Category: AI Strategy

The CIO of a mid-market financial services firm thought she had a handle on AI adoption. Her team had sanctioned three tools, trained 200 employees, and built a governance policy around them. Then she ran an AI visibility audit. The audit found 23 AI tools running across the organization — seven times what she expected. Customer service had adopted a chatbot through a free trial. Marketing was using three different content generators. Two engineering teams were running code assistants that had never been security-reviewed. And an entire business unit had been piping client data through an AI summarization tool that stored data on external servers.

She’s not unusual. According to the Torii 2026 Benchmark Report, 84% of organizations consistently discover more AI tools than expected during audits. And 31% find new unsanctioned tools every single month.

This is the SEE problem — the first and most foundational step in the SEE, MEASURE, DECIDE, ACT framework for proving AI ROI. (This is the first of four companion deep-dives — see also MEASURE, DECIDE, and ACT.). You cannot measure what you cannot see. And in most enterprises today, the AI landscape is far larger, more fragmented, and more exposed than anyone in the C-suite realizes.

The Visibility Crisis by the Numbers

The scale of unsanctioned AI usage has grown faster than most security and IT teams anticipated. A 2025 UpGuard study found that more than 80% of workers — including nearly 90% of security professionals — use unapproved AI tools on the job. That last part bears repeating: the people responsible for protecting the organization are themselves using tools that haven’t been vetted.

Deloitte’s 2026 State of AI survey tells the supply side of this story. Workforce access to AI tools expanded by 50% in a single year, from fewer than 40% of employees to roughly 60%. But that figure only counts sanctioned tools. The actual adoption rate — including shadow AI — is far higher. Research from Portal26 found that 73.8% of ChatGPT accounts used in the workplace are non-corporate accounts that lack enterprise security and privacy controls. For Gemini, that figure is 94.4%.

The result is an AI ecosystem that leadership cannot see, security cannot govern, and finance cannot account for. Only 38% of organizations report knowing which AI applications their employees actually use.

What Invisibility Actually Costs

The cost of this visibility gap isn’t hypothetical. IBM’s 2025 Cost of a Data Breach report found that breaches involving shadow AI add $670,000 to the average breach cost compared to organizations with low or no shadow AI exposure. The average organization now experiences 223 AI-related data security incidents per month — incidents that range from sensitive data shared with external AI services to policy violations that create compliance exposure.

But security costs are only one dimension. Hitachi Vantara research estimates that data infrastructure issues — many driven by ungoverned AI tooling — contribute to $108 billion in wasted annual AI spend across enterprises. When teams adopt AI tools independently, they duplicate capabilities, fragment data flows, and create redundant infrastructure costs that nobody tracks because nobody can see the full picture.

Then there’s the opportunity cost. If you don’t know what AI your organization is running, you cannot measure whether it’s working. You cannot identify which tools deliver value and which ones burn budget. You cannot rationalize spending, consolidate licenses, or negotiate enterprise agreements. And you cannot answer the one question the board increasingly cares about — what’s the return on our AI investment — because you don’t even know what the investment includes.

Why Traditional Discovery Fails

Most IT organizations approach AI discovery the same way they approach software asset management: check the procurement records, run a network scan, send out a survey. None of these methods work for AI.

Procurement records miss AI tools that employees adopt through free tiers, browser extensions, or personal accounts. Network scans miss browser-based AI tools that look like regular web traffic. Surveys depend on employees self-reporting usage they may not think of as “AI” — or usage they know isn’t sanctioned and don’t want to disclose.

The deeper problem is velocity. Employees adopt new AI tools faster than security teams can evaluate them. Eighty-three percent of organizations report that employees install AI tools faster than security can track, according to industry surveys. A quarterly discovery audit is fundamentally mismatched against a weekly adoption cycle.

And the challenge is getting more complex, not simpler. Embedded AI features — AI capabilities built into tools employees already use, like email clients, CRM platforms, and productivity suites — fly under the radar entirely. An employee isn’t “adopting a new AI tool” when their email client adds AI-powered reply suggestions. But the data exposure risk is real, and the cost shows up in per-seat licensing increases that finance sees but can’t attribute.

What a Real AI Visibility Audit Looks Like

A proper AI visibility audit goes beyond inventory. It answers four questions that are prerequisites to everything else in the AI ROI playbook:

What AI is running? A complete catalog of AI tools, models, and capabilities across the organization — including assistive AI (copilots, chatbots, content generators), agentic AI (autonomous agents executing workflows), and embedded AI (features within existing software). This isn’t a one-time list. It’s a continuously updated inventory that captures new tools as they appear.

Who is using it? Usage patterns by team, department, role, and individual. Not to police employees, but to understand where AI adoption is concentrated, where training gaps exist, and where usage patterns suggest risk or opportunity. If 60% of your customer success team uses an AI tool daily but 5% of your sales team does, that’s a signal worth understanding.

What data is it touching? The critical question from both a security and compliance perspective. Which AI tools have access to customer data, financial records, intellectual property, or regulated information? Are employees sharing sensitive data with external AI services? The shadow AI risk isn’t just that unauthorized tools exist — it’s that unauthorized tools often handle the most sensitive data, because employees turn to AI precisely when they’re working with complex, high-value information.

