Start an AI work register before you automate it

Most teams asking how to record AI contribution are really asking which tool to buy, and the honest answer is that they are not ready to choose one. They do not yet know what they would record, which parts of it anyone would look at, or what a month of entries would even show them.

So start with a spreadsheet. Not as a stopgap you apologise for, but because a register you fill in by hand for a month teaches you what your record needs to contain, and no evaluation call will. The cost of being wrong is a deleted column rather than a migration.

Why by hand, first

Automated capture answers the question of how work happened. It cannot answer whether the resulting record is one your team would defend to a client, because that depends on judgement calls nobody has made yet. Which involvement counts as material. What review means for your work. Which contributions belong to a workstream and which are noise.

Doing it by hand forces those decisions into the open in week one, when they are cheap. It also surfaces the disagreements. Two people recording the same kind of work differently is not a data quality problem, it is a definition your firm has not made yet, and you want to find that with eleven columns rather than after an integration is live.

The other reason is smaller and more practical. A manual register makes the volume visible. Teams are routinely surprised by how much or how little AI touches their delivery once they have to write it down, and that number changes what they build next.

The eleven columns

These map cleanly onto what an automated system would capture later, which is the point. A register you outgrow should import, not be retyped.

Involvement and contribution are doing different jobs and both are needed. Involvement says how much of the thinking was machine assisted. Contribution says which activity it was. Collaborative research and collaborative drafting carry very different weight in a client conversation, and a single blended field loses that. If you want the same vocabulary in a form a client can read, the attribution statement builder writes one from the same fields.

The two columns everyone skips

Review status and outcome are the ones that get dropped in week two, because they cannot be filled in at the same moment as the rest. The work is recorded when it happens and reviewed later, which means going back to a row, and going back to a row is the habit that fails.

They are also the only two columns that turn the register from an activity log into something that answers a question. Without them you have a record of how much AI you used, which is the least interesting fact available. With them you can see review coverage and discard rate, and those two numbers are what tell you whether the tooling is pointed at the right problem.

The fix is a weekly pass rather than better discipline. One person, fifteen minutes, closing out the rows from last week. Anything still unreviewed after two weeks is a question worth asking out loud.

Keeping it a record of work, not of people

A register with a person's name in every row is one bad quarter away from becoming a performance measure, and the moment people suspect that, the entries get careful. Careful entries are worse than no entries, because they look like data.

Two rules keep it honest. The owner column names who is accountable for the deliverable, not who typed, and the register is never read by row. Every question you ask of it should be about a workstream, a project, or a month, and if a question can only be answered by filtering to one person, it is a question about a person.

• No prompt or response text: keep a pointer in the evidence column, never a transcript.
• No time-per-tool tracking: this measures the person and tells you nothing about the deliverable.
• No individual leaderboards, in either direction: high AI use is not a virtue or a failing.

What a month of entries tells you

Three things become visible, and none of them are visible from a monthly tooling bill. The first is where AI actually lands in your delivery, which is usually more concentrated than people assume, sitting in two or three workstream types rather than spread evenly.

The second is review coverage: the share of client-facing rows with a completed review. Teams tend to discover this is lower than their policy implies, not through negligence but because review was never anyone's scheduled work.

The third is the discard rate. Work that was produced, recorded, and then thrown away still cost time and money, and a workstream with a high discard rate is telling you the tool is being used for something it is not good at.

When to stop doing it by hand

The register has done its job when filling it in becomes the bottleneck rather than the learning. In practice that is when the definitions have stopped changing, the weekly pass is routine, and someone has asked a question the spreadsheet cannot answer without an afternoon of work.

At that point you know what to automate, which is a much stronger position than the one you started in. You can evaluate tools against a record you have already been keeping, and you can tell whether an automated capture matches what your team knows to be true, because you have a month of manual entries to check it against.

Keep the columns stable when you make the move. The register is the thing that makes the first automated month legible, and a schema change at the same moment as a tooling change means neither can be trusted.