A new study from Domino Data Lab should give pause to every business leader who’s been pouring money into AI tools without a clear plan for what happens after the model is deployed.
The Fifth Annual Domino Enterprise AI Report, published July 21 and based on an independent survey of 639 senior enterprise AI leaders at companies with at least $100 million in annual revenue, found that 57% of enterprises still can’t generate AI ROI that outpaces their spend. That number is exactly where it was in 2025. Two years of continued AI investment, and the majority of large organisations are still underwater on returns.
What makes this finding particularly striking: it coexists with genuine progress. A record 93% of those same leaders report improved AI production capability in 2026, up from 88% last year. Businesses are getting better at building and deploying AI. They’re just not closing the gap between deployment and actual business value.
The Last-Mile Problem
Domino calls this the “last-mile gap,” and it’s a useful framing. The distance between a model sitting in production and a business user actually doing something different because of it turns out to be enormous in most organisations.
The COO of Domino put it plainly: the real milestone isn’t deploying a model. It’s the moment a business user can act on what the model found. For most enterprises, that moment still isn’t happening at the pace or scale of the business.
This gap shows up clearly in the governance numbers. Companies that have implemented governed data products are dramatically more likely to have AI projects running in production: 85% of respondents with company-wide data products report three or more AI projects in production. That compares to just 25% of companies that lack that governed data infrastructure.
The message is hard to miss. You can’t shortcut your way to AI ROI by skipping the data foundation.
Governance Is Lagging Behind Agentic AI
The report also highlights a separate but related problem in the agentic AI space. Businesses are moving fast on autonomous agents, but governance hasn’t kept up. Only 43% of organisations have agentic AI running in a governed production environment. Meanwhile, 41% are piloting or scaling agentic AI without proper governance frameworks in place.
That’s a significant exposure for companies building agent-driven workflows. Agents that can take action, query systems, and surface decisions to business users carry a different risk profile than a chatbot. Running them without governance isn’t a temporary shortcut; it’s the kind of thing that creates costly mistakes and erodes trust in AI systems across the organisation.
What This Means for Business
If you’re in the 57% who aren’t seeing ROI that justifies your AI spend, the report points to some specific places to look:
Your data infrastructure probably isn’t ready. The gap between companies with governed data products and those without is too large to ignore. Three or more AI projects in production at 85% vs 25% is not a subtle difference. If your organisation doesn’t have a clear data governance story, you’re working against yourself before you’ve even started optimising the AI layer.
Deployment is not the same as value delivery. Moving a model to production is a technical milestone. Getting a business user to change their behaviour based on what that model surfaces is a business milestone. Most teams are measuring the first and hoping the second follows. It doesn’t happen automatically.
Agentic AI requires more governance, not less. The temptation when building autonomous agents is to move fast and clean it up later. The research suggests this is exactly backward. The organisations getting results from agentic AI are the ones who put governance infrastructure in place first.
Upskilling matters more than tooling. A capable AI tool in the hands of a team that doesn’t understand how to interpret and act on what it produces delivers roughly the same value as not having the tool. The last mile is almost always a human and skills problem, not a technology problem.
The EDNA Perspective
This is where the data literacy and AI literacy investments that Enterprise DNA has championed for years show up in the numbers. You can deploy the most sophisticated AI stack in the world, but if your team lacks the skills to interrogate outputs, identify where models are wrong, and turn AI findings into business decisions, you end up in the 57%.
The organisations pulling ahead aren’t just buying better software. They’re building teams that can actually use it. That means understanding data quality, knowing how to prompt and evaluate AI outputs, and having enough foundational data skills to trust the model when it’s right and catch it when it’s not.
The ROI problem documented in this report is, at its core, a data literacy and AI readiness problem. And that’s a solvable one.
If your organisation is wrestling with this gap, Enterprise DNA’s learning platform is built specifically for teams that need to close it. Or if you need help thinking through your AI strategy from a governance and deployment perspective, Omni Advisory can help you map a path from AI spend to AI ROI.