At a room full of CIOs two months ago, AWS CEO Matt Garman asked how many were seeing materially positive AI ROI — or had a clear path to it within months. Ninety percent of hands went up. Then Wedbush Securities' analysts talked to enterprise executives at their Disruptive Technology Conference last week and found a different truth: most of those companies can't actually prove it.
This is the central paradox of enterprise AI in mid-2026. Optimism is at an all-time high. Confidence is up. Investment is accelerating. And yet, when boards and CFOs ask the simple question — "What are we getting back from this?" — the majority of organizations go quiet.
Understanding why this gap exists, and what to do about it, is the most important strategic question enterprise technology and business leaders are facing right now.
The AWS CEO's Confident Take
Matt Garman's 90% stat deserves unpacking. In an interview on the Platformer podcast, the AWS CEO described a fundamental shift in how enterprise leaders are experiencing AI — from a research experiment to a genuine business tool with measurable impact.
The context matters. Garman wasn't surveying random IT managers. He was in a room of CIOs who are actively deploying AI at scale, many of them AWS enterprise customers. These are organizations with the infrastructure, data governance, and technical talent to actually implement AI thoughtfully. That population skews optimistic by definition.
Still, the directional signal is real. A year before, Garman said, the response would have looked dramatically different. The shift from single-digit to 90% confidence in ROI — even within a select group — represents genuine progress in how enterprise AI is being deployed and measured.
Amazon is backing this conviction with capital. The company has committed $200 billion in capital expenditure this year, underpinned by what Garman describes as customer visibility, not speculation. "We have a lot of mitigations in there and we really think intentionally about how we can reduce risk," he said. That's a company making infrastructure bets based on actual demand signals, not hype cycles.
The Wedbush Warning That Changes the Story
Here's where it gets complicated. Dan Ives and the Wedbush Securities analyst team spent time at their Disruptive Technology Conference last week talking to enterprise executives. Their Friday investor note carried a message that cuts against the optimism:
Many enterprises have invested in AI pilots without a framework for gauging success. Without such a framework, they are likely to encounter difficulties justifying the investment, identifying which approaches are actually working, and building organizational confidence in AI-driven decision-making.
Ives put it directly: "Many executives noted that customers are feeling increased pressure from their boards and CFOs to demonstrate actual returns from AI, and the inability to answer this question presents a real barrier to additional investments in long-term technological buildouts."
This isn't a fringe concern. When you put both findings together, you get a troubling picture: the people with the confidence are the ones who've already built measurement systems. The majority of enterprises — still running pilots, still experimenting, still scaling — are operating without the benchmarks they need to defend their investments when the next budget cycle comes around.
Why 71% Are Blaming the Organization, Not the Technology
One of the more clarifying data points comes from a PYMNTS Intelligence study that surveyed enterprise executives about what's actually limiting their AI performance. Seventy-one percent pointed to their organization's people, processes, or data readiness — not the technology itself.
The executives cited an average of four to five organizational barriers simultaneously. The most common bottlenecks were data quality, budget limitations, and governance processes.
This matters for how you diagnose your own situation. If your AI investments aren't producing measurable returns, the most likely culprit isn't the model you chose or the vendor you partnered with. It's the scaffolding around the deployment: Are your data pipelines clean enough to produce reliable outputs? Do you have governance processes that let the AI actually make decisions rather than just generate recommendations that humans re-review anyway? Is there organizational clarity on who owns the outcomes?
Fixing the technology is the easy part. Building the organizational infrastructure to use it is where most enterprises stall.
The Measurement Framework Most Organizations Are Missing
Here's what distinguishes organizations that can prove AI ROI from those that can't: they defined what success looks like before they started, not after.
This sounds obvious. In practice, it rarely happens. Most AI pilots begin with a vague mandate to "explore what's possible" or "demonstrate value" — neither of which gives you anything concrete to measure. When the quarterly review comes, the team reports on outputs (number of prompts processed, hours saved in one workflow, user adoption percentages) without connecting them to the financial outcomes the CFO actually cares about.
A meaningful AI ROI framework tracks three layers:
Operational efficiency metrics — how much time, cost, or headcount has been removed from a specific process? These should be calculated in dollar terms, not percentage improvements. "We reduced invoice processing time by 40%" is a progress metric. "Invoice processing cost dropped from $18 per invoice to $7, saving us $2.2 million annually across 200,000 invoices" is a ROI metric.
