereignty requirements across geographies is challenging. Seventy-one percent said switching their primary AI vendor or model would be difficult. Ninety-one percent admitted they do not fully understand their AI dependencies across vendors, models and infrastructure. Only 7% operate at what IBM calls an advanced level of AI control, and that small group protects 55% more operating profit from AI-driven disruption than everyone else. Perhaps most telling, 72% said they would accept a 20% cost increase in exchange for more strategic flexibility.

Ana Paula Assis, IBM’s Senior Vice President and Chair for EMEA and APAC, summarized the study this way: “AI has introduced new forms of dependency that evolve faster than traditional governance, procurement, or technology cycles were designed to handle.” Self-hosted Bob is IBM’s product answer to its own research, and it lands in a market that was already moving. The repatriation and sovereign-cloud data we covered last week, from Cloudian, Broadcom and Gartner, point the same way: the buyer has decided that the model, the data and the tooling will sit inside a boundary they control, and vendors can either meet them there or lose the deal.

The layer IBM’s release does not mention

Here is what the announcement describes: where the agent runs, what data it keeps inside the perimeter, which models it can call. Here is what it does not describe: how an enterprise will know whether that agent is earning its cost, what it touched, which developers are using it well, and how its output compares with the three other coding tools the same engineering organization is almost certainly running. That is not a criticism of IBM. A coding agent’s job is to write software, and measurement and governance belong to a different layer. But it is the layer that decides whether the deployment was worth it, and it has a location too.

Follow the data path. An enterprise moves its coding agent on-prem so that source code, application context and regulated customer information never leave. Then it adopts a SaaS tool to measure that agent’s productivity and govern what it does. That tool ingests pull requests, repository metadata, developer identities, spend, and in many cases the prompts and sessions themselves, and sends all of it to a vendor’s multi-tenant cloud. The code stayed home. The record of everything that happened to the code went back out the door. Every argument that justified self-hosting the agent applies with at least equal force to the system that watches the agent, because that system holds a more complete picture of the engineering organization than any single tool ever will.

This is why Olakai Enterprise runs the same platform on-prem that it runs as SaaS: one Linux VM you own, your choice of analysis model, and nothing leaving your network when you route that analysis to a server inside it. A regulated enterprise that has just brought its coding agent inside the boundary should expect the measurement and governance layer to come with it, under the same deployment review and the same network allowlist, rather than being the one AI tool in the stack that is still phoning home.

What this means for a multi-vendor tool stack

Two more numbers from the IBM study matter here. Seventy-three percent of executives describe their AI environment as intentionally multi-vendor, and 71% say switching their primary vendor would be hard. Both are true at once in most engineering organizations we see: Cursor on one team, GitHub Copilot on another, Claude Code for the platform group, and now perhaps Bob for the modernization program that touches the mainframe. Nobody is going to standardize on one, and nobody can easily leave any of them. The only way to compare them fairly is to measure from a vantage point none of them controls.

That vantage point is the repository. Olakai Agentic detects AI-assisted work from pull-request data in your own version control: bot authors, co-author trailers and PR markers, with a classifier reading the gray-area pull requests that carry no machine-readable signature. Because the signal comes from your repositories rather than from any tool’s own telemetry, it does not matter whether the agent that wrote the code ran in a vendor’s cloud or on a server in your basement, and it does not matter when you swap one agent for another. The spend, adoption cohorts and cycle-time deltas are computed against the same baseline for every tool. That is what a vendor-neutral AI system of record is for: it is the one part of the stack that is supposed to outlive every vendor decision, which is exactly the flexibility 72% of IBM’s respondents said they would pay a premium for.

What leaders should do this quarter

For a VP of Engineering or CISO evaluating self-hosted Bob, or any coding agent moving inside the perimeter, the practical rule is that governance telemetry follows the workload. If the agent passes a deployment review, a network allowlist and a data-classification check, the system that measures and governs it should pass the same three, in the same environment, before either goes live. Ask every vendor in the stack the question IBM’s buyers evidently asked IBM: where does our data go, and can we make the answer “nowhere”?

