Case studies22 Apr 2026 · 12 min read

Why 70% of AI-agent pilots die — 9 mistakes that kill deployments

Gartner says 30% of GenAI projects will be abandoned after pilot by end of 2025. Our breakdown of 14 client engagements that almost died — and what we did to ship them anyway.

In December 2024 Gartner predicted that "at least 30% of generative AI projects will be abandoned after proof of concept by end of 2025, due to poor data quality, inadequate risk controls, escalating costs or unclear business value." We see the same number in our practice — but not because of poor data quality. The real killers are organizational, not technical.

Here is the post-mortem of 14 pilots from Polish, German and US clients in 2023–2025. Three died. Two went into permanent zombie state ("we are still evaluating"). Nine shipped. The patterns are surprisingly consistent.

Mistake #1: pilot scope is too wide

A retailer asked us for "an AI agent that can handle all incoming questions." That's not a pilot, that's a 3-year roadmap. We narrowed it to "answer questions about order status and delivery times in Polish, escalate everything else." Pilot delivered in 5 weeks. Original brief would have taken 9 months and cost €120K.

tip
Pilot rule: one channel, one language, one intent. Resist scope creep with a one-page written agreement before kickoff.

Mistake #2: no baseline metrics

You can't prove an AI agent saved money if you don't know what was happening before. We had a logistics client who proudly reported "AI handles 60% of driver calls now." Great. What was the previous percent? Nobody had measured. The CFO killed the renewal.

Spend 1–2 weeks at the start measuring: call volume, average handle time, escalation rate, FCR, cost per ticket. Without these numbers, you have no story for the steering committee at month 6.

Mistake #3: bus factor of one

Three pilots stalled because the one engineer who understood the system left for another job. AI agents have implicit context: prompt engineering decisions, edge-case handlers, monitoring rules. Document everything in a runbook from week 1.

Mistake #4: synthetic data instead of production logs

Engineering team tested with 50 synthetic call transcripts they wrote themselves. Production calls had 4× more interruptions, accents the team didn't anticipate, background noise from open-plan offices, and 7% of callers using mixed languages (Polish + English). Accuracy dropped from 94% in tests to 71% in production.

warning
Always test on at least 200 real production samples before launch. Anonymize, but don't synthesize.

Mistake #5: no fallback path

When the LLM is down, the agent silent-fails. We had a Friday evening incident where Claude API rate-limited us for 23 minutes. Without a fallback to human, the bot just hung up on customers. 240 lost calls.

Every production agent needs three fallbacks: a) retry with another model, b) graceful "let me transfer you" with hold music, c) post-incident SMS to follow up.

Mistake #6: paying for consulting, not for code

A vendor charged €40K for a "discovery phase" — 60-page document with personas, journey maps, and a roadmap. Zero code. Six months later the client had nothing in production. We rebuilt it in 5 weeks for €18K including the actual agent.

Discovery has value, but cap it at 10–15% of total budget. If a vendor wants more for "research," they're selling a deck, not a deployment.

Mistake #7: no human-in-the-loop

A pilot in healthcare auto-cancelled appointments based on AI interpretation of patient messages. After two weeks the patient relations team was buried in complaints. The fix was a 2-line code change: any cancellation triggers a confirmation message, then a 30-min hold before action. No more complaints.

Mistake #8: ignoring edge cases

Edge cases in voice agents we've hit in production:

  • Caller hands phone to a child mid-conversation
  • Background TV news triggers "transfer to a human" intent
  • Caller speaks 3 languages in one sentence (PL + UA + RU)
  • Phone rings on speaker in a car — engine noise breaks STT
  • Caller hangs up mid-sentence; agent keeps talking to dead line
  • Called number was reassigned — wrong customer answers

You won't catch these in testing. You catch them in the first 2 weeks of production with active monitoring. Allocate 20% of your week-1 budget for triage.

Mistake #9: no measurement post-deploy

Three of our clients shipped agents and then... never looked. Six months later they couldn't answer if it was working. Always set up: error rate alerting, weekly transcript sampling (at least 50 calls), and a monthly steering review with one person accountable.

Pre-kickoff checklist

ItemStatusOwner
Single intent, single channel scope writtenRequiredProduct
Baseline metrics measured for 1+ weekRequiredOps
Two engineers familiar with the systemRequiredEng
200+ real production samples availableRequiredData
Fallback path defined and testedRequiredEng
Discovery budget capped at ≤15%RequiredCFO
HITL approval flow for sensitive actionsRequiredCompliance
Monitoring + weekly review scheduledRequiredOps

If you're already stuck

Most "stuck pilots" we adopt have one fixable cause — usually #2 (no baseline) or #5 (no fallback). Three weeks of focused work usually unstucks them. Send a 30-min screen recording of the agent failing — we'll tell you which of the 9 you hit.

Newsletter

Like this kind of breakdown?

Once a week — a short letter with numbers from our deployments.