Consult · AI implementation
I work out whether AI will pay in your business. Where it will, I build it and run it in production. I run these systems in my own company, so I know what they cost and how they break.
What an AI engagement covers
AI demos well. The projects I get called into all worked once, in a room, on inputs somebody picked; then they went live and met the other half of the business, the half-filled form, the customer who put the delivery address in the notes field, the Sunday night with nobody there to catch a wrong answer. Closing that distance is the engineering and most of the bill.
So AI strategy, the way I do it, starts with what a wrong answer would cost and who’d notice; the automation article has the test in full. A lot of what reaches me as an AI project is a search box, a report somebody rebuilds in a spreadsheet every Monday, or an approval step nobody trusts. Fixing that is cheaper than a build and you can do it this week.
Where a model is right, what it costs to keep running is a design decision taken early. Put a right-sized model on hardware you own and a per-call charge turns into a fixed monthly cost.
Typical work
Call recordings, notes and documents, read in full for the first time. Nobody has the spare month it would take by hand, so the reasons people said no and the question every buyer asks first stay in the archive and out of your reporting. I pick one question the business would act on and answer it from the lot.
What people type into chat, email and forms becomes structured data, and every line keeps a trail back to the words it came from.
Speech-to-text and open-weight language models run on hardware you control, on your premises or in a private cloud, so the data never leaves your estate.
An answer is reconciled against the ledger before anyone sees it, and anything that fails to tie out is held for a person to look at.
The candidates run over the same input, and the one in production is the one that did best on your data.
MCP servers over live business data, so an assistant asked about a customer answers from the record, under the permissions that person already has.
Where it starts
All prices exclude VAT.
Already have a brief I can build from? Ask for a quote. If you think AI could take cost out of the business and can’t yet say where, the fortnight finds what’s costing you and proves the risky part before anyone commits to a build.
Questions
It isn’t, and part of what you’re paying for is somebody willing to say so. If what I find is a process that needs fixing first, you’ll hear that before you hear a model name. Where AI changes the economics of a workflow, I’ll build it.
Yes. The question that follows is whose hardware: mine, yours, or a private cloud your people run. Where your policy draws the line decides it, and working that out is part of the job.
I run my own platform this way; the coding education case study has the detail.
Defaulting to the biggest is how AI budgets get spent with little to show. A narrow, well-specified task runs on something much smaller, far cheaper, and fast enough at volume.
The size of the model and where it runs decide the bill at volume, more than the price per token does. On the platforms I build the running cost is known per call.
From Insights
Anyone using Claude can ask my business whether I'm worth a call, through a small server I run. The fit check turned away my ideal client before I stopped letting wording decide.
Read itA refused cloud AI request is a constraint on one stage of the pipeline, not the end of the project. How I run private AI in my own business, and how to work out what yours needs.
Read itThe pilot worked and nothing shipped. Five reasons that happens, all of them visible before the pilot started, and what to agree on day one so yours is not the next.
Read itOther ways to work together
Build
Partner
Consult
Work with me
I reply within one working day to arrange a free 30-minute call.