For a CFO, the IBM study’s most useful line is the 7% who protect 55% more operating profit through control of their AI stack. Control is not a security slogan; it shows up in the P&L. Insist that the ROI evidence for every AI coding investment is computed where the data already sits, by a system that does not depend on the vendor being measured, so the number survives both the audit and the next vendor change. Paul made the cost half of this case in July, when IBM’s own AI bill was the story; the sovereignty half is what October added.

IBM has now said out loud what regulated buyers have been saying in procurement for two years: the agent runs where the code lives. The organizations that get this right will govern and measure AI from inside the same walls, with Olakai or something like it. The ones that do not will eventually find that the single AI tool still sending data out of the building is the one they bought to keep an eye on all the others.

Bringing a coding agent inside your perimeter? Talk to an expert about measuring and governing it from the same place.

From the AI ROI Series, recorded 9 September 2026. Two stories broke this week that most people filed under separate headlines: OpenAI’s Astra, and Anthropic walking away from a six billion dollar acquisition. I want to make the case that they are the same story, and that the story is about the one number that decides who wins in enterprise AI.

Anthropic was reportedly ready to pay up to $6 billion for Decart, a lab whose optimisation engine runs agents at about eight times the industry average. Then, after full due diligence and right before its IPO, it reportedly walked. Both halves of that are reported rather than confirmed, so hold them accordingly.

The number underneath both headlines

The mechanism before the arithmetic, because the phrase doing the work here is one most buyers never see on an invoice. Cost to serve is what it costs a vendor to answer your request, the compute burned turning your tokens into their output. It sits underneath the rate card you are quoted, and it separates a model business that compounds from one that merely grows. An engine running agents at eight times the throughput moves that number directly, which is why it is worth billions to somebody who serves inference for a living.

Anthropic’s own cost to serve, halved in a year

A year ago Anthropic’s margin on serving inference was around 38%. Today it is around 70%. In plain dollars, running the AI used to eat about 62 cents of every revenue dollar and now it eats about 30, which is the same fact said twice rather than two separate findings. They cut their own cost to serve nearly in half in a single year. These are reported figures, so hold them loosely, and note what they cover: this is the margin on serving inference, before training, research, and staff, and it is not net profit.

That qualifier matters if you read what I wrote about Anthropic’s valuation a fortnight ago, where the figure in play was a company-wide gross margin of around 44%. The two are different measures and they sit comfortably together, since gross margin carries a great deal of cost that serving a token does not. The direction of travel is the interesting part. My argument then was that every enterprise pushing work down the price curve takes a point of the vendor’s margin. Here is the same vendor defending that margin from the other side by making each token cheaper to serve, one variable squeezed from both ends.

Why they walked, three honest reads

So you can see why an engine that gives you eight times the throughput is worth billions, and you can also see why they might still walk. There are three honest reads, and they do not contradict each other.

1. They already proved they can cut the cost themselves, so why pay $6 billion for more of a lever they are good at pulling?
2. The engine comes bundled with a whole video business, which is not their strategy, and you cannot cleanly buy just the meter.
3. Walking away from your biggest deal ever, in the week the market started grading on return rather than spend, is a discipline signal, and that is good finance management.

Pick whichever you like. What survives all three is the reason Astra and Decart are the same story. The model on your desk will keep flipping, OpenAI this quarter, Anthropic the next, somebody else after that. But whoever is winning the capability race, every one of them is measuring its cost to serve down to the cent, and one of them ran $6 billion of diligence to move it. The capability race is loud. The economics race is silent, and it is permanent.

Now turn it around, because you live on the other side

You are going to switch model this year based on who is best, and most enterprises will. But can you tell me what any of them actually cost you per outcome, per task? I am not asking about invoice totals. What did one completed task, one shipped feature, one resolved ticket cost you in AI, and what did it give back? For almost everyone I talk to, the answer is no.