What is it costing? The total cost of AI across the organization, including sanctioned licenses, API consumption, infrastructure, and the hidden costs of shadow AI — duplicate tools, wasted capacity, and the remediation costs when things go wrong. Until you can see the full cost picture, you cannot calculate ROI.

From Visibility to Value

The SEE step isn’t an end in itself. It’s the foundation that makes everything else possible. Once you have visibility into your AI ecosystem, you can move to MEASURE — connecting AI activity to business outcomes. You can identify which tools are delivering value and which are creating risk. You can rationalize spending, consolidate tooling, and negotiate from a position of knowledge rather than ignorance.

The enterprises that close the AI revenue gap — the 20% who prove AI drives results, according to Deloitte’s 2026 survey — start here. Not with measurement. Not with governance. With visibility. Because every dollar of AI ROI you can prove is built on a foundation of knowing what AI you have, who’s using it, what data it touches, and what it costs.

The visibility audit typically reveals three immediate value opportunities: tool consolidation (reducing redundant AI spending by 20-30%), risk reduction (identifying unvetted tools handling sensitive data), and measurement readiness (instrumenting high-value AI workflows for ROI tracking). Most enterprises find that the audit pays for itself through spend rationalization alone.

Ready to see what AI is actually running across your organization? Talk to an expert and we’ll show you how Olakai provides unified visibility across your entire AI ecosystem — sanctioned and shadow, assistive and agentic.

Half of CEOs believe their jobs are on the line if AI doesn’t pay off. Yet according to BCG’s AI Radar 2026 survey, 90% of chief executives believe agentic AI will deliver measurable ROI this year. That’s a remarkable level of conviction given what the data actually shows: IBM found that only 29% of executives can confidently measure their AI returns, and just 16% have scaled AI initiatives enterprise-wide.

The confidence is there. The measurement capability is not. And that gap — between what leaders believe AI can do and what they can prove it has done — is where budgets get cut, pilots stall, and competitors pull ahead.

This is why we built the SEE, MEASURE, DECIDE, ACT playbook — a four-step framework that takes enterprises from “we think AI is working” to “here’s exactly what it’s worth.” It’s the same methodology we use with every enterprise we work with, and the same framework that separates the 20% of organizations seeing real revenue impact from AI from the 74% who want it but can’t prove it.

The Playbook Gap

Deloitte’s 2026 State of AI survey captured the problem in a single data point: 74% of enterprises say they want AI to drive revenue growth. Only 20% have achieved it. That’s 3,235 business leaders across 24 countries essentially saying the same thing — we’re investing heavily, but we can’t connect the investment to results.

The issue isn’t the technology. AI models are more capable than ever. The issue is that most enterprises lack a systematic approach to proving value. They launch pilots without defining what success looks like. They measure activity (tokens processed, queries handled) instead of outcomes (revenue influenced, costs avoided). And when the CFO asks “what’s our return?”, the answer is a shrug wrapped in a slide deck full of usage charts.

BCG found that companies plan to double their AI spending in 2026, pushing AI investment to roughly 1.7% of total revenues. CEOs are committing more than 30% of their AI budgets specifically to agentic AI. The money is flowing. But without a measurement playbook, most of it flows into a black box.

Step 1: SEE — Map Your AI Ecosystem

You can’t measure what you can’t see. And in most enterprises, the AI landscape is far more sprawling than leadership realizes.

Workforce access to AI tools expanded by 50% in just one year, according to Deloitte — from fewer than 40% of workers to roughly 60% now equipped with sanctioned AI tools. That’s just the sanctioned ones. Factor in the tools employees adopt on their own — the shadow AI that bypasses procurement and IT review — and the real number is significantly higher.

The SEE step is an AI visibility audit. It answers three questions: What AI tools and models are running across the organization? Who is using them? And what data are they touching? This isn’t a one-time inventory. It’s an ongoing discovery process, because AI adoption in enterprises is a moving target — new tools appear weekly, usage patterns shift monthly, and the risk surface evolves with every new integration.

Most enterprises discover during this step that they have three to five times more AI touchpoints than they thought. Customer service teams running chatbots that marketing doesn’t know about. Engineering teams experimenting with code assistants that security hasn’t reviewed. Sales teams piping prospect data through AI tools that legal hasn’t vetted. Until you see the full picture, every other step in this playbook is built on incomplete information.

Step 2: MEASURE — Connect Activity to Business Outcomes

Once you can see what’s running, the next step is measuring what matters. And “what matters” is almost never what teams measure first.

The natural instinct is to track operational metrics: response time, tokens consumed, uptime, error rates. These are useful for engineering but meaningless to the CFO. The measurement step connects AI activity to the business KPIs that drive budget decisions — revenue influenced, costs reduced, risk mitigated, time recovered.

This is where most enterprises stall. IBM’s research found that while 79% of organizations see productivity gains from AI, only 29% can measure ROI confidently. The productivity is real but unquantified. A customer success agent saves each rep 45 minutes per day — but nobody has connected that time savings to the additional accounts each rep can now manage, or the churn reduction that comes from faster response times.

