Anatomy of an AI Mandate: A Timeline
Most AI-adoption horror stories don’t start with a bad idea. They start with a good demo. Here’s how one played out, told the way it actually unfolded — week by week, from two sides of the same company that stopped trusting each other’s version of events.
The setup: A mid-sized fintech company. 80 engineers, a decade-old monolith, and a CEO who spent a weekend building a working prototype app entirely through AI prompts.
Week 0: The Demo
What leadership saw: A working app, generated from natural-language prompts, in a weekend. Clean code, functioning UI, done.
The conclusion drawn: If one person with no engineering background can produce this in 48 hours, an 80-person engineering org running the same tools should be unrecognizably faster.
The mandate that followed:
- All new code must be written using AI coding assistants — not encouraged, required
- Pull requests must show AI-assisted commits
- A new internal metric: “AI-generated code as % of total output,” reported weekly, visible on a dashboard up to board level
Weeks 1–3: It Actually Works (For a While)
What engineers experienced: For narrow, well-scoped tasks, the tools genuinely delivered. New API endpoints, migration scripts, isolated React components — all noticeably faster to ship than before.
What leadership saw: The AI-generated-code metric climbing every week. Velocity charts trending up. Confirmation of the thesis.
What nobody was tracking: Whether any of these fast, isolated pieces were actually built to fit together as one system.
Week 4: The Payments Team
A team began rebuilding a payments reconciliation service using AI-assisted development, as mandated.
What the AI produced:
- Working code that compiled cleanly
- Tests that passed — because the AI had also written the tests
- Currency rounding handled three different ways across three different code paths
- Timestamps silently assumed to be UTC, though two upstream systems weren’t
- A retry mechanism that could, under specific failure conditions, double-process a transaction
None of this was visible on a quick read. All of it looked, individually, like reasonable code.
Week 7: The Incident
A customer support ticket: a transaction charged twice.
Time to find the root cause: most of a week, spent by a senior engineer tracing the double-charge through a chain of AI-generated modules that each solved their own local problem correctly, but had never been designed with each other in mind.
Time to fix it, once found: one afternoon.
Week 8: Two Very Different Meetings
In the leadership meeting, the question was: “Why did this take four days if AI writes the code?”
The engineer’s honest answer: The code needed real architectural review. That review is slow, careful work — and generating code faster doesn’t make reviewing it faster.
How that answer landed: Not well. To a leadership team that had internalized “AI writes code instantly,” a four-day fix sounded less like an inherent cost of engineering and more like someone slow-walking the mandate.
In the engineering team’s private Slack channel, the question was: “Is the AI-code percentage metric being used to figure out who gets cut?”
How that suspicion spread: Some engineers started pasting boilerplate through the AI tool purely to inflate their reported percentage — regardless of whether it saved any real time — just to stay off whatever list that dashboard was quietly building.
Where the Trust Actually Broke
| What leadership started to wonder | What engineers started to wonder |
|---|---|
| Are engineers avoiding AI tools to protect their own job security? | Is this metric actually a layoff list in disguise? |
| Is the four-day fix real, or exaggerated to resist the mandate? | Does anyone up the chain understand what “architecture” even means? |
| Why isn’t velocity compounding the way the demo promised? | Why is my most important work — catching what the AI missed — invisible on every dashboard? |
Nobody in either column was acting in bad faith. Both were reacting rationally to incomplete information, filtered through a tool that had made one part of the job dramatically faster and left everything else exactly as slow as it always was.
Week 12: The Fix Wasn’t Technical
The company didn’t abandon AI-assisted coding. It changed what it measured and who had to sign off on what:
- Every AI-assisted PR touching shared infrastructure or cross-service logic now required a named human architect’s sign-off before merge — no exceptions
- That sign-off was tracked as its own separate metric, so it couldn’t be misread as engineers “slowing things down”
- The public AI-percentage dashboard was quietly retired
The Actual Lesson
The AI didn’t fail. It did exactly what today’s coding assistants are good at: generating correct code for a well-defined, local problem, faster than any human could. What failed was the assumption that architecture and system-level judgment — the work of making a hundred locally-correct pieces behave as one coherent system — compress at the same rate that code generation does.
They don’t. And a company that measures success only by the metric that did compress, while treating the metric that didn’t as evidence someone is dragging their feet, doesn’t just misjudge the technology. It manufactures distrust between exactly the two groups who most need to stay on the same side: the leaders setting expectations, and the engineers living inside the gap between “the AI wrote it” and “the system actually works.”