And it is not because the teams are not sharp. It is because that data was never captured as a record in the first place. It sits scattered across four vendor consoles that do not talk to each other and were not built to tell you very much, so you cannot even see where to spend less, and none of them know what that token was actually for. This is the same gap that lets a falling rate card sit next to a rising bill, and the same one that made a vendor’s pricing change land under an agent budget without anyone noticing for a week.

Same tokens, two lenses, and they almost never meet

The dilemma is the same one whichever chair you sit in, and it splits cleanly down the middle of most organisations. If you are in finance, you watch the invoice climb and you cannot say whether that is a problem or a good investment. If you are in engineering, you watch the workload climb and you cannot put a dollar on it. It all runs on the same tokens. They are completely different lenses on one number, and the two almost never meet in one place, which is how you end up with two teams arguing from two screens and two spreadsheets.

I do see a shift here that I like a great deal, with people building their own dashboards and their own small solutions, and some of it is genuinely good work. There is always a story about somebody saving 20 minutes on a task. It is real, and it might even be statistically significant, but a pointy saved minute is a long way from a firm-level answer, let alone a board-level one, especially once you are past the pilot and into scale.

Put the two sides together and the asymmetry is the whole point. The seller measures every token to the cent, across every model, and will spend $6 billion of diligence to move the number by a few points. The buyer measures a good afternoon on whichever model is fashionable this quarter.

That gap is the reason Olakai exists. We capture every AI interaction and every outcome across an organisation, down to the token, structure it into one attributed record, and make it usable through your own AI, so you can see cost per completed task and value per outcome by team, by agent, by model, and by vendor. It is one record read through whichever lens is yours: finance reads it as return and budget, engineering reads it as throughput and the cost of what shipped. Data is the product. The record is the instrument.

FY27 budget season, and the questions that matter

The model on your desk is going to keep changing. The one thing that should not change is your ability to measure what it is worth. We are all getting the same question this year whether we like it or not, which is what did it return, and most of us cannot answer it cleanly yet. That is a measurement gap rather than a failing on anyone’s part, and measurement gaps are fixable.

So these are the table stakes questions for this year, and they are for anyone who owns a piece of the AI budget, which means finance, engineering, and AI leaders alike.

1. When you switch to Astra, or to whatever comes next, will you know whether it costs you more or less per outcome than the model it replaced?
2. Did it make your agents better, and can you show the difference?
3. If you are in finance, could you put that number in front of the board on Monday with the evidence behind it?
4. If you are in engineering, could you show which agents and which models earned their cost, and which did not?

If any of those is a no, that is the work.

Directional as always. The data is public, so check my math and tell me if you see it differently, because I welcome that all day long. I’m Paul, co-founder of Olakai. Your AI is an investment, so let’s measure it like one.

From Enterprise AI Weekly, recorded 28 August 2026. Three things happened in enterprise AI that week, and read together they settle an argument the industry has been having for about two years.

Nvidia reported $96.2 billion of revenue for the quarter, up 106% year on year, guided to 70% growth for next year, and said it is still supply-constrained. That last part is the interesting one. A company can sell $96.2B of anything in ninety days and still tell the market it cannot build fast enough, which puts the constraint somewhere physical, in fabs and power and memory, rather than anywhere a purchase order can reach.

Almost all of the coverage took these as three separate stories, filed to three separate desks: a semiconductor story, a CRM story, and a valuation story. I look at them from where I sit, which is with the people who have to put a number in a 2027 AI budget and then defend it, and from that seat they are quite obviously one.

Capability stopped being the interesting variable

Before the arithmetic, the mechanism, because most people still carry the wrong mental model here. The assumption underneath a great deal of enterprise AI procurement is that the frontier models are separated by large capability gaps, and that paying more buys you proportionally more machine. On the current numbers that assumption no longer holds. Ten frontier models now sit about 16 points apart on capability and are priced about 12 times apart, according to the Artificial Analysis leaderboard.