Effective AI measurement requires three elements. First, a baseline: what was the metric before AI? Without a counterfactual, you’re reporting output, not impact. Second, attribution: which portion of the improvement is actually due to AI versus other factors? Third, a time horizon that matches the business cycle. An AI agent that qualifies leads doesn’t show revenue impact in week one. It shows impact when those leads close, which in enterprise B2B might be 90 days later.

The 20% of enterprises that prove AI revenue impact aren’t using more sophisticated models. They’re using more sophisticated measurement. They define the success KPI before deployment, not after. They instrument their AI systems to capture business outcomes, not just technical telemetry. And they present results in the language the CFO speaks — dollars, not tokens.

Step 3: DECIDE — Turn Data Into Scaling Decisions

Measurement without decision-making is just reporting. The DECIDE step uses the data from MEASURE to answer the questions that actually move AI forward in an organization: Which pilots get promoted to production? Which get sunset? Where should the next investment go?

This is where the 30-to-45-day structured pilot becomes critical. Rather than running open-ended experiments that drift for months, a time-boxed pilot with predefined KPIs produces a clear decision point. At the end of 30 days, you have data. Not opinions, not anecdotes — data that shows whether the AI investment is generating the business outcome you defined in the MEASURE step.

The enterprises stuck in pilot purgatory almost always lack this decision framework. They have pilots running for six, nine, twelve months with no clear criteria for what constitutes success or failure. The result is the worst possible outcome: continued investment without conviction, where the AI initiative is too expensive to ignore and too poorly measured to champion.

A proper DECIDE framework answers four questions with data: Is the AI system delivering the outcome KPI we defined? Is the cost-to-value ratio favorable? Can the governance and risk profile support scaling? And does the organization have the operational readiness to absorb the change?

Google Cloud’s research found that top-performing enterprises generate $10.30 in value for every dollar invested in AI, while the average is $3.70. The difference isn’t luck. It’s disciplined decision-making about which investments to scale and which to cut — and that discipline is only possible with measurement data.

Step 4: ACT — Scale With Confidence

The final step is where measurement pays off: scaling the AI investments that prove their value while governing the entire portfolio continuously.

Deloitte found that 25% of organizations now report AI having a “transformative” effect — up from just 12% a year ago. These are the enterprises that have moved through SEE, MEASURE, and DECIDE, and are now deploying AI at scale with the data to back every decision. They’re not guessing which use cases deserve investment. They know, because they measured.

But scaling introduces new challenges that require continuous measurement. An AI agent that performs well with 100 users may behave differently with 10,000. Cost structures change at scale. Risk profiles shift as AI touches more sensitive data and higher-stakes decisions. The ACT step isn’t a one-time event — it’s an ongoing cycle of deploying, measuring, governing, and optimizing.

This is where governance and measurement converge. The enterprises with the strongest ROI data are also the ones with the most rigorous governance frameworks. Not because governance is a checkbox exercise, but because governance forces the discipline that measurement requires: defining what AI is allowed to do, instrumenting how it performs, and maintaining the accountability structures that ensure continuous improvement.

BCG reports that 72% of CEOs are now the primary decision-makers on AI, double the share from a year ago. These executives don’t want dashboards full of technical metrics. They want a portfolio view: which AI investments are generating returns, which ones need intervention, and where the next opportunity lies. The SEE, MEASURE, DECIDE, ACT framework gives them exactly that.

Building Your Playbook

The 74-to-20 gap Deloitte identified isn’t permanent. But it won’t close on its own. It closes when enterprises stop treating AI measurement as an afterthought and start treating it as the foundation of every AI initiative.

Start with SEE: audit your AI ecosystem. You’ll likely find more than you expected. Move to MEASURE: define the business outcomes that matter and instrument your AI systems to capture them. Progress to DECIDE: use 30-day structured pilots to generate decision-quality data. And then ACT: scale what works, govern what runs, and keep measuring.

The enterprises in the 20% didn’t get there with better AI. They got there with better measurement. The playbook isn’t complicated. The hard part is committing to it before the CFO asks the question you can’t answer. Our AI ROI measurement framework breaks down the methodology step by step, and Future of Agentic’s KPI library offers specific metrics by use case to get you started.

Ready to build your AI ROI playbook? Talk to an expert and we’ll show you how enterprises are turning AI activity into measurable business outcomes.

When JPMorgan Chase launched its LLM Suite platform in summer 2024, something unusual happened: within eight months, 200,000 employees were using it daily. No mandate. No compliance requirement. Just organic adoption at a scale that most enterprises can only dream about.

Meanwhile, at most other organizations, a very different story was playing out. MIT’s 2025 “GenAI Divide” report, based on 150 executive interviews and 300 public AI deployments, found that 95% of generative AI pilots fail to deliver rapid revenue acceleration. Not 50%. Not even 80%. Ninety-five percent.

The gap between JPMorgan and everyone else isn’t about technology, talent, or even budget. It’s about something far more fundamental: whether you can prove your AI is working.

