Reading what nobody has time to read
Calls, notes and documents that hold the answer to a question you already pay people to guess at, read in full for the first time.
Consult · AI implementation
AI that has to work on a Tuesday, with nobody watching.
I decide whether AI will pay for itself in your business, and where it will, I build it and put it into production. I run these systems in my own company, so I meet the bill and the failure modes before you do.
What an AI engagement covers
A demo is a happy path with a person standing beside it. Production has worse inputs, nobody explaining, and a cost that lands on somebody when the answer is wrong. The gap between the two is where an AI budget goes.
AI strategy, the way I do it, starts with one commercial question: what does being wrong cost, and who finds out? Where a mistake is cheap and visible, build it and get on. Where it’s expensive and invisible, the model is the easy part and the checking is the work.
A fair share of what reaches me as an AI project turns out to be a search box, a report nobody wrote or a process with too many approvals in it. I’ll say so, because that answer costs you less than the build would have.
- AI strategy: whether it will pay, before anything is built
- Choosing the technique, not only the model
- Private AI where data can’t leave
- Answers checked against the record
- Spend measured per call
Typical work
What this looks like in practice
Free text into a record you can audit
What people type into chat, email and forms, turned into structured data with a trail behind every line.
Private AI for data that can’t leave
Speech-to-text and open-weight language models on hardware you control. I run my own this way because my business handles children’s data.
Answers checked against the record
Where a number will be acted on, the AI’s answer is reconciled against the system of record to a stated tolerance, and a miss is a failure, not a warning.
The right-sized model, measured
Candidates run over identical input and judged side by side, so the model in production is there because it won.
Your assistants, connected to your systems
MCP servers over live business data, with its permissions, so staff asking an assistant about a customer get the record and not a guess.
Where it starts
The Leverage Diagnostic.
Two weeks, fixed, £7,500.
All prices exclude VAT.
Two weeks to find the costly constraint, prove the risky assumption, and price the right response.
If you suspect AI could take cost out of the business and can’t yet say where, the diagnostic is how that gets answered: it finds what is costing you, and proves the risky part before anyone commits to a build. If you already have a brief I can build from, ask for a quote, which costs nothing.
You leave with one agreed proof, and a decision you can act on.
Questions
The things people ask first
Is it always the right answer?
No, and being willing to say so is most of the value. A great deal of what gets sold as AI is a process problem with a model bolted onto it. Where AI changes the economics of a workflow I’ll build it; where it does not, saying so is the cheaper advice.
Can my data stay out of a third-party model?
Yes. I run speech-to-text and open-weight language models on hardware I own, because my own business handles children’s data. The same pipeline can run on my hardware, on yours, or in a private cloud, and working out which is right for your data is part of the job.
Is a frontier model always the right one to use?
No, and defaulting to one is how AI budgets get spent without much to show. A narrow, well-specified task doesn’t need the largest general-purpose model available. It needs the smallest one that clears the bar, which is dramatically cheaper to run and fast enough to use at volume.
The way to settle it’s to measure instead of assuming: run the candidates over identical input and judge the output side by side. That’s how the local model in my own pipeline was chosen.
What does it cost to run at volume?
That depends far more on the architecture than on the price per token. Work sized for a smaller model on hardware you already own turns a per-call charge into a fixed cost, which is what makes high-volume processing viable at all.
Either way it should be measured. On the platforms I build, every model call is logged with its token usage against a budget, so AI spend is a line you can read and not a surprise at the end of the month.
From Insights
What I’ve written about this
Running AI locally when the data cannot leave
A 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 itWhy an AI pilot stalls before production
The 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 itWhen a database query beats a language model
A good deal of what gets specified as AI is a query nobody wrote. How to tell the difference before you pay for the expensive version.
Read itOther ways to work together
Not quite the right shape?
Build
Hands-on engineering
Bespoke software, SaaS products and internal systems built from the ground up and delivered to production.
- Product design & architecture
- Full-stack development
- Desktop, browser-extension and chat clients
- APIs & integrations
- Cloud & DevOps
- AI-enabled features where useful
Partner
Technical leadership
A technical co-founder or fractional CTO to own the technology function: what gets built, in what order, and on what architecture.
- Technical/product strategy
- Architecture
- Team leadership
- Roadmaps & execution
- Technical due diligence
Consult
Complex technical problems
Architecture reviews, AI strategy and technical advisory for difficult or high-value decisions.
- AI & automation strategy
- System architecture
- Feasibility & prototyping
- Technology due diligence
- Modernisation planning
Work with me
Losing time or money to something you can’t fix?
I reply within one working day to arrange a free 30-minute call.