Trust and value are the missing layers in enterprise AI

The most expensive line item in enterprise AI right now is not compute. It is the gap between what AI promised to do and what it is actually delivering.

 

After two years of pilots and proofs-of-concept, boards have stopped asking what AI can do and started asking what it is delivering. The shift sounds subtle. It is not. It is the start of a recalibration that will reshape how large organisations think about their AI investments.

 

By 2027, Gartner forecasts, 80 per cent of data and analytics leaders will be calculating and communicating new AI value expectations and initiative trajectories as a result of setbacks and AI fatigue. 80 per cent. Effectively a market-wide reset. The companies that come out of that reset stronger will not be the ones that ran the most experiments. They will be the ones who learned to enforce trust in their data and translate it into measurable value before the audit, not after it.

 

The setbacks are not technical

 

When boards talk about AI failure, they typically describe it as a model problem. Closer inspection almost always tells a different story. Models drift because the data they consume is not maintained. Pilots stay in pilot because compliance teams cannot trace which decisions were made, by which agent, on which version of which data set. And the quietest failure of all is adoption: the people meant to use the system do not, because they cannot see where its outputs came from, who is accountable or whether to trust them. None of those are issues the model can fix. They are governance problems wearing technical clothes.

 

This is what makes the next 18 months structurally different. The pressure on AI is no longer just to demonstrate capability. It is to demonstrate accountability. Insurance carriers automating claims, manufacturers running predictive maintenance on critical assets, technology firms embedding agents into customer-facing products, banks deploying decision systems for credit and fraud. Every one of them is operating in a market where trust and value are now graded on the same scorecard.

 

Trust has stopped being a checkbox

 

The EU AI Act is the most visible illustration of this shift as it places liability on the deployer, not the developer. For high-risk systems, it expects organisations to be able to show what data trained an agent, what decisions it made, with what level of confidence and under whose accountability. The UK’s evolving sectoral approach is moving in a parallel direction. Regulators across financial services, healthcare and critical infrastructure are converging on the same expectation: organisations must be explainable and accountable for their AI deployments.

 

The implication for any large enterprise is that trust cannot be added at the end. It has to be built into the foundation. That means treating data lineage, quality and ownership as production-grade infrastructure rather than as documentation chores. It means giving every business-critical data asset a known owner, a defined definition and a continuous quality signal. And it means doing the same for the AI initiatives that consume them: recording what each one does, who is accountable for it and how it behaves when conditions change.

 

Value is not assumed, it is engineered

 

The harder conversation, and the one driving the 2027 recalibration, is about value. Edosa Odaro, in his recent book The Values of Artificial Intelligence, argues that most enterprise AI failures are not capability problems but value problems: organisations measure adoption, not defensible outcomes, and the resulting fog only lifts when leaders make trade-offs explicit.

 

That diagnosis has a familiar shape. Most AI portfolios today are loose collections of initiatives, started by different teams, scored by different KPIs, with no shared view of what they cost or what they return. When boards ask what AI is actually returning, the honest answer in many organisations is that nobody knows.

 

Closing that gap requires the kind of discipline that has long been applied to capital projects, but rarely to AI: portfolio management. Each initiative needs a clearly stated business outcome it is meant to influence, a metric that ties it to that outcome, a reusable data or AI product and a review cadence that retires the ones that do not perform. The exercise is unfashionable. It is also the only thing that turns AI from a budget line item into a balance-sheet asset.

 

What makes this difficult is that the data layer and the value layer have historically lived in different parts of the organisation. Today, the Chief Data Officer owns the data stack (tech) and data governance (definitions and policies), the Chief AI Officer owns the models (tech and AI governance) and the business owns assets (including quality, initiatives and value). The link between them is too often missing. The 2027 recalibration will force that link to be built.

 

The companies that will quietly outperform

 

A pattern is emerging among enterprises that are pulling ahead of the AI velocity gap. They have stopped treating AI governance as a project and started treating it as part of their operating model. They have a single, agreed-upon catalogue of business-critical data, with owners and definitions. From that catalogue, every AI initiative inherits its inputs, known data assets, known owners and known governance. They map every AI initiative to a measurable outcome. They retire what does not work without ceremony. They are not necessarily the loudest about their AI ambitions; they are simply the ones quietly delivering on them.

 

The lesson the next two years will teach is that scale is not the reward of speed. It is the reward of trust expressed as infrastructure and value expressed as discipline. Trust without value is paranoia. Value without trust is reckless. The companies that compound their AI advantage will be the ones that learn to do both at once.

 

A reset, not a retreat

 

Four in five data and analytics leaders re-examining their AI value expectations is not a failure of AI. It is the conversation finally catching up with the reality of deployment. The organisations that emerge as winners will not have the largest model libraries. They will be the ones whose data and value layers are connected, and connected in a way that drives the link from tech to business and holds up not only to an auditor, but also the board.

 

That is the work of the next 18 months. It is less glamorous than launching a new agent. But it’s what will determine whether the next agent stays employed or hits early retirement.

---

To find out more, click here

---

By Nicolas Averseng, Chief Product Officer, DataGalaxy