Put those two spreads next to each other and the shape of the decision changes. A 12x price range across a 16 point capability range means the premium at the top of the menu is buying a modest amount of measurable capability at a very immodest multiple, and it means the model you pick matters far less to your outcome than the discipline with which you route work to it. I have written the arithmetic on that routing decision at length before, and none of the week’s news changes it. If anything the convergence makes the case stronger, because the cheaper end of the ladder keeps closing the gap while the expensive end keeps its price.

Compute is being funded by the trillion

The second variable is the one being solved with capital rather than with cleverness. Nvidia cannot build fast enough at $96.2B a quarter, and Anthropic is reportedly targeting a $2 trillion IPO, per the Financial Times. Whatever you think of that number, and I have written separately about what that kind of raise signals to buyers, the direction of travel is not ambiguous. The compute layer is being financed at a scale that no individual enterprise participates in and no individual enterprise needs to, because you rent it by the token. It is a commodity with a public rate card, and what you buy when you sign an AI contract is a place in a queue somebody else is paying to lengthen on your behalf.

Then Salesforce put Claude on your CRM data

The third story is the one that ties the other two off. Salesforce announced ClaudeForce, putting Claude directly onto customer CRM data, and the market liked it: the stock rose about 22% on the day and about 40% across recent weeks, on a strong quarter and the Anthropic partnership, per The Motley Fool. The line I would keep from Salesforce’s own announcement is this one, because it is a vendor conceding something vendors do not usually concede:

Probabilistic intelligence alone doesn’t run a company.

The reasoning has to be fused with trusted data, with workflows, and with governance. Which is to say that the model on its own, however capable, is an input rather than an outcome, and the thing that turns it into an outcome is the material you point it at and the controls you wrap around it.

The one input with no market

Here is the comparison that reorganised how I think about 2027 budgets. Capability is converging and is available to your competitors on the same terms it is available to you. Compute is a commodity, priced publicly, and financed by people with more capital than any of us. Your own data, and the record of what your own AI has actually done with it, has no market at all. Nobody sells it, nobody can lend it to you, and there is no procurement cycle that gets you it faster. You either have it or you spend the next two years building it.

That asymmetry is the whole reason measurement stopped being a reporting exercise and became an asset question. If two companies rent the same models at the same prices, the difference between them is what each can see about its own usage, its own outcomes, and its own costs. This is the missing piece, and it is a system of record for enterprise AI rather than another dashboard. A dashboard shows you a view of something. A record is the thing itself, kept over time, in one data model, across coding tools, assistants, and autonomous agents alike.

What to check before you write the 2027 number

Directional as always, and check my math. But if the three stories above are one story, then a few things follow for anyone building next year’s budget, and they are checks rather than recommendations. Can you say, without a project, which models your organisation is currently paying for and what each one produced? Can you attribute AI spend to a team, a workload, and an outcome, rather than to a vendor invoice? If the capability gap between the cheapest capable model and your default model is genuinely 16 points, do you know what your own workload loses by moving down the ladder, or are you paying the 12x premium as insurance against an uncertainty you have never measured?

Those are answerable from your last quarter, and the answers are usually more uncomfortable than the headlines are. Most organisations find they are carrying the premium and cannot say what it bought, which is the same thing I keep running into when I look at AI ROI properly, and it is why the measurement layer is the part I would fund first.

So the question I would put to you, and you can answer it from what you already know: if every model your competitors use is available to you at the same price tomorrow morning, what is left that is actually yours?

I’m Paul, co-founder of Olakai. Measuring what AI actually costs and what it actually returns, on your own workload, is the work I spend my days on. Tell me if you see it differently, and if you would rather see it than argue about it, your AI is an investment, so let’s measure it like one.