Why 40% of agentic AI projects will be canceled, and the data foundation that prevents it

The mood around agentic AI has shifted from unbounded optimism to hard questions about return, and the numbers now driving boardroom conversations are sobering. Gartner has predicted that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Around the same time, an MIT study of enterprise deployments found that roughly 95% of generative AI pilots delivered no measurable impact on profit and loss. Two very different research efforts arrived at the same uncomfortable place: most of what is being built does not pay for itself. For anyone deploying agents against real commercial decisions, the useful question is not whether the technology works, but what separates the projects that survive from the ones that get quietly switched off. A large part of that answer sits in the data these systems are given to reason over.

What Gartner and MIT actually found

It is worth being precise about the two findings, because they are often flattened into a single "AI is overhyped" headline. Gartner's forecast, published in mid-2025, is specifically about agentic AI: systems that are meant to plan and act with some autonomy rather than simply answer prompts. Its analysts point out that most current efforts are early-stage experiments or proofs of concept driven by hype, and that today's models often lack the maturity to pursue complex goals reliably over time. The MIT work, from its NANDA initiative, looked more broadly at generative AI in business and found that despite very large investment, the overwhelming majority of pilots never crossed into measurable value. The report frames this as a "learning gap": generic tools impress individuals but stall inside organisations because they do not adapt to real workflows and real context. Both findings point past the model itself and toward everything around it.

Agent washing and the return-on-investment gap

Part of the cancellation wave is simply the market correcting for exaggeration. Gartner has described a pattern of "agent washing," where existing chatbots, robotic process automation, and assistants are rebranded as agents without any genuine change in capability. When the label promises autonomy but the product delivers a scripted workflow, the pilot underwhelms and the budget disappears. Strip out the rebadged products and the field of genuinely agentic offerings is far smaller than the vendor count suggests. That matters for buyers because the failures are not evenly distributed. Projects tend to collapse where the value was assumed rather than demonstrated, where costs compounded quietly across model calls and integration work, and where nobody could point to a decision that measurably improved. The projects that endure are the ones wired into a specific, repeatable decision with a clear before-and-after, and those decisions almost always depend on trustworthy external context that the model does not hold on its own.

Why the data foundation decides which projects survive

A large language model brings reasoning and language, but it does not know your market, your neighbourhoods, or the state of the world this quarter. When an agent has to decide which customers to prioritise, how to price for a region, or whether an application looks risky, it needs grounded facts, not just fluent text. If that context is missing, the agent guesses, and confident guesses are exactly what produce the "inadequate risk controls" Gartner warns about. If the context is present but stale, the agent reasons correctly over a picture of Britain that is two or three years out of date, which is its own quiet failure mode. And if the context is individual-level personal data assembled without a clear lawful basis, the project inherits compliance exposure that can stop it before it scales. The through-line across the failed pilots is not weak models. It is thin, stale, or legally fragile grounding.

This is where a structured context layer earns its place. Area-level geodemographic data describes neighbourhoods rather than named individuals, so it grounds an agent's reasoning without carrying the consent and ownership risk that individual-level enrichment brings. Because it is refreshed continuously rather than on the annual or biennial cycles typical of legacy providers, it reflects the world the agent is actually operating in. And because it is delivered in a consistent, machine-readable structure, an agent can consume it reliably instead of scraping it back together from inconsistent sources.

What a structured context layer looks like in practice

Concretely, the grounding an agent needs to move from demo to production tends to share a few properties. A dependable context layer is:

• Broad enough to answer real questions: Cogstrata derives 5,000+ attributes and a 24-group, 8-supergroup classification, so an agent can find the specific signal a decision needs rather than a handful of coarse categories.
• Current by design: attributes are refreshed continuously, so the data an agent reasons over reflects the present, not a snapshot from a previous update cycle.
• Privacy-safe at the source: because it describes areas rather than individuals, it is GDPR-safe by construction, which keeps the risk controls that kill projects out of the critical path.
• Structured for consumption: delivered through API, batch, and marketplace routes in a consistent shape, so both agents and human teams draw on the same grounded facts.

None of this makes a weak use case strong. It does remove the failure modes that have nothing to do with the use case: guessing where there should be grounding, reasoning over stale inputs, and carrying avoidable compliance risk. Those are the failures the research keeps surfacing, and they are the ones a good data foundation is built to prevent.

Making agentic projects survive 2027

The Gartner and MIT findings are a warning, but they are also a fairly practical checklist. Tie each agent to a decision you can measure. Be honest about whether a "smart" workflow is genuinely agentic or simply automation with a new label. Watch the compounding costs, and put risk controls in early rather than bolting them on when a regulator or a customer asks. And treat the data foundation as a first-order design choice, because an agent is only as good as the context it can trust. The organisations still running their agents in 2028 will not necessarily be the ones with the largest models. They will be the ones that grounded those models in current, structured, privacy-safe data and pointed them at decisions that actually moved. For more on this, see Why your AI agent needs curated data, not just more data and The trust layer: why agent-curated data outperforms traditional pipelines.