TL;DR
- 93% of enterprise AI teams improved production capability in the past year, but 57% report ROI growing no faster than investment — identical to 2025.
- Up to 50% of organizations are deploying agentic AI with zero governance infrastructure, exposing themselves to regulatory sanctions and competitive loss.
- Response refinement loops — not model pricing — consume 60% of agentic AI operating costs, yet most enterprises have no way to measure whether those costs deliver value.
- Companies that build governance and cost-tracking infrastructure before production rollout are 3.9x more likely to achieve governed deployment and deliver 75% faster than peers.
Speed Is the New Vanity Metric
There's a particular kind of organizational pride that comes from shipping fast. In enterprise AI right now, that pride is everywhere — and it may be masking a slow-motion crisis.
According to Domino Data Labs' Fifth Annual Enterprise AI Report (July 2026, 639 enterprise AI leaders across North America, the UK, and Europe), a remarkable 93% of respondents improved their ability to move AI from experimentation to production in the past twelve months. Forty-three percent already have agentic AI running in production. Another 41% are actively piloting or scaling it. By every deployment metric, enterprise AI is having its moment.
And yet. The ROI numbers look almost identical to last year. 57% of enterprise AI leaders report that returns are growing at the same pace as investment — or slower. That's not a rounding error. That's the same finding, from the same survey, two years in a row. The needle hasn't moved.
So what exactly is everyone speeding toward?
The Governance Void Nobody Is Talking About Loudly Enough
Here's where it gets uncomfortable: a significant slice of organizations aren't just missing ROI — they're missing the basic infrastructure to even understand why. Roughly 40% of North American respondents and 50% of European respondents who are deploying or scaling agentic AI are doing so without governance infrastructure in place. No audit trails. No model lineage tracking. No oversight controls.
That's not a technical debt problem. That's a liability.
The survey respondents seem to know it, too. 44% cite regulatory sanctions from ungoverned AI outputs as a top risk, and 42% flag revenue loss from competitors who are deploying faster with better guardrails. Shadow AI (38%) and board-level scrutiny after failed ROI demonstrations (34%) round out a risk picture that looks less like an innovation roadmap and more like a list of things that could go very wrong, very publicly.
The irony is sharp: enterprises are racing to scale agents partly out of competitive anxiety — and in doing so, they're creating the exact governance gaps that will slow them down or sanction them later. It's the AI equivalent of skipping the foundation to finish the roof faster.
The Cost Architecture Problem: It's Not the Tokens, It's the Loops
Meanwhile, CFOs are getting a separate kind of surprise. McKinsey's May 2026 Enterprise AI FinOps Survey (75 qualified respondents) found that 93% of enterprise AI teams exceed their AI budgets. Token prices have fallen roughly 99% over two years. Enterprise AI bills have tripled. Something doesn't add up — until you look at where the money is actually going.
Response refinement loops — the validation, checking, and correction cycles baked into agentic architectures — consume approximately 60% of agentic AI operating costs. This isn't a model pricing problem. It's an architecture problem. Every time an agent second-guesses itself, validates an output, or loops back for correction, that's compute spend. And most enterprises have no instrumentation to measure whether those refinement cycles are producing outputs that are actually accurate, actually used, or actually worth more than what they cost.
The practical result: 20% of organizations are already constraining AI use specifically because of operating costs. The tool that was supposed to unlock productivity is getting throttled because nobody built the financial visibility layer to justify the spend.
"Most orgs are tracking token spend. Few understand the full cost of running AI at scale. The economics of agentic AI extend across tech, ops, governance & people. In this @HBR article, @EY shares why Agentic FinOps is emerging as a critical discipline."
— @danielmellen
McKinsey frames this squarely: the real metric for agentic AI isn't cost-per-token. It's cost-per-completed-accurate-task — and almost nobody is measuring it yet. As one practitioner in the FinOps space is already betting on:
"Every cloud dollar — AWS, GCP, Vercel, your AI stack — in one reconciled ledger, watched 24/7 by an agent that catches the spike, explains it, and drafts the fix."
— @Pavel_FFP
The race to build agentic FinOps tooling is real, and it reflects a genuine market gap. Enterprises are flying financially blind.
The Governance Advantage Is Real — and Measurable
Here's the thing that should reframe this entire conversation: governance isn't a slowdown. The data says it's a speed multiplier.
Among companies that have fully integrated AI governance, 75% report significantly improved delivery velocity. Among organizations with only partial governance? That number drops to 23%. Governance-first organizations are 3.9x more likely to have agents operating in governed production environments.
The regulated verticals — Financial Services, Public Sector, and Life Sciences — are actually leading agent deployment, not lagging it. That's counterintuitive only if you assume governance is a brake. In practice, it turns out that knowing your regulatory constraints in advance makes you build better systems the first time. Who knew that the compliance-heavy sectors had something to teach everyone else?
"Nutanix to demonstrate enterprise AI innovations for agentic AI era at @AMD Advancing AI 2026 — demonstrating how organizations can deploy agentic AI at enterprise scale."
— @MikeLongTerm
The infrastructure conversation is maturing. The question is whether most enterprises will catch up before the audit letters arrive.
The Pattern (And How to Break It)
The failure mode here follows a recognizable script:
- Pressure to deploy fast drives organizations to skip design-phase governance and cost architecture work.
- Agents go into production without audit trails, cost attribution by workload phase, or model lineage.
- Cost overruns emerge — not from token prices, but from refinement loop architecture that nobody modeled in advance.
- ROI justification fails because there's no instrumentation to connect agent activity to business outcomes.
- Governance gaps surface during incidents, audits, or regulatory reviews — precisely when it's most expensive to retrofit them.
The organizations breaking this pattern aren't doing something exotic. They're simply doing governance and cost architecture work during the design phase, not as a post-launch scramble. That means building audit trails before agents go live. It means attributing costs by workload phase so you can see which loops are burning budget without producing value. It means knowing your regulatory exposure before your agents make decisions at scale.
For implementation partners and IT consultancies, this is a defining moment. Enterprises that recognize the gap are actively looking for design-phase audits — not post-launch firefighting. The value of catching a governance blind spot before production is exponentially higher than patching it after a regulatory notice or a board inquiry.
The Honest Assessment
The headline numbers — 93% improved production capability, 43% in production deployment — are genuinely impressive. Enterprise AI has moved from lab curiosity to operational reality at a pace few predicted two years ago.
But production velocity without governance architecture and cost visibility is a bit like driving very fast with your headlights off. You're covering ground. You just don't know what's in front of you.
The 57% ROI stagnation figure is the tell. After two years of aggressive production deployment, enterprise AI is still struggling to demonstrate returns that outpace investment. The problem isn't the technology. The problem is that the organizational infrastructure required to govern, measure, and optimize that technology at scale is lagging badly.
The silver lining — and there genuinely is one — is that the path forward is well-lit by the organizations who've already walked it. Governance-first builds faster. Cost visibility enables ROI justification. Design-phase audits prevent production disasters. The data is unambiguous.
The only question is whether enterprises will read it before or after their next governance crisis.
Sources: Domino Data Labs Fifth Annual Enterprise AI Report, July 2026 (639 enterprise AI leaders); McKinsey Enterprise AI FinOps Survey, May 2026 (75 qualified respondents); McKinsey 2026 State of AI global survey (1,719 participants, fielded May 4–June 8, 2026).
Published in Stream · Dispatch #460 · July 22, 2026 · 7 min read.
Reply to paolo@mont3.ch - every email gets a human answer within 24h.