One year ago, a CIO might have raised their hand to say AI was delivering ROI. Now, nearly all of them are. That's the signal buried inside a Platformer podcast interview with AWS CEO Matt Garman this week — and it represents one of the most significant sentiment shifts in enterprise technology in recent memory.
"I was talking to a room full of CIOs just a couple of months ago," Garman said. "I asked, 'How many of you are either seeing materially positive ROI today or have a path in the next couple of months to really high ROI?' 90% of hands went up, which is totally different than a year before."
That single data point is worth sitting with. A year ago, most enterprise AI conversations were about experimentation, pilots, and "exploring the technology." Today, the conversation has moved to optimization, scaling, and returns. The question is no longer "should we invest in AI?" — it's "how do we maximize what we're getting back?"
The Shift From Experimentation to Production
Garman draws a direct comparison between the current AI moment and the early cloud era. He joined Amazon before AWS launched, was among the first product managers for EC2, and has watched cloud adoption reshape enterprise IT over two decades. His read on AI adoption speed is unusually well-informed.
"Cloud still hasn't displaced the vast majority of enterprise workloads after 20 years," Garman noted. "AI is moving faster, because cloud infrastructure is already in place."
That infrastructure advantage matters more than most leaders appreciate. The reason AI deployment timelines are compressing is that enterprises aren't starting from scratch. Data center investments, API-first architecture, and cloud-native development practices built over the past decade are now the foundation for AI production workloads — and the ramp time has collapsed.
For CIOs and CTOs evaluating where they stand: if your organization is still in pilot mode while 90% of your peer group is measuring ROI, the gap has become a competitive liability, not a prudent pause.
The 3 Tactics Separating High-ROI Enterprises
Garman didn't just share the headline number — he outlined what enterprises generating real returns are actually doing differently. These aren't abstract principles; they're operational decisions that directly affect cost structures and outcome velocity.
Use the Right Model for the Job — Not the Most Expensive One
The single biggest avoidable cost in enterprise AI deployments, Garman said, is model misallocation. Organizations default to the most powerful, most expensive model for every task — regardless of whether that task requires frontier reasoning capability or not.
"One of the things that's driven up a bunch of cost of AI is that people were trying to use the best model for every single thing," Garman told Platformer. "We pick the right model and then appropriately help customers budget to get the results they want faster, and less expensively."
AWS's agentic development environment Kiro does this automatically — routing code generation tasks to lighter models and reserving high-reasoning models for complex, nuanced requests. The cost differential between these tiers can be 10x-50x per token, which at enterprise scale translates to millions of dollars annually in avoidable spend.
For CFOs reviewing AI spend: if your organization isn't implementing intelligent model routing, you're almost certainly overpaying. The ROI math changes significantly when you match model capability to task complexity. For CTOs and VPs of Engineering: this is an architecture decision, not just a cost optimization. Building model selection into your AI pipeline is becoming table stakes for production-grade systems.
Measure Outcomes, Not Token Consumption
This sounds obvious. It isn't. The default reporting metric in most enterprise AI dashboards today is token usage — how many API calls were made, how much compute was consumed, how many credits were spent. These are cost metrics, not value metrics.
Garman's prescription is to shift the measurement frame entirely: what did the AI actually deliver? How many support tickets were resolved faster? How many contracts were reviewed? How many lines of production code shipped?
"Let employees behave like owners over their AI usage — focus on what AI delivers rather than how much compute it burns," Garman said.
The supporting data validates this framing. McKinsey's Global AI Survey 2026 found that knowledge workers using production AI agents recover a median of 6.4 hours per week per seat. Software engineers using AI coding assistants are shipping 26-55% more code per sprint. Customer support teams are resolving tickets 25-40% faster.
None of these numbers show up in a token consumption report. They require connecting AI usage to workflow-level outcomes. Organizations that make this measurement shift are the ones discovering that AI isn't just a productivity multiplier — it's a cost-per-unit reducer that compounds over time. For CFOs: the ROI calculation needs to live in the business unit P&L, not the IT budget.
Scale What Works, Cut What Doesn't — Fast
The third tactic is the most culturally difficult: making fast, decisive calls on which AI use cases to double down on and which to kill.
"Double down on what works; quickly shut down what doesn't. The fastest path to ROI is a disciplined approach," Garman said.
In conversations with business and technology leaders across industries this year, the pattern I've consistently seen is organizations running 15-20 AI pilots simultaneously, with none getting the organizational focus or resource commitment needed to reach scale. The result is modest ROI spread thin — small percentage improvements across a wide surface area rather than dramatic gains in a few high-leverage workflows.
