Over 40% of Enterprise AI Agent Projects Will Be Cancelled: Check Whether Yours is One

Agentic AI projects following contrasting paths towards successful deployment or cancellation due to rising costs, unclear value and late risk controls.

Gartner forecasts that up to 40% of enterprise applications will include task-specific AI agents by the end of 2026, and that over 40% of agentic AI projects will be cancelled by the end of 2027. Both are probably right, and the space between them is the most useful thing said about enterprise AI agents this year.

Two Gartner forecasts, placed side by side, look like a contradiction.

The first, published in August 2025: up to 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from fewer than 5% in 2025. The second, from June 2025: more than 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value and inadequate risk controls.

They are not in conflict. They describe different populations. The first is about features shipped inside products that vendors build once and sell many times. The second covers agentic AI projects in general, and in our reading it bites hardest on the bespoke projects that individual organisations attempt. Rapid diffusion of a capability alongside a high failure rate among individual implementations is the ordinary signature of a technology in its scaling phase.

What are enterprise AI agents?

Enterprise AI agents are AI systems that carry out multi-step business tasks with limited human input, using an organisation's own data, tools and systems. Gartner separates them from AI assistants, which help with a task but depend on a person to drive it. A task-specific agent can take a piece of work from start to finish, such as monitoring network activity and starting a response to a security threat.

They reach organisations in two ways: as features built into the software they already buy, and as bespoke agents built for an individual organisation. That distinction is why the two Gartner forecasts can both be right.

Is enterprise AI agent adoption real, and uneven?

McKinsey's State of AI 2026 survey, published in August and based on 1,719 respondents, describes a two-speed race. Among organisations with more than $1 billion in annual revenue, 40% report they are now scaling AI agents, up from 27% a year earlier. Among smaller organisations the figure is 22%, essentially unchanged. A year earlier, McKinsey's 2025 survey found 23% of all respondents scaling an agentic system somewhere in the enterprise, with a further 39% still experimenting.

So the demand is genuine and the movement from pilot to production is happening. The open question is what proportion survives contact with a profit-and-loss account.

Why do enterprise AI agent projects get cancelled?

Gartner gives three causes, and each deserves to be taken seriously:

  • Escalating costs. Agent loops consume tokens non-linearly. A workflow that costs pennies as a single call can cost pounds when an agent plans, calls tools, evaluates and retries. OpenAI offers an extreme example: by mid-August 2026 its agents were logging 3.1 agent-workdays for every human workday across its research organisation, and the median researcher was spending more than $600 a day on inference at API prices, according to The New Stack. That is a frontier AI lab rather than a typical business, but the pattern holds at smaller scale. Most business cases are built on per-call arithmetic and then meet per-task reality.
  • Unclear business value. The familiar problem in a new costume: capability acquired before a measurable baseline was identified. If you cannot state what the process costs today, you cannot demonstrate that the agent improved it. Our analysis of AI document processing returns shows what a measurable baseline looks like in practice.
  • Inadequate risk controls. The most expensive of the three, because it tends to surface late. A system that works is stopped at the security, legal or compliance gate after the money has been spent, not before. Our breakdown of the SalesBleed vulnerabilities shows the kind of risk that gate exists to catch.

Is agent washing part of the explanation?

Gartner also points to a supply-side problem. Vendors have rebranded existing products, such as AI assistants, robotic process automation and chatbots, as agentic without substantive agentic capability. Gartner's estimate is that only around 130 of the thousands of vendors claiming the label are the real thing.

That figure, if roughly right, explains a meaningful share of the projected cancellations on its own. A buyer selecting from a market where most labels are inaccurate will sometimes buy something that cannot do what the category name implies. The project is then cancelled, correctly, but the failure was in procurement, not in the technology.

Handing the build to a vendor does not remove the risk either. On 29 September 2026, Gartner predicted that 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineers by 2028, because they end up with soaring costs and no ability to evolve the system themselves. Its advice is to agree knowledge transfer, intellectual property and exit terms before signing.

A cancellation rate is not a verdict on a technology. It is a measurement of the gap between what was promised, what was bought, and what was actually needed.

What's the honest counter-argument?

It is worth resisting the temptation to read the cancellation figure as proof that agents do not work.

