Budget the Cancellation: 40% of Agentic AI Projects Won't Survive 2027

Most 2027 AI budgets now carry a line for agents. Very few carry a line for what happens if the agent gets cancelled — and Gartner expects that to be the fate of more than 40% of agentic AI projects by the end of 2027. If two in five of something are expected to fail, the cancellation belongs in the budget before the build does.

What the forecasts and the surveys say

Gartner's June 2025 prediction was specific about the causes: escalating costs, unclear business value, or inadequate risk controls. It also put a number on the vendor noise. Of the thousands of companies selling "agentic" products, Gartner estimated only about 130 offer genuine agentic capability; the rest are rebranded assistants, chatbots, and automation tools, a practice it called agent washing. In the same release, a Gartner poll of 3,412 webinar attendees found 19% of organisations had made significant agentic investments, 42% conservative ones, 8% none, and 31% were waiting to see.

Fifteen months later, the survey data shows the pullback already under way. KPMG's Q2 2026 Global AI Pulse surveyed 2,145 senior leaders at organisations with at least $50 million in revenue across 20 countries, fielded from April 28 to May 25, 2026. When the expected costs of agents began to outweigh the value, 49% had rephased their deployments: 24% scaled back or narrowed them and 22% delayed or paused further rollout.

The cost side is moving in the wrong direction for anyone who budgeted off a pilot. In August 2026, Gartner predicted that inference costs per agentic workflow will increase more than fivefold through 2028, even as the price per token keeps falling. Routing a task to an agentic reasoning model costs a provider at least five times what a basic chatbot exchange costs, and often much more as the task grows complex. Cheaper tokens, in other words, are being spent on far more of them.

And on September 29, Gartner added a prediction aimed squarely at the build-it-for-me model: 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineering by 2028, mostly because operating costs climb and the company never builds the internal capability to run what the vendor left behind. Fewer than 20% of those engagements, Gartner expects, will turn into permanent product capability.

The cancellation budget: a worked calculation

Put those findings together and the useful number for a mid-market operator is an expected write-off: what a single agent project is likely to cost you in the scenario where it does not survive.

Expected write-off = probability of cancellation × (build cost + monthly run cost × months until someone decides to stop)

Here is the calculation with illustrative inputs. The 40% comes from Gartner's forecast and the 5× multiplier from Gartner's inference analysis; the dollar figures are our assumptions for a company in the $50–$200 million range, not benchmarks. Replace them with your own.

• Build cost: $150,000, whether paid to a vendor or absorbed internally.
• Run cost: the pilot ran at $2,000 a month on chatbot-style usage. Production routes work through agentic reasoning, so apply the 5× multiplier: $10,000 a month.
• Months until someone decides to stop: nine, which is typical when nobody wrote a kill criterion and the decision waits for the next budget cycle.

Expected write-off: 0.40 × ($150,000 + $10,000 × 9) = 0.40 × $240,000 = $96,000. On these inputs, every dollar of planned build carries about 64 cents of expected loss that never appears in the budget.

Now change one input. With a written kill criterion reviewed at month three, the same project's expected write-off falls to 0.40 × ($150,000 + $30,000) = $72,000. The probability of failure hasn't moved; the cost of finding out has.

Two caveats. This combines an enterprise-wide forecast with cost assumptions of our own, so treat it as directional rather than precise. And it assumes the cancellation probability is fixed, when the KPMG data suggests it is not.

The odds you can actually move

KPMG's same Q2 survey found that only 13% of organisations say their AI operating costs are fully visible and actively monitored. Those that have that visibility report established ROI at 15%, against 3% for those without it. Organisations with clear CEO accountability for AI report established ROI at 14%, against 4% elsewhere. KPMG's Q3 2026 Pulse (2,131 leaders, fielded July 23 to August 26) points the same way: 86% of organisations with established returns run a formal cross-functional or enterprise-wide AI management layer, against 31% of those still experimenting.

Visibility, ownership, and a management layer are the variables that separate the 49% who pulled back from the minority who are banking returns. None of them is a technology purchase. All of them can be in place before the first agent is built. That is the argument we made about who owns the 2027 AI budget, applied to one line of it.

Four lines to write next to the agent line

1. A monthly cost ceiling per workflow. Run cost moves after launch. Gartner's fivefold forecast means the pilot invoice is the floor, not the estimate.
2. A written kill criterion and a review date. In the calculation above, it was worth $24,000 of expected loss on a single project, and it costs nothing but a decision.
3. A named owner with authority over the P&L line. The 14% versus 4% split is the strongest single signal in the KPMG data.
4. An exit plan if a vendor builds it. Gartner's 70% abandonment forecast is about companies that rented the build and never owned the operation. Decide before signing who runs it in year two.

A quick test for agent washing

If Gartner is right that only around 130 vendors sell genuine agentic capability, most of the pitches a mid-market company hears this year will be something else wearing the label. Three questions sort them quickly. First, what decision does the system make on its own, and what happens when it is wrong? A product that only drafts or summarises for a human to approve is an assistant, which may be exactly what you need, at a fraction of agentic run cost. Second, what does a month of production usage cost at your volume, not the pilot's? A vendor who cannot answer has not run it at scale. Third, who at your company will be able to change the workflow after the vendor's engineers leave? If the honest answer is nobody, you are buying into the 70% Gartner expects to walk away.

None of those questions needs technical depth to ask. They need someone in the room whose job is to ask them, which is the gap most mid-market companies have and few budgets name.

What this is not

This is not an argument against agents. Gartner also expects at least 15% of day-to-day work decisions to be made autonomously through agentic AI by 2028, up from zero in 2024, and a third of enterprise software to include agentic capability. Agents that survive will matter. The point is that fewer bets, understood deeply, beat a portfolio of pilots — the same reason we run a handful of executive searches thoroughly instead of a hundred at once.

That is also where most mid-market companies are thinnest. The pilots that stall rarely fail on the model; they fail on the missing owner and the missing cost discipline described in The AI ROI Gap. Our diagnostic-first approach exists to put those four lines in place before money moves. At $9,500, the AI Opportunity Diagnostic costs roughly a tenth of the illustrative expected write-off above, and it also answers whether you need a fractional Chief AI Officer to own the work or simply a tighter brief for the team you have — the distinction we drew in AI consultant vs. fractional head of AI.