Your AI Rollout Is a Management Problem in a Technology Costume
The delegation frame and the reading tax, argued at essay length with the full story attached.
Claude Code is a coding agent you run from a terminal. You describe an outcome; it reads the codebase, plans, writes, runs and tests — then hands you a change to review. It is not autocomplete, and it is not a chatbot with a code block. The right mental model is a fast, tireless mid-level engineer with no memory of yesterday and no fear of being wrong. Everything else in this piece is about managing exactly that person.
90 words · rewritten twice since Feb — a reader talked me out of "pair programmer" and they were right
The pitch vs the parking lot
What they say · what showed upThe pitch
"Ship features while you sleep."
The parking lot
It ships drafts while you sleep. Someone senior still reads them in the morning — that's the asterisk on the verdict stamp, and it's a real line item.
The pitch
"It understands your whole codebase."
The parking lot
It reads your whole codebase, which is different. Give it an ambiguous brief and it will pick an interpretation and commit with terrifying confidence. Good briefs in, good code out.
The pitch
"Your team will move 10x faster."
The parking lot
The boring work got 3x faster — scaffolding, migrations, tests. The interesting work got maybe 1.2x faster, because the interesting work was never typing. Still a great trade.
The pitch
"Quality goes up because AI doesn't get tired."
The parking lot
Quality held — our defect rate stayed flat, which honestly surprised me. But it held because we rebuilt the review step, not because the model is careful. It isn't.
The receipts
Our team, five months, real tickets▼ LowerPR review cycle
Median review turnaround fell markedly. Real, but read the fine print: the queue moved, the reading didn't. See the third receipt.
▼ Much lowerGrunt-work time
Scaffolding, migrations and test boilerplate dropped by more than half a sprint's worth. The cleanest win in the whole file — nobody on the team misses that work, and nobody pretends to.
▲ HigherSenior review hours
The asterisk, made visible. Agent throughput converts junior writing time into senior reading time. Budget for it or discover it the hard way around sprint three, like we did.
— FlatDefect escape rate
Barely moved — statistically a shrug. The fear that agents would flood production with subtle bugs did not show up, once spec-level review replaced line-by-line reading.
Honesty corner: one team, one org, no control group — this is a field report, not a study. The deltas are directional, not audited statistics; the commentary is the point. Precise numbers land when the sample is honest enough to publish.
Where it bites
Learned so you don't have toConfidently wrong
Ambiguous brief in, polished disaster out. The output looks finished, which is exactly the trap — polish reads as correctness.
The fix — checkpoint at the spec, not the code. Two minutes of "here's my plan, okay?" saves a day of unwinding.
The reading tax
It doesn't remove work; it relocates it. Writing time falls, reading time rises — and it rises for your most expensive people.
The fix — re-plan senior capacity before rollout, not after the mysterious slowdown two sprints in.
The 2 a.m. idle
Its rhythm assumes someone can answer a checkpoint in minutes. Across Bengaluru↔US hours, an agent blocked at 2 a.m. IST naps for half a day.
The fix — batch briefs at handoff time and write them so complete the agent never needs to ask. Annoying. Works.
Get it / skip it
The honest sorting hatGet it if
- Your backlog has a thick layer of scaffolding, migrations and test debt nobody loves
- You have seniors who can review at the level of intent, and you'll actually give them the hours
- Your team already writes decent tickets — good briefs in, good code out
- You're willing to treat this as a management change, not a tool install
Skip it (for now) if
- Your codebase has no tests — you'll be reviewing blind and the polish will fool you
- Nobody senior has bandwidth to read more; you'll just move the bottleneck and blame the tool
- You measure success in story points and can't change that yet
- You want it because the keynote was exciting. Wait one quarter; it'll still be here