Why 40% of agentic AI projects get canceled
TL;DR: Gartner forecast that more than 40% of agentic AI projects will be canceled by the end of 2027. The reasons were cost, unclear business value, and inadequate risk controls, not a weak model. S&P Global reported abandonment of most AI projects rising from 17% (2024) to 42% (2025). MIT NANDA found 60% of orgs evaluated custom GenAI, 20% piloted, 5% reached production. Cancel the agent loop when the work is a known path. Keep the model for writing the path. Run the path as a playbook.
What Gartner actually said
On 25 June 2025, Gartner published Over 40% of agentic AI projects will be canceled by end of 2027.
Three causes:
- Escalating costs
- Unclear business value
- Inadequate risk controls
Analyst Anushree Verma said most projects are early experiments driven by hype, and that many use cases sold as agentic do not need an agent. Gartner also warned about "agent washing": chatbots and RPA relabeled as agents. They estimated only about 130 of thousands of "agentic" vendors were real.
This is a forecast, not a 2026 body count. Searchers still type the headline. The rest of this page is what to do with it.
The other public numbers
| Source | Number | What it measures |
|---|---|---|
| Gartner, June 2025 | 40%+ canceled by end of 2027 | Forecast: agentic projects, cost / value / risk |
| S&P Global, 2025 | 42% abandoned most AI projects (was 17% in 2024) | Companies already investing in AI |
| MIT NANDA, 2025 | 60% evaluated, 20% piloted, 5% in production | Custom enterprise GenAI tools |
| τ-bench (Yao et al.) | Retail ~61% once, <25% if you need 8/8 | Customer-service agents vs a policy and a database |
| IBM / Ponemon 2025 | 13% had an AI app/model breach; 97% of those lacked AI access controls | Security, not model IQ |
These are not Ballet scores. They are why pilots die after a good demo.
Full Ballet lab tables (54 of 54 vs 42 of 54): Deterministic operations, measured. First-year cost of one workflow: Ballet vs Claude code production cost benchmarks.
Why AI agents fail in production
The model is rarely the first cause.
Cost. A pilot uses one prompt. Production adds retries, longer context, extra tools, and multi-agent fan-out. Token bills move without a matching KPI. Gartner named this first.
Unclear value. The demo writes a CRM update. Nobody defined "done": same stage rule, same owner, same time-to-first-touch. Next year's budget goes to the team with a completed-task rate.
Risk controls. IBM's 97% figure is among orgs that already had an AI-related breach: almost none had proper AI access controls. Credentials in chat, agents acting as admins, no owner vs runner.
Integration. The pilot read a CSV. Production must write Salesforce or HubSpot with least privilege. Salesforce's own MCP guidance: start read-only, use mcp_api, do not hand out the full Platform API. See Salesforce in Claude.
Ownership. Platform built it. The business asked for it. Monitoring belongs to both, so it belongs to nobody. Quality drifts until the incident that quietly defunds the project.
Agent washing. If a Zap or a chatbot was relabeled an agent, you paid agent cost for workflow work. Anthropic's own split: a workflow has a written path; an agent picks the path as it goes. Start with the workflow. Building effective agents.
What to pilot before you scale
- Name the KPI (time-to-route, refund accuracy, cost per 10,000 runs).
- Decide workflow vs agent. Known refund window: workflow. Open research: agent.
- Scope credentials. Integration user, not a pasted admin password.
- Replay 100 to 200 historical records. Same input, same output, eight times if you can.
- Put stage changes, sends, and refunds behind a human on the same run ID.
- Assign one owner for the process and one owner for model upgrades.
- Only then raise write access.
If the work is Salesforce-in-chat, use the beta checklist in Salesforce in Claude. If the work is lists and sequences, see How to automate lead routing with AI.
Where Ballet fits
Ballet is the workflow side of Gartner's split.
- Claude writes the playbook.
- The playbook runs without paying the model again on every ticket (unless you add an AI step on purpose).
- Cost is a run, not an unbounded agent loop. Year-one example: $3,303 vs $13,873 to rebuild one Freshdesk reply on Claude. Cost benchmarks.
- Risk controls have a home: who edits, who runs, where secrets live, what the run log shows.
Longer argument: Claude writes the playbook. Ballet runs it..
FAQ
Why do AI agents fail in production? Cost, no KPI, weak access control, messy integration, and no owner. Model quality is usually not the first item on the postmortem.
What is agent washing? Selling a chatbot, RPA bot, or Zap as an autonomous agent.
Should you cancel the agent project? Cancel the unbounded loop on known work. Keep the model to write the path. Run the path.
What should you pilot first? Read-only CRM or helpdesk. 100 historical records. One KPI. Approval on writes.
Related articles
- Salesforce in Claude: what shipped, what it can write, what it cannot govern
- How to automate lead routing with AI (HubSpot and MarkOps)
- n8n vs Zapier vs Ballet for AI agents (2026)
- Claude writes the playbook. Ballet runs it.
- Deterministic operations, measured
- Ballet vs Claude code production cost benchmarks
- Why Ballet
Sources
- Gartner, Over 40% of agentic AI projects will be canceled by end of 2027 (25 June 2025)
- S&P Global, reported in CFO Dive
- MIT NANDA, The GenAI Divide
- Yao, Shinn, et al., τ-bench
- Anthropic, Building effective agents
- IBM / Ponemon, Cost of a Data Breach Report 2025