The Measurement Gap Is the Real Pilot Killer

Enterprise AI has an accountability problem. Organizations are spending aggressively — global generative AI investment tripled to roughly $37 billion in 2025 — but most cannot answer a simple question: What’s the ROI on our AI?

The numbers tell a stark story. McKinsey’s State of AI 2025 report found that 88% of organizations now use AI regularly in at least one business function. Yet only 6% qualify as “AI high performers” who can attribute more than 5% of total EBIT to AI. The other 82% are running AI, but they cannot connect it to business results.

Deloitte’s State of AI in the Enterprise 2026 survey — covering 3,000 director-to-C-suite leaders across 24 countries — revealed what might be the most telling statistic of all: 74% of organizations want AI to grow revenue, but only 20% have actually seen it happen. That’s not a technology gap. That’s a measurement gap.

Why “Pilot Purgatory” Is Getting Worse, Not Better

You might expect the pilot-to-production problem to improve as AI matures. It’s not. S&P Global data shows that 42% of companies abandoned most of their AI initiatives in 2025, more than double the 17% abandonment rate just one year earlier. The average enterprise scrapped 46% of AI pilots before they ever reached production — a pattern we first explored in From AI Experimentation to Business Impact. For every 33 prototypes built, only 4 made it into production — an 88% failure rate at the scaling stage.

The pattern is consistent: organizations launch pilots with enthusiasm, run them for three to six months, then struggle to justify continued investment. Without baseline metrics established before deployment, there’s no way to quantify what AI actually changed. Our AI ROI measurement framework provides the methodology for establishing those baselines and tracking outcomes. Without ongoing measurement, there’s no way to distinguish a successful pilot from an expensive experiment. And without clear ROI data, there’s no executive willing to sign off on scaling.

Gartner reinforced this trajectory in June 2025, predicting that over 40% of agentic AI projects will be canceled by the end of 2027, citing three drivers: escalating costs, unclear business value, and inadequate risk controls. The emphasis on “unclear business value” is telling — it’s not that the AI doesn’t work, it’s that nobody built the infrastructure to prove that it does.

What the 5% Do Differently

The companies that successfully move AI from pilot to production share a pattern that has nothing to do with having better models or bigger datasets. They build measurement into the process from day one.

JPMorgan didn’t just deploy AI — they tracked adoption rates, time savings, and productivity gains from the first week. Their AI benefits are growing 30-40% annually, and they know this because they measure it. Walmart didn’t just experiment with AI in their supply chain — they documented that route optimization eliminated 30 million unnecessary delivery miles and avoided 94 million pounds of CO2 emissions. Their customer service AI cut problem resolution times by 40%, a number they can report because they established baselines before deployment.

This is the pattern MIT’s research confirmed across hundreds of deployments: the companies that scale AI successfully are the ones that treat measurement as infrastructure, not an afterthought. They know which processes AI is accelerating, by how much, and at what cost. They can calculate the total cost of ownership — including the API costs, engineering time, and maintenance burden that most organizations bury in IT budgets. And they can present executives with a clear picture: here’s what AI costs, here’s what it delivers, and here’s why scaling it makes financial sense.

The Four Phases of Scaling (and Where Most Organizations Get Stuck)

Successfully moving AI from pilot to production typically follows four phases, each gated by measurement milestones rather than arbitrary timelines.

Phase 1: Validate value (weeks 1-4). Deploy the AI solution with a small group and establish clear baselines. What does the process look like without AI? How long does it take? What does it cost? What’s the error rate? Without these pre-AI measurements, you’ll never be able to quantify impact. Most organizations skip this step entirely and then wonder why they can’t prove ROI six months later.

Phase 2: Harden for production (weeks 5-10). Once you have evidence that the AI delivers measurable value, build the governance and monitoring infrastructure needed for scale. This means policy enforcement, access controls, audit trails, and cost tracking. It also means ensuring someone owns ongoing operations — not as a side project, but as a defined responsibility.

Phase 3: Controlled expansion (weeks 11-16). Roll out to a broader group while continuing to measure. Are the gains from Phase 1 holding at scale? Are costs scaling linearly or exponentially? Are new user segments finding different use cases? This phase is where many organizations discover that their pilot’s curated dataset doesn’t translate to messy real-world data — Gartner found that data quality issues derail 85% of AI projects at this stage.

Phase 4: Full deployment and continuous optimization. With validated ROI data from the first three phases, you have the evidence to justify enterprise-wide investment. But the measurement doesn’t stop — it shifts from proving value to optimizing it. Which teams are getting the most benefit? Where are costs disproportionate to returns? What new use cases are emerging?

The organizations that stall are almost always stuck between Phase 1 and Phase 2. They ran a pilot, it “seemed to work,” but they never established the baselines or tracking needed to prove it. So the pilot sits in limbo — too promising to kill, too unproven to scale.

Buy vs. Build: A Measurement Shortcut

MIT’s research uncovered a surprising finding about the build-versus-buy decision. Purchasing AI tools from specialized vendors and building partnerships succeeds roughly 67% of the time, while internal builds succeed only about 22% of the time. Our analysis of 100+ AI agent deployments confirms this pattern. The gap is striking, and measurement is a significant part of the explanation.