The enterprises generating outsized returns are typically doing the opposite: identifying two or three workflows with clear, measurable outcome potential, committing seriously to them, and treating everything else as a future roadmap item. The discipline is the differentiator — not the technology.
The Uncomfortable Counterpoint: IBM's 25%
Before celebrating the 90% headline, there's a critical data point that deserves equal attention.
IBM's 2025 CEO study found that only 25% of AI initiatives delivered expected ROI. IDC and Microsoft report a 3.7x average return per dollar invested in generative AI — but "average" conceals enormous variance. The top quartile of AI deployments is pulling that number up dramatically; the bottom quartile is barely breaking even.
What this tells us: the 90% of CIOs raising their hands includes organizations with materially different outcomes. Some are seeing 10x returns on specific workflows. Others are seeing modest productivity gains that barely justify the investment.
The gap is largely predictable. Organizations following Garman's three principles — right model for the job, outcome-focused measurement, disciplined portfolio management — consistently land in the top quartile. Organizations that defaulted to deploying the most powerful model everywhere, measuring token consumption, and running sprawling pilot portfolios are in the bottom half.
This isn't a technology problem. It's an operational discipline problem.
What Amazon's $200B Bet Signals About Enterprise Demand
Garman's ROI claims aren't executive optimism — they're grounded in customer purchase behavior. Amazon's $200 billion capital expenditure commitment in 2026, the bulk of it on AI infrastructure, isn't speculative investing. It's a direct response to signed customer commitments and forward-looking demand signals.
"We really think intentionally about how we can reduce risk," Garman said. He highlighted that server and chip commitments are only made months out, when customer visibility is high. Land and power investments are more durable — they retain value even if AI demand shifts. "If you really like the ROIC of a business, you want the 'C' to be as high as possible. It's not speculative."
For CIOs doing strategic planning: this level of infrastructure commitment from AWS, Azure, and GCP is a strong signal that enterprise AI is not entering a correction cycle. Vendor roadmaps, pricing strategies, and capability releases over the next 18-24 months are all being funded and built now. Organizations that delay production deployments aren't avoiding risk — they're creating it.
The Decision Framework for Q3 2026
If you're a CIO, CTO, or CFO reading this, the 90% data point creates a specific set of immediate decisions.
For CIOs: If you're not yet measuring AI ROI at the workflow level, this quarter is the deadline. You need outcomes data, not pilot reports. If you have more than five active AI initiatives without clear ROI targets, consolidate to two or three with serious resource commitment. The breadth-over-depth approach is the most reliable path to underwhelming returns.
For CTOs and VP Engineering: Model routing isn't an optimization — it's an architecture requirement for cost-effective production AI. Budget it accordingly in Q3. If you can't answer "what did this AI workflow deliver this week?", you don't have the right instrumentation. Outcome measurement requires instrumenting AI pipelines differently than most teams do today.
For CFOs: The IDC 3.7x return figure is an average that covers significant variance. The question isn't "is AI ROI real?" — it's "which specific initiatives are we running that are in the top quartile of returns?" If your AI investment sits entirely in IT's budget without corresponding productivity metrics from business units, the measurement framework is broken. Fix it before Q4 planning.
The Inflection Point Has Passed
The most significant thing Garman's 90% figure tells us isn't about ROI — it's about timing. The window for "wait and see" has closed. The leaders seeing real returns today built their data advantage, institutional knowledge, and vendor relationships while others were still deliberating.
That doesn't mean late movers can't catch up — but the catch-up curve is steeper. Organizations reporting positive ROI today are heading into the next phase: scaling proven use cases, building proprietary data advantages, and deploying autonomous agents against workflows that were previously too complex to automate.
In conversations with operations and technology leaders across industries this year, the consistent finding is that production AI deployments in 2026 are qualitatively different from 2024-2025 pilots. The tools are more reliable. Model costs have dropped 80-90% from 2023 levels. Integration tooling is more mature. And the organizational change management playbook is better understood.
The 90% of CIOs seeing ROI aren't lucky. They made specific operational decisions to get there. The 10% still figuring it out have access to the exact same models and the same cloud infrastructure. The difference is execution — and the three tactics Garman outlined are as good a framework as any for closing that gap fast.
Continue Reading
Sources:
- AWS CEO Matt Garman on enterprise AI ROI — Amazon, June 2026
- Platformer Podcast: Matt Garman interview — Platformer, June 2026
- State of the CIO 2026 — CIO.com, 2026
- Enterprise AI Agent Productivity Statistics 2026 — Digital Applied, 2026
- Enterprise AI Agents Adoption Statistics 2026 — Paul Okhrem, 2026
