TL;DR
- 93% of enterprises improved AI production capability in 2026, yet 57% still can't generate ROI that outpaces investment — a plateau now two years old.
- 41% of enterprises are actively piloting or scaling agentic AI right now with zero governance infrastructure in place.
- Enterprises with fully integrated governance are 3.9x more likely to reach governed agentic production and report 75% better delivery velocity than partially-governed peers.
- European enterprises are the most exposed: 50% are deploying or scaling agents without a governance framework, the highest gap of any region surveyed.
The Plateau Nobody Wants to Talk About at the All-Hands
There's a particular kind of discomfort that settles into a boardroom when the AI slide deck looks great and the P&L doesn't. Enterprises have spent the last two-plus years hiring data scientists, standing up MLOps pipelines, and confidently announcing that they are "AI-led organizations." And yet, according to Domino Data Lab's Fifth Annual Enterprise AI Report (published July 21, 2026), 57% of enterprises still report AI ROI that fails to outpace their investment — the exact same figure from 2025.
That's not a dip. That's a plateau. And it's a stubborn one.
To be fair, there's genuine progress happening underneath the headline. 93% of enterprises report improved AI production capability in 2026, up from 88% the prior year. Models are reaching production faster. Infrastructure is maturing. The machinery is getting better. The problem is that better machinery moving in the wrong direction doesn't actually get you anywhere useful.
So what's going wrong?
It's Not the Model. It's the Architecture.
Here's the uncomfortable truth that the industry is slowly digesting: the AI ROI gap isn't a model problem. You can swap in the most capable LLM on the market, fine-tune it on proprietary data, and wire it into the fanciest agentic orchestration framework — and still land in the same 57% bucket.
McKinsey's parallel research (published July 20, 2026) reinforces this point sharply: cost architecture and cost transparency are the operational bottlenecks, not the underlying model choice. The CIO and CFO are often pulling in opposite directions — the CIO chasing capability, the CFO demanding accountability — and without governance infrastructure bridging that gap, AI investments drift into ambiguity.
"Gartner: 40%+ of agentic AI projects may be canceled by 2027. Cause: cost, unclear ROI, weak controls — not the model. McKinsey: only 23% are actually scaling agents."
— @heymrun
That tweet neatly summarizes an industry-wide reckoning. Cancellation isn't a technology failure — it's a governance failure wearing a technology mask.
The 41% Problem: Governance-Free Zones in Agentic AI
The most striking finding from the Domino report is the near-parity between governed and ungoverned agentic deployments. Among enterprises currently engaged with agentic AI:
- 43% have agentic AI running in governed production.
- 41% are piloting (12%) or actively scaling (29%) agentic AI without the governance infrastructure to manage it.
Read that second bullet again slowly. Nearly four in ten enterprises are not just experimenting without guardrails — they are scaling without them. The organizations actively scaling ungoverned agents outnumber those merely piloting by more than two to one. That's not a pilot program risk; that's an operational risk at enterprise velocity.
This matters because agentic AI isn't a passive analytics tool. Agents make decisions, trigger workflows, interact with external systems, and increasingly take actions with real business consequences. Running them without governance infrastructure isn't just imprudent — in regulated industries, it may not be legal for long.
"AI is no longer just helping financial services — it's starting to do the work. Trust, governance & regulation will be critical."
— @InnFin
The Numbers That Should Change Your Roadmap
The Domino data doesn't just identify the problem — it quantifies the reward for solving it. The governance maturity gap creates a measurable, compounding advantage:
| Governance Status | % Reaching Governed Agentic Production | Significantly Improved Delivery Velocity |
|---|---|---|
| Fully integrated governance | 67.5% | 75% |
| Partially keeping pace | 17.2% | 23% |
| Multiplier | 3.9x | ~3.3x |
Fully governed organizations are nearly four times more likely to successfully reach governed agentic production than their partially-governed peers. And three-quarters of them report significantly improved AI delivery velocity — compared to fewer than one in four partially-governed organizations.
This isn't a marginal efficiency gain. This is a structural divergence. Enterprises that treat governance as a post-launch checkbox are not just slower — they are on a fundamentally different trajectory.
Thomas Robinson, COO at Domino Data Lab, put it plainly: "Getting a model into production used to be the milestone that mattered. Our research shows that's not enough anymore. The real milestone is the moment a business user can act on what the model found."
The Last-Mile Gap: When AI Reaches Production but Not People
Even among the enterprises generating value from AI, there's a second, quieter failure mode: the last-mile gap. Production and delivery, it turns out, are not synonyms.
