Will Your AI Agent Survive 2027? A One-Page Review

Many AI agent projects started in 2026. Fewer may still be running at the end of 2027. Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, because of cost, unclear value and weak risk controls. That is a forecast, not a count, and we read it as quoted in press coverage because Gartner's own page did not open for us.

The numbers we could check show the same tension: a lot of building, and a lot of tidying up still to do. This page gives you a one-page review you can run on any agent you use or plan to use, so yours is one of the ones that stays.

As of 2 November 2026:

• KPMG, Q3 2026 AI Pulse (314 US C-suite and business leaders, surveyed 24 July to 25 August): 62% said their organisation is building, deploying or developing AI agents, up from 53% a quarter earlier. 25% are developing multi-agent systems, up from 6%. 43% have set usage or token budgets, and 74% include cost reviews in AI approvals.
• IDC (two surveys: 903 respondents in February 2026 and 1,009 in March to April): the share of organisations running agent pilots fell from 17.8% to 12.8%. IDC also says firms expect to go from an average of 23.5 agent types in production in April 2026 to 43.2 by the end of the year. That second figure is an expectation, not a count.

Read together: many organisations are building, fewer are piloting, and cost reviews are spreading. "Building" is a wide word. It does not mean an agent does real work every day.

Our reading of the pattern in these sources is that an agent is easiest to cut when it lacks one of these four things:

1. No cost cap. Nobody knows what it costs per month, so the first big bill is a surprise.
2. No success number. Nobody can say what "working" means, so nobody can say it is not.
3. No named owner. When it breaks, nobody is responsible for it.
4. No stop rule. There is no agreed point at which you switch it off.

This list is our own summary, not a finding from any one study. Forrester's 2027 predictions (21 October 2026) point at related risks. It predicts that firms will uncover at least 10 employee-built "shadow" AI agents for every officially sanctioned one (from its Asia-Pacific release), and that an AI-generated bug or regression will cause a global outage across multiple industries. Both are predictions.

Copy this into a document. Fill one page per agent. Leave nothing blank; write "unknown" if you do not know, because an "unknown" is the thing to fix first.

Agent name:
Owner (one named person):
Task in one sentence:
Data it can read:
Actions it can take (read only / draft / send / spend):
Actions that need a human OK first:
Monthly cost cap (money or tokens):
Success number, measured weekly:
Stop rule (when we switch it off):
Last reviewed (date):

If you cannot fill in the owner, the cost cap and the stop rule, the agent is not ready to run on its own.

1. List your agents (10 minutes). Include automations, scheduled workflows, browser agents and coding agents. Count anything that acts without you clicking each step.
2. Fill one page each (15 minutes). Use the template above.
3. Set a cost cap (5 minutes). Pick a monthly number you would not mind losing. Check how your tool shows usage, and set its own limit if it has one.
4. Pick one success number (5 minutes). For example "replies drafted that I sent without edits" or "leads sorted correctly out of 20 checked by hand". One number you can measure each week.
5. Add a human OK for anything you cannot undo (5 minutes). Sending, paying, deleting and publishing all need your approval.
6. Put the next review date in your calendar (5 minutes). Monthly is a good start.

Forrester's prediction about shadow agents is a reminder to look. In a team, ask: who has built an automation, a custom assistant or a scheduled workflow, and which tools and accounts does it use? Check shared automation tools, scheduled tasks, API keys and browser extensions that act on a page for you.

Ask kindly. People who built something useful should be able to register it without being punished. The goal is one list, with an owner and a cost for each.

A stop rule is a sentence you agree on before the agent runs. Examples you can adapt:

• Pause it if it goes over the monthly cost cap in any month.
• Pause it if the weekly success number is below your target for 4 weeks in a row.
• Pause it after any irreversible action that was done without approval, and find out why.

These are our suggestions, not a standard. Writing one down matters more than the exact numbers. Related pages: AI Agents: Start with One Bounded, Reviewable Workflow, Set a Small Budget for an AI Coding-Agent Session and the lesson Stop an AI Automation From Burning Through Your Credits.

The KPMG survey covers 314 US leaders, so it does not describe every business or country. IDC's "expected" figure is a forecast. Gartner and Forrester items are predictions, and predictions can be wrong. We did not test any agent for this page, and the review template is a checklist, not a compliance or legal standard. If an agent touches personal data, customer money or health or legal decisions, ask a qualified person before it runs on its own.

Reviewed 2 November 2026. Figures were re-read on the pages below that day.

• KPMG: Q3 2026 AI Quarterly Pulse (24 September 2026)
• IDC: FutureScape 2027, The Great Rewrite (27 October 2026)
• Forrester: 2027 technology and security predictions (21 October 2026)
• Gartner: over 40% of agentic AI projects will be cancelled by end of 2027 (25 June 2025)