What AI actually changes about migration cost
The honest version: where the 3–5× is real, and where a human still signs.
Every consultancy now claims AI-accelerated delivery, so the claim itself carries no information. Here's the concrete version for data migration.
Where the leverage is real
Profiling. Schema discovery, distributions, null analysis, candidate keys, referential integrity findings — machine-generated in hours, consistently formatted, exhaustive rather than sampled.
Mapping drafts. A first-pass source-to-target mapping specification, field by field, with transformation logic proposed and edge cases flagged. This used to be weeks of spreadsheet archaeology. It's now a review exercise.
Transform code and tests. Generated from the approved mapping spec — with the tests generated alongside the code, not promised afterwards. Documentation falls out as a by-product rather than a chore.
Reconciliation. The economics of exhaustive, row-level reconciliation used to be prohibitive, which is why the industry normalised sampling. Generated reconciliation harnesses make "check everything" the cheap option. This is the quiet revolution: AI leverage shows up as more assurance, not just less cost.
Where it isn't
Nothing generated goes to production unreviewed. Mapping decisions on regulated data are judgement calls — an engineer approves every spec, and the approval is recorded. Cutover decisions, exception dispositions, sign-offs: humans, every time, on the record. The machine does the grunt work; the accountability stays human.
The result isn't a cheaper day rate. It's a different shape of firm: a migration that used to need a team of eight for nine months becomes a fixed six-week engagement. That's not a discount. It's a different product.
Got a cutover date? Tell us the source system, the target, and the deadline — we'll tell you within 48 hours whether we can hit it and what the Assessment will cost.
freddie@godwit.uk