Gartner's own framing is that most current agentic efforts are early-stage experiments or proofs of concept, mostly driven by hype and often misapplied. Cancelling an experiment that has answered its question is a success, not a failure. The statistic is only alarming to organisations that were treating experiments as commitments, which is a governance problem rather than a technical one. Our CTO's piece on moving AI from proof of concept to production covers what it takes to turn one into the other.

There is also a reasonable argument that the forecast is self-defeating in the useful sense: published widely enough, it changes the behaviour it predicts, because boards start asking harder questions before approving the next pilot.

What separates the enterprise AI agent projects that survive?

1A bounded scope with a measured baseline

The strongest returns continue to come from high-volume, repetitive work with an unambiguous success criterion and a known current cost. That is unglamorous, and it is where the evidence points.

2Cost modelled per completed task

Not per call, and not per token. Model the loop, including retries and failures, and set a ceiling that stops a runaway before it reaches the invoice. Our piece on LLM pricing makes the same case for knowing what one unit of delivered work costs.

3Risk controls designed in, not appended

Bring security, legal and compliance into the design conversation. The gate they represent is cheaper to pass in week two than in month nine.

4Human checkpoints where being wrong is expensive

Full autonomy is a choice to be made per decision type, on the basis of consequence, rather than adopted wholesale because the technology permits it.

What's the honest summary?

The two forecasts together describe a market where the capability is arriving regardless of what any individual organisation decides, and where a large share of deliberate attempts to build with it will be abandoned.

The differentiator is not ambition or access to the technology. Both are widely available. It is the discipline of scoping to a measurable problem, costing the loop rather than the call, and involving the people who can stop the project before the money is spent rather than after.

None of that is novel advice. It is the same advice that applied to every previous wave of enterprise software, which is precisely why it keeps being ignored.

Figures are quoted in US dollars as published in the underlying research.

Deciding which agent projects to fund?

Before approving the next pilot, check that it has all four of these:

  1. A measured baseline: what the process costs and how long it takes today.
  2. A cost per completed task: including retries and failures, with a ceiling.
  3. Risk sign-off at design stage: security, legal and compliance involved from the start.
  4. Named human checkpoints: for every decision where being wrong is expensive.

If any of the four is missing, fix that first. Our free AI Impact Audit is a starting point for identifying where AI can deliver measurable value in your business.

Talk to us about your agent project

Frequently asked questions

What are enterprise AI agents?

Enterprise AI agents are AI systems that carry out multi-step business tasks with limited human input, using an organisation's own data, tools and systems. They arrive either as features built into enterprise software or as bespoke agents built for a specific organisation. Unlike AI assistants, which depend on a person to drive each step, a task-specific agent can complete a piece of work from start to finish.

Do Gartner's two agentic AI forecasts contradict each other?

No. The 40% enterprise-app figure describes agent features built into vendor products and sold at scale. The cancellation figure covers agentic AI projects in general, and in our reading it falls hardest on bespoke builds for individual organisations. Rapid feature diffusion alongside a high failure rate in individual implementations is normal for a technology in its scaling phase.

Why are agentic AI projects getting cancelled?

Gartner points to three causes: escalating costs (agent loops consume tokens non-linearly, so per-call cost estimates badly understate per-task reality), unclear business value (no measurable baseline existed before the project started), and inadequate risk controls (the project gets stopped at a security, legal or compliance gate after the budget is already spent).

What is agent washing?

It is vendors rebranding existing, non-agentic products, such as AI assistants, robotic process automation and chatbots, with the agentic label without substantive agentic capability. Gartner estimates only around 130 of the thousands of vendors making this claim actually have the real capability, which means some project cancellations are really procurement failures rather than a sign the technology does not work.

Is a 40% cancellation rate actually bad?

Not necessarily. Gartner's own framing is that most current agentic efforts are early-stage experiments or proofs of concept. Cancelling an experiment that has already answered its question is a successful outcome, not a failed one. The rate is only a problem for organisations that treated experiments as commitments.

What do successful enterprise AI agent projects have in common?

Four things: a bounded scope with a known, measured baseline cost; cost modelled per completed task rather than per call or token; risk controls (security, legal, compliance) designed in from the start rather than added later; and human checkpoints placed specifically where a wrong decision would be expensive.

Sources

Have a project in mind? Let’s get to work.

Let’s chat about how we can help you. Fill in the details and we’ll get back to you as soon we can.