AI automation, with a person checking
We're an AI automation agency that builds AI into the systems a business already runs: reading documents, sorting email, pulling data out of forms and drafting replies, with a person checking anything that matters. We build it as software we host and look after, with tests that prove it still gets the answers right.
We use AI where it's good and keep it away from where it isn't. If a rule or a lookup does the job more reliably, we use that instead.
Good first jobs for AI
- Documents that arrive in many layouts but always carry the same facts.
- An inbox where most messages fall into a handful of types.
- Free-text fields that need to become proper data.
- Replies and summaries that follow a pattern but need a human touch.
Where AI helps
Language models are good at reading messy text and turning it into structured data: an emailed order into order lines, a PDF invoice into a supplier, amounts and dates, a CV into skills and experience. They're good at sorting and summarising, and at writing a first draft for a person to check.
They are not reliable judges of anything with legal, financial or safety consequences, so those decisions stay with people. We design every AI step so that it's clear what it decided, what it was unsure of, and who checks it.
What we build
Document reading
Invoices, delivery notes, purchase orders, application forms and contracts read and turned into data your systems can use, with uncertain fields flagged for a person.
Email triage
Incoming email sorted by what it is and who should deal with it, with the key details pulled out and a draft reply ready to check.
Data extraction and clean-up
Free text, old records and imported lists turned into consistent, structured data you can search and report on.
Drafting
First drafts of quotes, summaries, reports and replies, written from your data and your templates, for a person to approve.
Classification
Enquiries, tickets, transactions or products put into the right category, with the reason recorded so it can be checked.
How we keep it accurate
Every AI step we build gets a set of test cases from your real work, checked by your team: these documents should produce these answers. The tests run before any change to a prompt, a model or a provider goes live, so an update that makes the answers worse doesn't reach you.
Where it matters we use more than one model provider, so an outage at one doesn't stop the work, and we keep a record of what the AI was given and what it returned.
Your data
We tell you which AI providers your data goes to and why, and use terms under which it isn't used to train their models. Personal data is kept to the minimum the job needs, stored for a set period, and covered in your data protection paperwork.
An example: a supplier invoice arrives
A PDF invoice lands in the accounts inbox. The AI reads it and pulls out the supplier, invoice number, date, lines, VAT and total, then checks them against the purchase order and the goods received. If everything matches, it's queued for approval with the evidence attached. If a price differs or a field is unclear, it goes to a person with the difference highlighted.
Approved invoices go into the accounting system without being retyped, and every decision, made by the AI or a person, is recorded.
How a project runs
We start with a sample of your real documents or emails and show you what the AI gets right and wrong before anything is built into your systems. Then we agree the scope and a fixed price for each stage.
Once it's live we host it, watch the error and fallback rates, rerun the tests when anything changes, and improve it as your team finds the edge cases. That's covered by a monthly fee.
Software we built and still run
Our own products, live today. We designed, built and host every one of them.
-
4dayweek.io
4dayweek.io's AI service reads job ads all day and pulls out the salary, location, remote or hybrid status, schedule, skills and seniority. Two model providers cover for each other, implausible salaries are dropped, and golden tests against the real models must pass before any change goes live.
Industries we build it for
Questions
Which AI models do you use?
Whichever does the job best for the cost, usually models from Anthropic (Claude) and Google (Gemini). We don't tie you to one: the tests tell us when a different model is better or cheaper.
Can AI make mistakes?
Yes, which is why we design for it. Anything uncertain or important goes to a person, every result can be traced back to its source, and the tests catch a change that makes answers worse.
Do we need lots of data to start?
No. This work uses existing models rather than training new ones. A few dozen real examples of the documents or emails you handle are enough to build the first tests and show you results.
How is this different from using ChatGPT?
ChatGPT is a tool a person uses one conversation at a time. We build AI into the process itself: it runs on every document or email as it arrives, writes the results into your systems, and is tested and monitored like any other software.
Tell us what you want to build
A few lines is enough to start. Phil replies within one working day.
Start a project