Specialized vendors have already solved the measurement problem for their specific domain. They’ve established the benchmarks, built the tracking, and validated the ROI across hundreds of customers. When an enterprise buys rather than builds, they’re importing not just the technology but the measurement framework that proves it works.

Internal builds, by contrast, require organizations to solve two problems simultaneously: making the AI work and building the infrastructure to prove it works. Most teams focus entirely on the first problem and neglect the second.

From Science Experiment to Business Case

Harvard Business Review captured the core challenge in November 2025: “Most AI initiatives fail not because the models are weak, but because organizations aren’t built to sustain them.” Their five-part framework for scaling AI emphasizes that the bottleneck is organizational, not technical — and at the center of every organizational bottleneck is the inability to prove value.

The path from pilot to production isn’t about better technology. It’s about building the measurement infrastructure that turns an AI experiment into a business case. That means establishing baselines before deployment, tracking outcomes continuously, calculating total cost of ownership honestly, and presenting results in terms executives care about: revenue impact, cost reduction, risk mitigation, and time to value.

Without that measurement layer, every AI pilot is a science experiment. And enterprises don’t scale science experiments — they scale proven investments.

Ready to move your AI from pilot to production? Talk to an expert to see how Olakai helps enterprises measure AI ROI, govern risk, and scale what works across every AI tool and team.

Conversational AI is projected to save $80 billion in contact center labor costs by 2026. That number is staggering — but it also tells a narrow story. Most enterprises still think of voice AI as a call deflection tool: something that answers the phone so a human doesn’t have to. That framing misses what’s actually happening.

Voice AI has quietly become one of the most versatile technologies in the enterprise stack. It’s writing medical notes in real time. It’s scoring sales calls for sentiment and coaching reps mid-conversation. It’s authenticating banking customers by analyzing 100 vocal traits in under a second. And the economics are compelling: companies implementing voice AI in customer support are seeing 68% reductions in cost per interaction, from $4.60 to $1.45 on average, with leading organizations reporting ROI as high as 8x their initial investment.

The question for enterprise leaders isn’t whether voice AI works — it’s whether they can measure, govern, and scale it responsibly across every department that’s already experimenting with it.

The Accuracy Turning Point

For years, accuracy held voice AI back. Anyone who has shouted “REPRESENTATIVE” into a phone tree understands the frustration. But 2025 marked a genuine inflection point. Word error rates in noisy environments — the kind you’d encounter in a hospital, a factory floor, or a busy sales bullpen — dropped from over 40% to near zero. Recognition of non-native accents improved from 35% WER to 15%. Multi-speaker scenarios went from “largely unusable” at 65% WER to “practically viable” at 25%.

These aren’t incremental improvements. They’re the difference between a technology that frustrates users and one that earns their trust. Healthcare saw it first: specialized speech models now produce 70% fewer transcription errors in clinical workflows, according to Stanford Medicine research. Meanwhile, latency has dropped to the natural conversational rhythm of 500 milliseconds — fast enough that talking to an AI agent no longer feels like talking to a machine.

This accuracy revolution explains why 80% of businesses plan to integrate AI-driven voice technology into customer service by 2026, and why the voice AI agent market is on track to grow from $2.4 billion to $47.5 billion over the next decade.

Beyond the Call Center

The real story of enterprise voice AI isn’t about replacing call center agents. It’s about what happens when voice becomes a data layer across your organization.

In healthcare, ambient listening technology is quietly transforming clinical documentation. AI scribe systems listen to patient-provider conversations and automatically generate structured SOAP notes that sync directly with electronic health records. A 2025 study published in JAMA Network Open found that clinicians using ambient AI documentation reported self-reported burnout dropping from 42% to 35%, spent less time writing notes both during and after appointments, and — crucially — felt they could actually listen to their patients. Microsoft’s Dragon Copilot, launched in March 2025, now combines real-time dictation with ambient listening in a single clinical workflow.

In financial services, voice AI handles two mission-critical functions simultaneously: authentication and compliance. Biometric voice analysis can verify a customer’s identity by analyzing over 100 vocal characteristics, cutting identity checks from minutes to seconds while satisfying KYC and AML requirements. At the same time, real-time compliance monitoring flags potential regulatory violations during live calls — an agent recommending an unauthorized product, a missing disclosure, a sanctions-list match — alerting supervisors instantly rather than catching issues in a post-call review weeks later. Over 60% of financial firms plan to increase voice AI investment to boost both automation and fraud detection.

In sales, conversation intelligence platforms are turning every call into structured data. Real-time sentiment scoring helps reps adapt their pitch based on a prospect’s emotional state. Post-call analytics identify which talk tracks convert and which don’t. AI-assisted outbound campaigns enable round-the-clock prospect engagement, with some enterprises reporting 35% higher first-visit conversion rates. This isn’t replacing salespeople — it’s giving them the kind of coaching and analytics that used to require a dedicated enablement team.

The Consolidation Signal

The investment landscape tells its own story. Meta acquired Play AI for $23.5 million to embed voice capabilities into Meta AI products and smart glasses. SoundHound acquired Interactions for $60 million, bringing Fortune 100 clients into its voice portfolio. NICE acquired Cognigy in September 2025. ElevenLabs raised $180 million at a $3.3 billion valuation. Uniphore secured $260 million from Nvidia and AMD.