Revenue impact metrics — has AI enabled the business to capture opportunities it couldn't before? Faster sales cycles, higher conversion rates from better customer data, reduced churn because of proactive service. These are harder to attribute directly, but if you can't attempt the calculation, you don't have a ROI story.
Risk-reduction value — the hardest to quantify but often the largest. Compliance catches, fraud detection, security threat identification. Actuarial firms routinely value risk reduction in dollar terms for insurance purposes. Enterprise AI leaders need to apply the same discipline.
The trap most organizations fall into is measuring activity instead of outcomes. Token consumption, model API calls, active users — these are operational metrics for IT. They tell you nothing a CFO needs to know to sign off on next year's budget.
What Matt Garman's Three Tactics Actually Mean
Garman offered three practical approaches during the Platformer interview, and they're worth translating from CEO-speak into operational guidance.
Use the right model for the job, not the most expensive one. This is a cost management principle disguised as a technical decision. Most enterprise AI overspend happens because teams default to their most capable (and most expensive) model for everything. AWS built routing logic into its Kiro development environment specifically to address this — lighter models handle code generation, higher-reasoning models handle complex decision support.
For enterprise buyers, the equivalent decision is about model tiering in your vendor contracts. If you're using a frontier model API at scale, you should have a clear policy on which tasks justify that cost and which should run on smaller, faster, cheaper alternatives. Without this discipline, your AI infrastructure costs will grow faster than the value you're capturing.
Measure outcomes, not token consumption. Garman's framing here was about treating employees like owners over their AI usage — focusing on what the AI delivers rather than how much compute it burns. The CFO translation: your AI performance dashboards should show business metrics, not infrastructure metrics.
This requires a rethink of how you set up reporting. Your engineering team will naturally instrument what they can measure — API calls, latency, model responses. Your business team needs a parallel dashboard that shows what those calls are producing in terms the P&L recognizes.
Double down on what works; quickly shut down what doesn't. The fastest path to ROI is ruthless prioritization. Garman's word was "disciplined" — scale the use cases showing results, cut the ones that aren't. This is harder than it sounds because enterprise organizations have inertia around projects once they've started. The pilot that isn't working still has a team, a budget owner, and a sponsor who's publicly committed to it.
The organizations producing the clearest AI ROI are the ones that have built explicit review gates into their deployment process. Every AI initiative has a 90-day checkpoint where it has to demonstrate movement on pre-defined financial metrics or it gets paused or cancelled. No exceptions.
The Board Conversation That's Coming
Wedbush is right that this is becoming a board-level issue. The sequence of events is predictable: an enterprise spends two to three years experimenting with AI. The costs accumulate on the balance sheet. The board, increasingly composed of directors who've seen this story before with ERP implementations and cloud migrations, wants to understand the return.
If the answer is "we believe the value is there but we're still working on the measurement framework," that's a governance failure — not a technology one.
Executives facing this conversation in the next six to twelve months need to move now. The window to retroactively build measurement infrastructure is closing. Boards that were patient with AI exploration in 2024 and 2025 are entering a period of accountability in 2026.
The specific asks to expect: what's the dollar ROI on our three largest AI deployments, how does our AI investment compare to peers in our industry, and what's the plan to scale the investments that are working.
Organizations that can answer those three questions confidently will get more budget. Organizations that can't will face the cuts that Wedbush is warning about.
A Practical Starting Point for Enterprise Leaders
If your organization is in the "we have AI but can't prove the ROI" category, the path forward is not to wait for a better measurement system to magically appear. It's to pick one deployment that is already running and build the measurement framework around it retroactively.
Choose the AI deployment with the clearest process connection — something like invoice processing automation, customer support ticket routing, or sales lead scoring. These have obvious input/output relationships and existing cost data you can compare against.
For that deployment, document: what did this process cost before AI (in fully-loaded terms including labor), what does it cost now, and what is the quality differential (error rates, processing times, customer satisfaction scores if applicable). Calculate the annualized savings. That number, derived from one concrete deployment, gives you the template to apply across the rest of your portfolio.
It's not comprehensive. But it's a defensible starting point, and in board conversations, a defensible number is infinitely more useful than a vague confidence statement.
The 90% of CIOs who raised their hands for Matt Garman have this already. The rest of the market is catching up — or will face the consequences of not doing so.
Sources:
- AWS CEO Matt Garman interview — Platformer / About Amazon (June 2026)
- Wedbush: Missing ROI Metrics Threaten Enterprise AI Deployment — PYMNTS (June 26, 2026)
- The Enterprise AI Readiness Gap — PYMNTS Intelligence
