AI, put to work in real businesses.

1 August 2026

Research agents that aren't allowed to lie

Today's job: build a list of thirty established UK companies — owner-managed, £8M–£120M turnover, no in-house tech leadership — across construction, manufacturing, wholesale and logistics. The old way is a week of trade directories, or buying a stale list from a data broker. I ran three AI research agents in parallel instead, each under the same hard rule: no source, no row. Every company needed a Companies House number confirmed Active, a turnover figure with a named citation, and an owner verified from the officers register.

The interesting part is what got rejected. One "family-run" firm turned out to have a private-equity holding company behind the branding — visible in two clicks on Companies House, invisible in the marketing. Another quietly employed a Head of IT, which disqualified it. And one promising haulier was, on inspection of its accounts, mostly a poultry business that happens to own lorries.

• 35 candidates researched, 30 survived verification. The cuts were the point.
• I hand-checked a sample of nine against Companies House afterwards: nine matched.
• Stale figures didn't get quietly reused — anything older than two years is flagged for a refresh before anyone acts on it.

The lesson transfers to every business I talk to: AI research is only as good as the refusal rules you give it. Ask a model to "find companies" and it will pad the list to please you. Tell it that every claim needs a source or the row dies, and it comes back with fewer answers — the true ones. That discipline is configurable. Most people just don't configure it.

3 agents in parallel · 35 researched → 30 verified · built in an afternoon

1 August 2026

A readiness assessment, built and shipped in a day

First artefact of the new practice: a 12-question AI readiness assessment for business owners, scored across leadership, operations, data and people. Built with AI-assisted coding in a working day; a decade ago the same thing would have been a two-week job and an invoice most owners wouldn't sign.

• One self-contained page — no framework, no tracking, no cookie banner. Your answers never leave your browser.
• Scores against what UK businesses actually report, not a vendor's fantasy benchmark.
• Shipped the same day, live at roryphillips.dev/assessment.

The point isn't the tool. The point is that software this cheap changes what's worth automating: a job that used to need a development budget now needs an afternoon. Most of the businesses I talk to haven't priced that shift in yet.

~1,100 lines · built in a day · £0 to run

1 August 2026

Why this page exists

The official numbers are stark. Roughly half of UK businesses now say they "use AI" — but only about one in ten automate anything real. In construction, 88% have no AI plans at all. The companies with the most repetitive manual work, and so the most to gain, are precisely the ones with nobody whose job it is to look.

That's the job I've taken on: fractional AI officer for owner-managed UK companies — strategy with the owner, systems built alongside the team, and training so the capability stays in the building when I leave.

This page is the working log. I'd rather show the receipts than write the brochure: each week, one note on something built or learned, with the numbers attached — including the ideas that didn't survive contact with a real business.