In total, more than 200 voice AI startups raised over $1.5 billion in 2025 alone. This kind of capital concentration signals that voice AI is moving from experimental to infrastructural — and that enterprises need to start treating it accordingly.

The Governance Gap Nobody’s Talking About

Here’s the problem: as voice AI proliferates across departments, the governance complexity multiplies in ways that text-based AI never required.

Voice data is inherently biometric. Every conversation captures patterns unique to the speaker — making governance essential — patterns that fall under GDPR, CCPA, BIPA, HIPAA, and an evolving patchwork of state and international regulations. The FCC has already ruled AI-generated robocalls illegal without prior written consent. Financial services firms deploying voice AI must satisfy PCI-DSS, SOC 2, and local regulator requirements — and in many jurisdictions, public cloud-only deployments may not even be compliant.

Then there’s the bias question. Speech recognition models trained on limited datasets still struggle with certain accents and dialects. In a customer-facing context, that’s not just a technical limitation — it’s a discrimination risk. And as voice AI handles increasingly sensitive workflows (clinical documentation, financial advice, legal consultations), the stakes of getting it wrong compound.

Deepfake spoofing adds another layer. Voice biometrics that seemed secure a year ago now require multi-factor verification — OTP codes, device fingerprints, behavioral analytics — to guard against synthetic voice attacks. The technology that makes voice AI powerful also makes it vulnerable.

Most enterprises deploying voice AI today have no unified way to monitor these risks across vendors and departments. The call center team uses one platform. Sales uses another. Healthcare uses a third. Each has its own compliance posture, its own accuracy metrics, its own cost structure — and nobody has the full picture.

Measuring What Actually Matters

The standard voice AI metric — call deflection rate — is necessary but insufficient. It tells you how many conversations the AI handled, not whether those conversations produced good outcomes. Enterprises that are serious about measuring AI ROI need a broader framework.

That means tracking revenue impact (conversion rates, upsell opportunities, time-to-resolution), quality metrics (CSAT, accuracy, escalation rates), risk metrics (compliance violations, hallucinations, customer churn from bad AI experiences), and true cost beyond infrastructure — vendor switching costs, integration complexity, the human effort required for QA at scale. As we found in studying 100+ AI agent deployments, the organizations that prove ROI are the ones that instrument these metrics from day one, not the ones that try to retrofit measurement after the fact.

Voice AI makes this measurement challenge particularly acute because conversations are ephemeral by nature. Unlike a chatbot transcript you can grep through, voice interactions require real-time analysis or expensive post-processing. The enterprises getting this right are the ones building measurement into their voice AI stack from the start — tracking accuracy, sentiment, compliance, and cost per interaction across every vendor and department in a single view.

Getting Started

If your organization is deploying voice AI — or if teams are already experimenting without central oversight — the first step isn’t choosing a vendor. It’s establishing visibility. Map where voice AI is being used today, what data it’s processing, which regulations apply, and what success looks like for each use case. That foundation makes everything else possible: vendor evaluation, governance policies, ROI measurement, and the confidence to scale what’s working.

We explored the accuracy breakthroughs driving this shift in depth on our podcast episode Breaking Through Voice AI Accuracy Barriers — worth a listen if you’re evaluating voice AI for your enterprise.

Ready to measure and govern your voice AI deployments? Talk to an expert to see how Olakai gives you unified visibility across every AI tool in your organization — voice included.

A voice AI agent in a retail call center was handling thousands of calls per month. Costs were down. Resolution rates were up. The operations team was thrilled.

Then the CFO asked a question no one could answer: “How much revenue did this thing actually generate?”

The basic metrics — calls handled, cost per call, resolution rate — told an efficiency story. But efficiency doesn’t get budget renewed. Revenue does. When the team finally tracked qualified leads that converted within 30 days, the agent proved thousands of dollars in quarterly value. Not cost savings. Revenue.

That’s the gap hiding in plain sight across enterprise AI today. And after measuring more than 100 AI agent deployments across retail, financial services, healthcare, and professional services, we’ve seen the same pattern repeat with remarkable consistency.

The $2.5 Trillion Question Nobody Can Answer

Global AI spending is projected to reach $2.5 trillion in 2026, according to Gartner. AI now represents more than 40% of total IT spending. Yet MIT’s Project NANDA found that 95% of companies see zero measurable bottom-line impact from their AI investments within six months.

Read that again. Trillions in spend. Ninety-five percent with nothing to show the CFO.

The problem isn’t that AI doesn’t work. The agents we’ve measured do work — they resolve tickets, qualify leads, process documents, flag anomalies. The problem is that most enterprises never connect that activity to business outcomes. They measure what’s easy (calls handled, tokens processed, tasks completed) instead of what matters (revenue influenced, costs avoided, risk reduced, time recovered).

This is why 61% of senior business leaders now report more pressure to prove AI ROI than they felt a year ago, according to Fortune’s 2025 CFO confidence survey. The era of “trust us, AI is helping” is over.

