Most AI programmes do not fail at the model. They fail at the data nobody profiled, the exceptions nobody logged and the handover nobody scoped. I work out which of your workflows will survive automation, and which will quietly cost more than they save.
Independent consultancy for enterprise and public sector organisations across the UK.
Four kinds of work. Each one ends with a decision you can defend in a business case, not a deck that restates the question.
Which workflows are worth automating, which are not, and what each one genuinely costs to run once the exceptions are counted.
System design and hands-on build for organisations that need the thing to ship, rather than to be specified and shelved.
An independent read on a platform, a supplier claim or a business case before the budget is committed to it.
Standing access for the decisions that do not fit a project: build against buy, sequencing, hiring, and the risks nobody has named yet.
I hold no reseller agreements and no licences to place. That means the recommendation is the same whether the answer is build it, buy it, or stop.
Almost every AI pilot works. Very few survive contact with real data, real exceptions and real users. The distance between a demonstration and a system your teams can depend on is where the budget actually goes, and closing it is the work I am usually called in to do.
Five steps. The first carries a fixed price, so you can find out whether I am useful before committing to anything larger.
Four levels, from a fixed-price first week through to an embedded programme. Only the first carries a published price; the rest are scoped against your own requirements.
01 / START
Recommended
02 / ADVISORY
03 / BUILD
04 / PARTNER
Twenty years of hands-on delivery. Past clients include the Department for Work and Pensions. Most of the rest sits under agreements that prevent me naming the organisation, which is usually a sign of the kind of work it was.
Claims about AI are cheap and mostly unfalsifiable. These are things I have built and measured, from my own systems.
1,000 runs
One pipeline run a thousand times, producing around ten thousand reports, to separate the failures that were real from the ones that were noise.
Seven layers
Every failure is traced to one of seven layers before anything is changed: schema, data, function, pattern, journey, template, configuration.
No one-off fixes
A standing ban on patching a single case. If it cannot be fixed at the layer it belongs to, it is not fixed - it is hidden.
26 services
A production architecture of roughly twenty-six services, running local-first rather than renting someone else's inference by the token.
55 reports
A catalogue of fifty-five distinct reports generated from a single schema, because the schema was designed before the reports were.
139 activities
A monetisation registry covering one hundred and thirty-nine separate activities, each one costed rather than estimated.
One workflow. One week. A fixed price, and a written answer you can put in front of a board - including the answer that says do not build it.