UK SMEs are being told to “adopt AI” as if it were a software licence. It is not. It is a change to how a person in Leeds, Manchester or Birmingham spends Tuesday morning. If you start with a chatbot on the website, you will get a demo. If you start with a painful workflow, you might get hours back.
Pick a job, not a model
List the work that is repetitive, rules-ish, and expensive when it is late: invoice coding, inbound triage, quoting first drafts, checking a delivery exception. Score each by volume, risk, and whether a human must still sign it off. The winner is usually boring. That is a good sign.
Do not begin with “customer-facing generative AI” unless you have a retrieval set you trust and a tone of voice you are willing to defend to the ICO and to a client. Internal tools forgive more. Public tools forgive less.
UK GDPR is a design constraint
Know where data sits, which processor you are using, and whether anyone is using your documents to train a foundation model. Many United Kingdom SMEs can get value from retrieval and classification without ever fine-tuning anything. Write that down. Your lawyer will eventually ask.
Prove it in production-like conditions
A slide with a 90% accuracy number from a tidy spreadsheet is not a system. Run the workflow on last month’s messy reality. Measure time saved and error rate, including the time staff spend correcting the machine. If correction eats the gain, you do not have a product yet.
What “good” looks like in 90 days
One workflow in a supervised slice. Logging. A named owner in the business. A kill-criteria. Training that is not a PDF. If a vendor cannot talk about those, they are selling theatre. If you want a partner who starts from a risk register, that is how we run AI software development from Leeds.
The companies that win with AI in the UK will not be the ones with the flashiest demo. They will be the ones whose operations quietly got cheaper while remaining accountable.