What 100+ Deployments Actually Taught Us

Across more than 100 measured agent deployments, we’ve identified four patterns that separate the 5% who prove ROI from the 95% who can’t.

1. They Define the Success KPI Before Deployment

The retail voice AI example above illustrates this perfectly. The operations team measured what they controlled: call volume, handle time, resolution rate. All green. But the finance team needed to see qualified leads that converted — a metric that crossed departmental boundaries and required connecting the agent’s activity to CRM data 30 days downstream.

The enterprises that prove ROI identify this “success KPI” before the agent goes live. Not after. Not when the CFO asks. Before. It’s the single metric that answers the question: If this agent works perfectly, what business outcome changes?

2. They Measure the Counterfactual, Not Just the Output

One financial services firm deployed an AI agent to flag compliance anomalies. The agent flagged 340 issues in its first quarter. Impressive? The team thought so — until someone asked how many of those would have been caught by the existing manual process. The answer was 312. The agent’s real value wasn’t 340 flags. It was 28 catches that would have been missed, each representing potential regulatory exposure worth six figures.

Measuring output without a baseline is vanity metrics dressed up as ROI. The question isn’t “what did the agent do?” It’s “what would have happened without it?”

3. They Track Cost-to-Value, Not Just Cost-to-Run

Enterprise AI cost conversations almost always focus on infrastructure: compute costs, API calls, token usage. These matter, but they’re only half the equation. A customer success agent we measured cost $4,200 per month to run — and prevented an average of $47,000 in monthly churn by identifying at-risk accounts three weeks earlier than the human team. The cost-to-run looked expensive in isolation. The cost-to-value ratio was 11:1.

The enterprises that scale AI investment successfully present both numbers to finance. They don’t defend the cost. They contextualize it against the value.

4. They Build Governance Into Measurement, Not Around It

Here’s the pattern that surprised us most. The deployments with the strongest ROI data weren’t the ones with the most sophisticated AI models. They were the ones with the most rigorous governance frameworks. Why? Because governance forces you to define what the agent is allowed to do, which forces you to define what success looks like, which forces you to instrument the metrics that prove value.

Governance and measurement aren’t separate workstreams. They’re the same workstream. Organizations that treat them as separate end up with compliant agents they can’t prove are valuable, or valuable agents they can’t prove are compliant.

The SEE → MEASURE → DECIDE → ACT Framework

These four patterns map to a framework we’ve refined across every deployment:

SEE: Get unified visibility into what AI agents are actually doing across your organization. Not just which agents exist, but what they’re touching — which data, which workflows, which customer interactions. You can’t measure what you can’t see, and most enterprises have agents running in places they don’t even know about.

MEASURE: Connect agent activity to the success KPIs that matter to the business. This means going beyond operational metrics (tokens, latency, uptime) to outcome metrics (revenue influenced, costs avoided, risk mitigated). It also means establishing baselines so you can measure the counterfactual.

DECIDE: Use measurement data to make scaling decisions. Which agents get more budget? Which get sunset? Which workflows should be automated next? Without measurement, these decisions are political. With measurement, they’re strategic.

ACT: Scale what’s working, fix what’s not, and govern the entire portfolio continuously. This is where most enterprises stall — not because they lack the will, but because they lack the data to act with confidence.

The framework isn’t complicated. But it requires designing measurement and governance from day one, not bolting them on after deployment. Enterprises that bolt on measurement retroactively spend 3-4x more time and money instrumenting metrics than those who build it in from the start.

Why This Matters Now

Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026 — up from less than 5% in 2025. That’s an 8x increase in one year. Meanwhile, 58% of organizations still cite unclear ownership as their primary barrier to measuring AI performance, and 62% lack a comprehensive inventory of the AI applications they’re running.

The math is straightforward. Agent proliferation is accelerating. Measurement capability is not keeping pace. The gap between AI activity and AI accountability is widening every quarter. And the organizations that close that gap first will be the ones who scale AI investment while their competitors are still stuck in pilot purgatory, unable to answer the CFO’s question.

In 2026, AI is being judged less on promise and more on proof. The playbook for providing that proof exists. It starts with seeing what you have, measuring what matters, deciding with data, and acting with confidence.

If your enterprise is deploying AI agents and struggling to prove their value, you’re not alone — but the organizations pulling ahead aren’t waiting for better AI. They’re building better measurement. Our AI ROI framework breaks down the methodology, and Future of Agentic’s success KPI library offers specific metrics by use case.

Ready to see what your AI agents are actually worth? Talk to an expert and we’ll show you how enterprises are turning AI activity into measurable business outcomes.

Forget the gym membership. Here’s the 2026 resolution that will actually transform your organization.

Every January, leadership teams gather to set priorities for the year ahead. They review budgets, realign strategies, and make bold promises about what they’ll accomplish. But if your organization launched AI initiatives in 2024 or 2025, there’s one resolution that matters more than all the others: this year, you’re going to measure what your AI is actually doing.

It sounds simple. It isn’t. According to Gartner, at least 30% of generative AI projects were abandoned after proof of concept by the end of 2025—not because they failed, but because teams couldn’t demonstrate clear business value. The AI worked. The measurement didn’t.

