The problem
Six ideas, and then the inbox
The company ran an AI workshop and came away with six ideas. Nothing was built. Nobody refused; daily work simply felt more important, as it does at 25 employees, and a list of maybes loses to a full inbox every time.
A year later one problem stood out. Requests for quotation arrive as specification packs, anywhere from three pages to three hundred, padded with attachments that have nothing to do with the job. Two or three people read each one, pull out what matters and annotate the PDF. The managing director checks the result before the bid moves on. The work needs context that exists nowhere on paper.
The belief
“Too complicated for AI”
A first try in a plain ChatGPT chat returned a summary. A quotation needs values in the right fields, not a summary. That was the wrong tool for the job, but the company drew a bigger conclusion: that the documents were too complicated for AI, and that AI had no real use in the business.
That conclusion was the real obstacle. It was not a skills gap, and another training session would not have moved it. A belief formed by one bad experience changes the same way it formed: through experience, this time on the company’s own documents.
Co-building
Build with them, not at them
The second attempt had two parts: a prototype, and one hour a week. Staff bring whatever is annoying them that week. The problem is fixed within the hour, or becomes a plan to build the fix together. The people who read the specification packs shape the tool that reads them, which is also why they trust it.
The early lessons were less about technology than about habit. A prompt is not a spell. Asking for a five-minute film gets nothing useful; asking for the script, then the characters, then the music, gets somewhere. A three-hundred-page specification obeys the same logic: break it into steps, and decide what is reasonable to ask of each one.
Change management
A rhythm, not a project
Small companies do not stall for lack of ideas. They stall because every new initiative competes with the day job and loses. So the work runs in sprints of eight weeks on and four weeks off, fitted around the order book rather than on top of it.
- Workshop Six ideas A list, and no time to build any of it.
- A year later One problem Pulling the relevant values out of specification packs.
- Now A prototype on real documents An interface and a database the company hosts itself, so the value stays in the company and the data can serve other purposes.
- Every sprint Eight weeks on, four off Each sprint closes with what was done, what is fixable next, what goes on the back burner and what is new.
The sprint review is short on purpose. Four questions, the same every time, so the habit outlasts the enthusiasm.
Trust and data
The tool shows its work
The usual objections to AI in a small company are about control: where the data goes, and whether anyone can tell when the answer is wrong. Both were design decisions from the start. The database is hosted by the company, not by a vendor.
Design rule
Every value points to its source
Every extracted value links back to the page it came from, so a person can check it in seconds. Where a document is unclear, the interface flags it for a person instead of quietly picking an answer. Judgment does not disappear; it moves to where it can be seen.
Measuring it
Real bids, checked blind
The company already had a form with fifty fields, so the tool knew what to look for. A small trial set of real documents has been tested so far. Each finished bid went into a fresh chat that knew nothing about the build, and the extraction was rated against it field by field.
Success is defined in the company’s terms, not the tool’s: less reading time per bid, fewer values missed, and a managing director who checks a list instead of a three-hundred-page PDF.
What happens next
A test, not a verdict
The pilot is meant to show whether the idea can work in a company of this size, suggest what to build next and find other uses for the data it gathers. There is no headline number yet. That is what in progress means.
The workshop left six ideas on a list. This attempt leaves one tool, a set of checked results and an hour in the diary every week.