The brief
Eli-May Martin has spent 25 years in document control on EPCM mega projects. She had just been offered a consulting role, had decided to start her own business around it, and wanted to walk in on the front foot rather than the back.
Her work is authority checking. Not whether a document is any good, but whether it is the authorised version, approved by the people who were supposed to approve it, matching what every other system says about it. On a mega project that means hundreds of thousands of documents, checked and rechecked as the project moves. She was doing it by hand, in spreadsheets.
She had already tried AI. Copilot, and a little ChatGPT. Neither got her anywhere near the result she needed on this particular problem, and she wanted to know whether the ceiling was the technology or the way she was using it.
What she was nervous about
Data. Everything else was secondary.
Her clients are government and major project consortiums. Some of them contractually forbid project data leaving the country, let alone sitting on a server owned by someone else. If the answer to that question was unsatisfactory, nothing else about the day mattered.
So we spent real time on it before we built anything: the different levels of security available, what actually leaves a machine and what does not, and the legal position on where data sits and who is responsible for it. That was not a detour from the session. For her industry it is the qualifying question, and she needed to be able to answer it herself, in front of a client, without us in the room.
How the day ran
A single-day one-on-one intensive, just the two of us.
First half, set up.
- Installed the tooling on her own machine and walked through what each part of it actually does
- Built her context layer, which is roughly an hour of teaching the AI her company, her industry, her clients and how she works
- Organised her folders so the work had somewhere to live
We deliberately connected nothing. No CRM, no third-party tools, no integrations. Partly because none of it was her quick win, and partly because every connection is another data question she would have to answer for a client later.
Second half, build.
We started on prompting: not tips, but what a prompt is actually doing and why the results change. Then she wrote one and turned it into her first working program.
That program was the thing she came in for. It reads the documents, checks them for authority across her sources, and organises the result the way she needs it, at a scale she cannot reach by hand.
The only part that fought us was cosmetic. Some text overlapped on the dashboard and needed tweaking. That was the hardest thing that happened all day.
The moment it clicked
It was not during the explaining. It was while she was building it, when the thing in front of her started doing the work and she could see how far it went.
Her words afterwards: “I knew but I didn’t know.”
The result
By the end of the day she had reviewed hundreds of documents with a tool she had built that morning.
She costed the same job by hand at three weeks of full-time work. It ran in about an hour.
That is her figure, from her own experience of doing this work for 25 years, not a modelled estimate from us.
She also left with the thing that is harder to put a number on. She is about to run document control on a project overseas, as a brand new consultancy, and she can now walk in organised, across her documents, and able to answer the data question herself.
What we are honest about
She finished the day with a working tool and the ability to keep building. She has since flown out to run the project, and the eight weeks of follow-up consultation included in the intensive are still running. Commercial outcomes from her business are hers to report, and when there are any, we will not be claiming them here unless she tells us herself.
The point of the day was never a deliverable we hand over. It was that she can now build these herself.