Gartner says 40% of agentic projects get cancelled. Here is the pattern.

Across the engagements we've reviewed, cancellations cluster around five causes. They compound, meaning most cancelled projects have three or more of these stacked on top of each other.

1. No named owner. The project reports to a committee, a working group, or "the AI team" collectively. When something breaks, nobody has the authority to cut scope or kill it, so it drifts until someone above the committee finally notices the spend.
2. No cost ceiling. Nobody wrote down what a task is allowed to cost. Token spend creeps upward as the prompts grow and the agent retries failed steps, and eventually finance asks a question nobody can answer with a number.
3. No eval suite. Without a regression baseline, every fix is a guess and every demo is anecdotal. Six months in, nobody can say whether the agent is actually getting better or just different.
4. Agent washing. The project was scoped as a deterministic workflow, then rebranded as an "agent" to fit a budget category or an executive mandate. It gets judged against agentic expectations it was never built to meet.
5. Wrong use case for an agent at all. Some tasks are better solved with a script, a form, or a lookup table. Wrapping one in an LLM adds latency, cost, and nondeterminism with no offsetting benefit.

This is the part that surprises executives the most. Ownership, cost ceilings, eval design, and use-case fit are all decisions made in the kickoff meeting, not the engineering sprint. By the time a developer opens an editor, the project's fate is largely set.

We've watched two teams build functionally identical customer-support agents. One had a named owner, a documented $0.40-per-ticket ceiling, and twenty labeled example tickets before week one ended. The other had a Slack channel, a rough budget number nobody would commit to in writing, and a demo scheduled for the following month. The first shipped. The second is one of the cancellations in Gartner's number, and the model underneath was the same one.

This is also why a rescue engagement usually starts with a scoping conversation, not a code review. Our AI Audit exists specifically to surface these gaps in 48 hours, before the sunk cost gets any larger.

Agent washing deserves its own section because it's the cause we see most often in mid-market companies specifically. It happens for a boring reason: agentic AI has budget attention right now, and a proposal labeled "customer support agent" gets funded faster than one labeled "workflow automation."

The problem shows up six months later, when leadership expects the system to handle judgment calls, exceptions, and multi-step reasoning, because that's what "agent" implies. If the underlying build is actually a rules engine with an LLM bolted on for text generation, it can't do any of that, and it gets cancelled for "underdelivering" against expectations it was never designed to meet.

The fix isn't complicated. Be honest about what the system actually is at kickoff. If it's deterministic, call it automation and measure it like automation: throughput, error rate, cost per unit. If it genuinely needs judgment across ambiguous inputs, then it's agentic, and it needs the harness, evals, and escalation design that comes with that. Our agents vs agentic AI explainer covers this distinction in more depth, because the label decides how the project gets funded, staffed, and judged.

Before committing budget to a new agentic project, run this checklist. We use a version of it on every intake call.

• Owner. One named person accountable for the go/no-go decision, not a committee.
• Cost ceiling. A written dollar figure per task and per month, reviewed weekly against actual spend.
• Ten to twenty real examples. Actual cases from your business, with the expected correct output written down, before any prompt is drafted.
• Judgment test. Can this task be solved with a deterministic script or lookup table? If yes, it isn't an agentic project, and pretending otherwise sets up the wrong expectations.
• Kill criteria. A written threshold, agreed before launch, for what "not working" looks like and who has authority to pull the plug.
• Eval baseline. A regression suite built from the ten to twenty examples above, run before every change ships.

Any project missing three or more of these is a strong cancellation candidate, regardless of the model or framework chosen.

The common read on Gartner's number is that agentic AI overpromised and the technology needs another cycle to mature. That take is comfortable because it puts the blame on the model vendors and buys everyone another year of experimentation without changing how projects get scoped.

It's wrong in a specific way: it treats a management failure as a technology failure. The models involved in most of the cancelled projects we've reviewed were perfectly capable of the underlying task. What was missing was ownership, a cost ceiling, and an honest label for what was being built. None of that improves with a better model. It improves with better project discipline, and that discipline is available today, not in some future model release.

If you want the more measured version of the failure-rate debate, including how MIT's, Gartner's, and vendor-reported numbers actually differ, see our companion piece on why AI pilots never ship.

If your agentic project already has some of these gaps and is six months in, the fix isn't to start over. It's to retrofit the missing discipline onto the smallest viable slice of the project:

1. Name one owner this week, even if it feels like taking territory from a committee.
2. Pull real production logs and build a twenty-case eval suite from what actually happened, not hypothetical scenarios.
3. Write the cost ceiling down and compare it against last month's actual spend. The gap, if there is one, tells you how much runway is left before finance intervenes anyway.
4. Decide honestly whether the task needs agentic judgment or a script would do. If it's the latter, cutting scope down to automation is a save, not a failure.

This is close to the same sequence we run in a pilot-to-production engagement: triage first, harden second, ship one narrow thing third.

Gartner's 40% is a real number, but it's measuring management decisions more than model capability. Ownership, cost ceilings, honest labeling, and eval discipline are cheap to install in week one and expensive to retrofit in month six. If your project has all four in writing, you are already outside the population most likely to get cancelled.