The Pilot Purgatory Problem

If you’ve been in enterprise technology for any length of time, you’ve seen this movie before. A promising technology emerges. Teams rush to experiment. Pilots launch across departments. And then… nothing. The pilots keep running, but they never scale. They become permanent experiments, consuming budget and attention without ever delivering the transformation they promised.

AI has accelerated this pattern dramatically. The barrier to launching an AI pilot is lower than ever—a team can spin up a chatbot or copilot integration in days. But the barrier to proving that pilot’s value remains stubbornly high. When the CFO asks “What’s the ROI on our AI investment?”, most teams can only offer anecdotes and assumptions.

This is pilot purgatory, and it’s where AI initiatives go to languish. A recent industry analysis found that on average, only 48% of AI projects make it into production, and it takes 8 months to go from prototype to production. The problem isn’t the technology. It’s the inability to answer the fundamental question: is this working?

Why 2026 Is Different

The pressure to prove AI value has never been higher. After two years of experimentation, boards and executive teams are demanding results. They’ve seen the hype. They’ve approved the budgets. Now they want to know what they got for their investment.

Meanwhile, AI capabilities are advancing rapidly. Agentic AI—systems that can autonomously plan and execute complex tasks—is moving from research labs to production environments. Organizations that can’t measure the value of their current AI deployments will struggle to make informed decisions about these more sophisticated (and more expensive) capabilities.

The teams that figure out measurement in 2026 will scale their AI programs. The teams that don’t will watch their pilots slowly fade away, replaced by the next wave of experiments that also never prove their worth.

Five Measurement Commitments for 2026

Making “measure AI” a meaningful resolution requires specific commitments. Here’s what the teams that escape pilot purgatory actually do differently.

First, they track outcomes, not just usage. Knowing that 500 employees used your AI assistant last month tells you almost nothing. Knowing that those employees resolved customer issues 23% faster, or processed invoices with 15% fewer errors—that’s actionable intelligence. The shift from counting interactions to measuring business impact is the single most important change most organizations need to make.

Second, they tie AI to existing business KPIs. Your organization already measures what matters: revenue, costs, customer satisfaction, employee productivity, error rates, cycle times. Effective AI measurement connects AI usage to these existing metrics rather than creating a parallel universe of AI-specific vanity metrics. When you can show that teams using AI tools have 18% higher customer satisfaction scores, you’ve made the business case.

Third, they monitor costs proactively. AI costs can spiral quickly—API calls, compute resources, vendor subscriptions, integration maintenance. Teams that measure well know their cost per outcome, not just their total spend. They can answer questions like “How much does it cost us to resolve a customer issue with AI assistance versus without?” This kind of granular cost visibility is essential for making scaling decisions.

Fourth, they document what’s working and what isn’t. The value of AI measurement isn’t just in proving ROI—it’s in learning. Which use cases deliver the highest value? Which teams have figured out how to get the most from AI tools? Which integrations consistently underperform? Organizations that systematically capture these insights can make smarter decisions about where to invest next.

Fifth, they build the case for scaling incrementally. The path from pilot to production isn’t a single leap—it’s a series of gates, each requiring evidence that the AI is delivering value. Teams that measure well can show steady improvement over time, building confidence with stakeholders and earning the resources needed to expand.

How to Actually Keep This Resolution

Unlike most New Year’s resolutions, measuring AI doesn’t require willpower—it requires infrastructure. You need systems that capture AI usage data, connect it to business outcomes, and present it in ways that executives and finance teams can act on.

This is where many organizations stumble. They try to build measurement capabilities from scratch, cobbling together logging tools, custom dashboards, and manual reporting processes. The result is fragile, incomplete, and almost never maintained once the initial enthusiasm fades.

The more sustainable approach is to implement purpose-built AI intelligence platforms that handle measurement automatically. These platforms integrate with your existing AI tools—chatbots, copilots, agent frameworks, AI-enabled SaaS—and provide unified visibility into usage, outcomes, and costs across all of them. Olakai, for example, was built specifically to solve this problem: giving enterprises the data they need to prove AI value and make confident scaling decisions.

The Payoff

Teams that measure scale. Teams that don’t stay stuck in pilot purgatory indefinitely. It’s that simple.

When you can show the CFO exactly how much value your AI initiatives are delivering—in terms they understand, tied to metrics they already care about—you transform the conversation. You move from defending your AI budget to advocating for expansion. You shift from “we think this is working” to “here’s the data proving it works.”

More importantly, you give your organization the information it needs to make smart decisions about AI. Not every pilot should scale. Not every use case delivers value. Measurement lets you distinguish the winners from the losers and concentrate resources where they’ll have the greatest impact.

2026 will be the year that separates the organizations that figured out AI from the ones still experimenting. The difference won’t be which AI tools they chose or how sophisticated their implementations were. It will be whether they could prove their AI was working—and use that proof to build something lasting.

That’s a resolution worth keeping.

Ready to start 2026 with visibility into your AI investments? Talk to an expert to see how Olakai measures AI ROI across your entire organization.