We cut AI down to four concrete jobs, then decided to switch it off
A law firm asked us for "AI for team productivity". We turned it into four concrete jobs, each with a person approving the result. A month later their case-management software launched its own AI and we closed the project. We think we got it right both times.
First, turn the brief into jobs
"AI for productivity" can mean a chat, a search box or an assistant that drafts, and with a brief like that there's no way to tell when it's finished or whether it works. So we asked what the team did by hand every week, and kept four jobs:
- Backfill: fill in the missing details on old cases. The result stayed a draft until someone approved it.
- Field fill: pull the details out of each new enquiry and propose them for the case fields, again with approval before anything was saved.
- Summaries: one page before each consultation, to replace ten to twenty minutes of reading with one. That was the design goal; we never got to measure it.
- Data control: spot returning clients, duplicates and possible conflicts, and log everything.
What we left out
We dropped the chat, the "ask your data anything", a screen for training the model, fine-tuning our own models and a fifth pillar for bilingual drafting. The firm's software already covered the core of all that. We estimate the cut saved about sixty hours of building.
What happened to each job
Field fill was the only one the team ever used. It passed its tests: eight of eight case fields and four of four contact fields. Months later we found the approve button had been failing for a while without anyone noticing, because the process behind it had been switched off. We removed the button and the suggestions kept working.
Backfill taught us something about cost. On the firm's own computer it would have taken about nine hours for some 3,400 cases; on a rented server, over anonymised text, we estimated around a dollar a run. We tested it on 47 cases but never on the full archive, because the privacy contract with that provider had to be signed first.
Summaries and data control stayed on paper. Two pieces of the second did ship: an option to exclude a client from any use of AI, and a switch to turn each feature on or off.
Along the way we changed AI provider three times in a month, each time over privacy: one wouldn't guarantee it kept no data, another wouldn't give us the capacity we asked for. With a law firm's client data, that outweighs any technical comparison.
Why we switched it off
In early June the firm's case-management software launched its own AI. It did much of what we had planned, inside the tool the team already used, under contracts already signed. Carrying on would have meant maintaining two systems for the same job, so we closed the project and helped the firm move to the software's own option.
We have no usage or time-saved figures.
What we'd apply to any business
- If you can't name the job and who checks the result, there's no AI project yet.
- Before building, look at what the software you already pay for has planned. In three months it may do it for you.
- With sensitive data, decide first who may process it, then compare.
- Keep checking what's already running. A feature nobody looks at can stay broken for months.
- If your software already does what yours does, close yours and save the upkeep.
Being sold AI and not sure you need it? Write to us and we'll break it down into concrete jobs, even if the answer is no.