What is actually worth automating with AI?

Three things have to be true before automation is worth the money, and they have nothing to do with technology.

The task has to be written down, or at least writable down. If the rules exist only in the head of the person who has done it for nine years, the first job is documenting the process, and that usually delivers value on its own.

The inputs have to be reasonably consistent. Emails, PDFs and form submissions are fine. Handwritten notes photographed at an angle in bad light are a much harder problem than the demos suggest.

And somebody has to own the output. Automation without a named person checking a sample of results is not a productivity gain, it's an unmonitored process, and those fail quietly.

Sending business data to a model provider is a data-processing decision, and if any of it is personal data, UK GDPR applies in exactly the way it does to any other processor. Write down four things before you build anything: what data the workflow touches, who processes it and where, how long anything is retained, and what your lawful basis is. That document is useful whether or not you proceed.

Then four practical defaults. Send the least data that makes the task work. Strip identifiers the task doesn't need, because most classification and drafting jobs don't need names at all. Use providers whose commercial terms exclude your data from model training, which the major API products do and the free consumer tiers often do not. And log what went in and what came back, so you can answer questions about a decision months later.

One decision deserves particular care. Where an automated output has a legal or similarly significant effect on a person, UK GDPR gives them rights including human review, so those workflows need a person in the loop by design, not as a policy nobody follows. The ICO's guidance on this is linked in the sources below, and it is more readable than its length suggests.

Start with off-the-shelf. For meeting notes, transcription and general writing assistance, mature products exist for a few pounds per user per month and no build will beat them. Try those before paying anyone to build anything.

A custom pilot on a single workflow starts at £1,500 with us and runs four to six weeks: the scoping session, the build, running it alongside your existing process, and the measurement at the end. Automation that belongs inside a larger system is quoted as custom software, from £5,000.

The ongoing cost is model usage, billed by the provider on your own account so you can watch it. At small-business volumes, classifying a few hundred emails a day or summarising a few dozen documents, that is typically a few pounds to a few tens of pounds a month. It is worth setting a spend cap on the account on day one, because the failure mode of an automation loop going wrong is a bill and not an outage.

You do not need a consultant for the first pass. This is the plan we'd follow, and you can run it with whoever in your business is most curious.

1. Days 1 to 5: inventory. List every repetitive task across the week and score each on frequency, time, judgement and blast radius. Pick exactly one. One.
2. Days 6 to 8: set the number. Decide before building what success means: hours saved a week, turnaround time, or error rate. Measure the current value now, by hand, so there is something to compare against.
3. Days 9 to 20: build the smallest version. It only needs to handle the common case. Route anything unusual to a person instead of trying to cover every branch.
4. Days 21 to 28: run it in parallel. Alongside the manual process, not instead of it. Have someone check every output for the first week and a sample after that, and write down every case it got wrong.
5. Days 29 to 30: decide honestly. Did it hit the number? If yes, expand it. If it was close, fix the specific failures. If it missed badly, stop, and you have spent a month instead of a budget.

The discipline that makes this work is picking one workflow. Every failed AI project we've been asked to rescue started by automating five things at once and being unable to tell which of them was working.

PolicyMind is our compliance platform, and it's a fair illustration because you can see both halves of the argument in one product.

AI does the drafting: turning a structured brief into a first-pass policy document, summarising long documents, and answering questions grounded in an organisation's own policy library rather than in whatever a general model believes about the subject. That work is high frequency, mostly mechanical, and always checked by somebody before it counts.

Approval, publication and staff acknowledgement are deliberately manual, because a policy nobody signed off is worse than no policy at all, and the whole point of the audit trail is that a human decision is recorded. Automate the drafting, keep the deciding. That line is the single most useful thing we've learned from running an AI product instead of writing about one.