The Domino report found a fragmented access picture across the enterprise:
- 34% of organizations report a patchwork of AI access methods that varies by business unit — the single most common response.
- 40% still rely on at least one fully mediated access method: a scheduled report from a data science team, or an analyst queue where business users submit requests and wait.
Imagine building a high-speed rail network and then requiring passengers to submit a formal written request 48 hours in advance to board a train. That's roughly the situation for AI delivery in a large slice of enterprise organizations today. The infrastructure exists. The value isn't flowing.
This is the architectural gap that governance frameworks are designed to close — standardizing delivery, ensuring auditability, and creating consistent access experiences across business units rather than leaving each department to improvise their own connection to the model layer.
"Most people talk about Agentic AI. Very few can actually design it."
— @MeenakshiYACS
Designing agentic AI — with its orchestration layers, tool-calling logic, and multi-step decision chains — is genuinely hard. But designing it with governance baked in from the start is the critical differentiator the data now clearly supports.
Europe's Particular Exposure
Regional data from the Domino report adds a geographical dimension to the governance gap that deserves attention. Among the three regions surveyed:
- Europe: 57.4% lack fully integrated AI governance; 50% are deploying or scaling agents without governance frameworks — the highest gap of any region.
- North America: 40% scaling agents without governance.
- UK: 38% scaling agents without governance.
European enterprises face a compounding challenge: higher regulatory exposure (GDPR, the EU AI Act's tiered risk requirements) combined with the lowest governance maturity rates of the regions studied. This is not a comfortable combination. Deploying ungoverned agentic systems in a regulatory environment that is actively developing enforcement mechanisms is a risk that will eventually price itself into the business.
For IT implementation partners operating in European markets, this represents a specific and time-sensitive opportunity. The organizations that need governance architecture the most urgently are often the ones that have moved the fastest on deployment — and the slowest on the framework to manage it.
Financial Services Is Showing the Way
It's not all grim news. The Domino report highlights that financial services, banking, and insurance organizations — among the most heavily regulated sectors — are leading on both governance maturity and production velocity.
This is counterintuitive only if you assume that regulation slows innovation. In practice, regulated sectors often build better infrastructure precisely because they have to. When the cost of a governance failure is a regulatory action, not just a missed KPI, the incentive structure changes. The NIST AI Risk Management Framework has become a key reference point for these organizations, providing structured guidance on identifying, assessing, and managing AI risk across the model lifecycle. (NIST AI RMF)
The implication for less-regulated sectors is worth sitting with: the governance discipline that financial services built out of regulatory necessity is now proving to be a competitive advantage in AI delivery velocity. Compliance frameworks, it turns out, are also delivery frameworks.
The Governance-First Imperative
The pattern in the data is unambiguous. The enterprises that will own the next wave of agentic AI scaling are not the ones with the most sophisticated models or the most aggressive deployment timelines. They are the ones that build governance infrastructure into initial agentic design — not as a compliance add-on, but as a foundational architectural decision.
The 41% of enterprises currently scaling agents without governance are not bad actors. Many of them are genuinely excited about the technology, moving fast in response to competitive pressure, and fully intending to "sort out governance later." The Domino data suggests "later" is where ROI goes to stall.
For IT implementation partners, this creates a clear positioning choice. The firms that show up as governance architects during the design phase — not post-launch fixers called in after the agentic deployment has already accumulated technical and compliance debt — will own the relationships that matter most as enterprise agentic AI matures.
The inflection is happening now. The two-year ROI plateau is a solvable problem. But solving it requires treating governance not as a constraint on AI ambition, but as the infrastructure that makes AI ambition sustainable.
Key Takeaways for IT Leaders
- Governance maturity is the strongest predictor of agentic AI success — not model selection, not compute budget, not team size.
- The scaling-without-governance window is closing. Regulatory frameworks (NIST AI RMF, EU AI Act) are raising the bar, and the operational costs of retrofitting governance onto ungoverned agentic systems are non-trivial.
- The last-mile gap is a delivery architecture problem, not a data science problem. Inconsistent access by business unit is a governance and infrastructure failure.
- Financial services offers a replicable playbook. Governance-first organizations in regulated sectors are demonstrating measurably better outcomes — their frameworks are worth studying regardless of your industry.
- European enterprises face the sharpest urgency: highest regulatory exposure, lowest governance maturity, highest rate of ungoverned agentic scaling.
The ROI plateau is real. So is the path out of it — and it starts with governance, not with the next model release.
Published in Stream · Dispatch #462 · July 24, 2026 · 9 min read.
Reply to paolo@mont3.ch - every email gets a human answer within 24h.