2 out of 100. That's the survival rate for enterprise AI projects after 18 months. Not a typo. New production data from across the industry shows that 88% of AI agent pilots never reach production at all.
We've been building AI implementations long enough to see the pattern. It's never the model. It's never the data. It's always the same seven things teams skip because they seemed optional at the time. They're not optional.
The Problem: Why AI Pilots Die
The numbers paint a clear picture. The top causes of AI project failure are remarkably consistent:
- 64% of failed projects never specified success criteria
- 61% built no evaluation set to measure performance against
- 52% had brittle tool boundaries that broke under real-world conditions
- Silent failures—agents that run without errors but produce useless output—are the fastest-growing failure mode
Here's what's happening in most organizations: a team gets excited about an AI use case, builds a demo that works on 10 carefully chosen examples, gets budget approval, and then watches the project slowly die over the next 6-12 months as it encounters real-world messiness.
The PwC survey from this week puts the executive perspective in sharp relief: 88% of companies use AI, but only 39% see actual EBIT impact. That 49-point gap isn't a technology problem. It's a discipline problem.
The Solution: The 7-Step Pre-Flight Checklist
After analyzing dozens of production deployments (both successful and failed), here are the seven steps that separate the 2% from the 98%:
1. Define Success Before You Write a Single Prompt
What does "done" look like? Not "improve customer service"—that's a wish. Something like: "resolve 85% of tier-1 support tickets without human escalation, with a customer satisfaction score above 4.2, in under 3 minutes."
If you can't measure it, you can't ship it. 64% of failed projects skip this step.
2. Build Your Evaluation Set First
Before you choose a model, before you design an architecture, build a golden test set of inputs and expected outputs. This is your ground truth. Every model change, every prompt iteration, every architecture decision gets measured against this set.
61% of failed projects never built one. They optimized for demos instead of outcomes.
3. Set Hard Tool Boundaries
Define exactly what your agent can and cannot do. Can it access the production database? Can it send emails? Can it make purchases? Can it modify customer records?
52% of failures trace back to brittle tool boundaries that worked in testing but broke under real-world complexity.
4. Implement Cost Budgets on Day One
Token costs are the silent killer of AI projects. Uber burned through their entire 2026 AI budget in 4 months. One enterprise client hit $500M in a single month. Consumption-based pricing overshoots budgets by 40% on average.
Set per-task, per-day, and per-week token limits. Cost discipline is project discipline.
5. Add a Verification Layer
Every agent output should be checked before it's acted upon. This can be a simpler, cheaper model verifying a complex one. It can be a rules engine. For high-stakes decisions, it can be a human checkpoint.
The "model + harness = agent" framework from Meta/Stanford's recent review paper confirms this: the harness is the product. Your verification layer is the most important part of that harness.
6. Plan for Scope Drift
Agents expand their mandate over time. A support bot starts modifying account settings. A data analysis agent starts making business recommendations. A coding agent starts refactoring systems it wasn't asked to touch.
Build explicit scope enforcement. When an agent exceeds its defined domain, it should escalate or terminate—not improvise.
7. Measure Impact, Not Usage
The 49-point gap between AI adoption (88%) and EBIT impact (39%) exists because most organizations measure whether people use AI, not whether it helps.
Define your impact metrics upfront. Track them weekly. If you can't connect your AI implementation to a business outcome within 90 days, something is fundamentally wrong with the approach.
Benchmarks: What This Checklist Delivers
Data from teams that follow these principles:
- 3x higher pilot-to-production conversion rate compared to industry average
- 70% fewer production incidents from agent failures
- 60-80% lower token costs through budgets and scoping
- 90-day time-to-impact for teams with predefined success criteria vs. 6-12 months for teams without
- Caveat: These numbers assume competent engineering teams. The checklist doesn't create capability—it structures it.
- Caveat: Steps 1-3 add 2-4 weeks to project kickoff. This feels slow. It's actually faster than a 6-month death spiral.
The 12-factor-agents framework (22K+ GitHub stars) aligns closely with these principles and is worth studying for engineering teams building production systems.
Impact: The ROI of Doing It Right
Let's be concrete about the business case.
The average enterprise spends $15-50M on AI in 2026. With an 88% pilot failure rate and a 49-point adoption-to-impact gap, most of that spend generates zero returns. That's not a budget problem—it's a process problem.
Teams that implement this checklist consistently see:
- Faster time to value: 90 days vs. 6-12 months
- Higher success rates: 3x the industry average for pilot-to-production conversion
- Lower total cost: 60-80% savings on token costs alone
- Real ROI: Measurable business outcomes within the first quarter
For every 10 hours of AI efficiency that most companies gain, 4 hours are lost fixing output. The checklist doesn't just prevent failures—it eliminates that 40% waste.
The Hard Truth
Most AI projects fail because teams treat implementation like experimentation. They start building before they know what success looks like. They skip evaluation because it feels like paperwork. They ignore cost controls because the pilot is "just a test."
None of these are technology problems. They're discipline problems. And they're completely preventable.
The 2% of teams that succeed don't have better models or more data. They have better processes. They do the boring work upfront so the exciting work actually ships.
Your AI pilot doesn't have to die. But it will if you skip the checklist.