Skip to content
Lacambra Web Portal Services
← All writing
AI · 10 November 2026 · 3 min read

Let AI read the labels, not move the records

The slowest part of a migration is working out what the old system's words mean. A language model is good at that, and the wrong tool for the part that has to be exactly right.

Dozens of blank manila tie-on tags sorted into five neat rows on a dark desk, with a small unsorted heap beside them and one oxblood tag in the heap.

On the legislative platform I built, four state legislatures between them used 123 status labels for one sequence of events, and only 44 of them were distinct strings. Most were the same stage in different words. A handful looked identical across two states and meant different things, which is the dangerous kind. Sorting out which was which had to come before anything else could be built.

Every migration I have done has a version of the same problem. Old statuses, old field names, codes somebody invented years ago, and free-text notes holding half the real information. Before a single record moves, someone has to decide what each of those means in the new system.

This part, a language model does well

Give a model the old labels and the stages they have to land on, and it will propose a mapping for every one, with a sentence on why, in minutes rather than days. Ask it to, and it will also list the labels it could not place. If I were mapping those 123 labels today, that is where I would start.

But what it produces is a draft, and a person approves every row. A wrong mapping is quiet. The record moves, it looks fine, and it sits in the wrong place for months. The labels that look identical and are not are exactly the ones a model will map the same way with complete confidence, which is why a person who knows the organisation has to read each line before it counts.

Where a language model belongs in a migrationIt mayread every old label and field namepropose a mapping, with a reason for eachlist the labels it could not placesummarise notes nobody has readpoint out records that look like duplicatesIt must notmove a recordapprove its own mappinghave the last word on any labeldecide what gets left behindThe model drafts the mapA person approves every rowOrdinary code moves itThe move itself stays ordinary code: same input, same output, every time.Where a model belongs in a migrationIt mayread every old label and field namepropose a mapping, with a reason for eachlist the labels it could not placesummarise notes nobody has readpoint out records that look like duplicatesIt must notmove a recordapprove its own mappinghave the last word on any labeldecide what gets left behindThe model drafts the mapA person approves every rowOrdinary code moves itSame input, same output, every time.
Drawn, not photographed. What a language model may do in a migration, what it must not, and where the person sits between them.

This part, it must not touch

Moving the records has to be exactly right, and exactly repeatable. I rehearse every migration on a copy of the live data before the real move, and a rehearsal is only worth something if the real move does the same thing, so that the counts reconcile down to the last record. A language model does not promise the same answer twice. Hand it ten thousand rows to transform and you cannot say which of them it quietly changed.

So the move itself stays ordinary code: same input, same output, every time. Once a person has approved it, the mapping becomes a plain table that the code reads. The model is never in the loop when a record moves.

Where else it earns its place

The same split works across the rest of a migration. A model can summarise years of free-text notes nobody has read, so the client can decide whether they are worth keeping. It can point out records that look like duplicates, for a person to confirm. It can draft the explanation for each group of records that did not move. What it cannot do is decide what gets left behind. That decision belongs to someone at the client who can still be asked about it in a year.

Which is why the report at the end does not change. It still lists every record that moved and every record that did not, with a reason for each. Where AI drafted the mapping, the report says so, and names the person who approved it.

Enrique Lacambra III builds internal business systems from Cainta, Rizal.

Email me← All writing

Tell me what is broken and I will tell you if it is worth building. A 20-minute call, no charge, and I will say so if you do not need me.