Greycell Technical · Case Study 01 · Operator evidence
Same Login. Same Tool. Two Completely Different Outcomes.
Thirteen weeks inside a live UK engineering firm. Every number below is read straight off the account records. Nothing here is modelled or estimated.
- Your team already has an AI subscription and the usage graph is quietly falling.
- You're heading into a bid and need a credible, specific AI-capability claim, not a marketing line.
- You want to know what "AI actually working inside an engineering firm" looks like before you spend anything on it.
Here's what it looks like when an AI subscription is actually wired into the way an engineering firm works. The inbox triaged every morning before anyone asks. A raw enquiry turned into a costed, client-ready proposal, in the firm's own voice, four times in the past month, each one checked before it went out. Every technical report scored against the standard and against the source numbers before a client sees it. Two hundred and eighty-seven finished, reusable pieces of work (proposals, reviews, transmittals, tools) filed and ready to use again, not read once in a chat window and lost.
That's not a forecast. It's thirteen weeks, measured, from inside a live UK engineering firm that Greycell serves under a services contract. And here's the part worth reading on for: it happened on the same £75-a-month subscription the rest of the team was already using, and getting less out of, every month that passed.
Same login. Same tool. Two completely different outcomes.
What 13 weeks actually showed
Over 8 April to 6 July 2026, we pulled the real usage data for this firm's AI subscription, together, to see what was actually working and what wasn't: every conversation in the team's shared chat window, and every session Greycell's own system logged on the same account. No figure below is estimated. It's read straight off the record.
- 71 conversations across 13 weeks.
- 65% were three questions or fewer. 27% were a single question and done.
- Monthly use falling: 27 conversations in April, 21 in May, 17 in June.
- 91 working sessions in the last 4 weeks alone.
- 287 finished pieces of work shipped since April, all filed and reusable.
- Output compounding, not flat: 3 pieces of work in April, 91 in May, 182 in June.
- 36 tools, each purpose-built for a specific job, plus 58 files holding the firm's own standards and the state of every job, so nothing has to be re-explained each time.
The subscription itself costs £75 a month, in total, for everyone on it. That's roughly half an engineering hour a month.
Two people, one login each, wildly different results
Three things this proved
Structure decays if nobody keeps pushing it.
Two shared workspaces were set up in the team's first week, in April. They were used for six weeks, nineteen conversations filed away. Then, from 21 May, nothing. Not because anyone objected to them. A habit that isn't wired into the way the day already runs doesn't survive on its own, however good the intent behind it was.
The team isn't idle. It's stranded.
Real, hard engineering work is going through that same chat window, by hand, one question at a time: a long, iterative session working step by step through statistical analysis of on-site test data. A 29-question session explaining a technical drawing. Sessions running to 15 and 18 questions on technical design points. That's system-grade work being pushed through chatbot-grade tooling, because chatbot-grade tooling is all anyone gave them. The appetite for the work is there. The tooling to carry it never arrived.
Hand people a real tool and they use it more than anything else.
Across all 71 team conversations, the single most common pattern, about a third of them, was one document-review prompt somebody had written and shared round the office. The one piece of capability the team was ever actually handed became their most-used feature, by a distance. Give people a blank box and usage falls away. Give them something built for the job in front of them, and they reach for it more than anything else. The gap here was never willingness. It wasn't even the AI. Nobody had built the tool and handed it over.
What it looks like wired all the way in
Wired all the way in, on the same subscription, this is what the system actually does:
Commercial production
An enquiry lands. The system reads the correspondence, pulls the comparable past jobs, builds the costing workbook, and drafts a client-ready proposal in the firm's own voice. Run four times in the past month, each one checked before it went out.
Quality assurance
Technical reports and proposals checked against the applicable engineering standards and against the source figures, findings scored red, amber or green, and filed as a review note, catching issues before a client ever sees them.
The routine work, unattended
The inbox triaged every morning, classified and drafted. A weekly sweep to clear the backlog. Time reconstructed and booked to the accounts. Triage alone ran seventeen times last month, without anyone asking it to.
The real gap. It isn't the £75.
One login shared by six people means nobody has their own memory in the system, nobody's work is attributed to them, and people sign each other out mid-session. That's not a tool problem. That's a setup nobody designed.
The subscription costs £75 a month, total, for the whole team. Roughly half an engineering hour a month. It was never the constraint. The constraint was that nobody had built a system around the team. On the other side of this comparison isn't a person working harder, it's the 36 tools and 58 files above: a system, built once, that runs the same way every day. The tool is already in the building and already paid for. What it's worth is the part nobody has costed yet.
If you're reading this before you spend anything
If you've already run this exact experiment, bought the subscription, watched a few people try it, watched the numbers quietly fall, and filed it under "we tried AI, it doesn't really work for a firm like ours", you haven't been wrong about what you saw. You've been looking at the wrong half of the story. The tool was never the variable. What gets built around it is.
If you're heading into a bid and need something better than a slide that says "we use AI", the honest version of that claim is above. Thirteen weeks, read straight off the account records, nothing modelled.
The part nobody has costed yet
Start with what Greycell's owner sees from inside this firm, and the figure is deliberately modest, because a small true number is worth more to you here than a big impressive one. He is on site about three days a week. Even at that, the system hands him back a few hours of those days: hours that used to go on the first pass of a proposal or a report, and now go on the work only a senior engineer can actually do.
Now a projection, and it's flagged plainly as a projection, not a measured figure. Every one of those long, by-hand sessions in the team's chat window is a one-pass job the system already does, one desk away. So the assumption is simple: if each of the five people reclaimed even a few hours a week, the way Greycell's owner has, that comes to close to a full engineer's working week handed back to the firm, every week. On a subscription that costs half an engineering hour a month.
That is the prize sitting untouched. Not a cost to be justified, but a week of senior time going unclaimed, every week, while the tool to claim it is already bought and already in the building. The gap was never the money. It was that nobody had built the system that turns the tool into the time.
Take the whole case study with you
The full piece as a PDF, 8 pages. Share with attribution, not for resale.
This is what Headroom is built to do for a firm's senior judgement, not just its admin. Over 12 months, side by side, the same kind of system gets built around your firm, until the work above (the triage, the proposal, the review, the routine that runs while you sleep) happens whether you're in the room or not.
The first step is free: a 30-minute AI Discovery Meeting, no pitch, just a look at where your firm's version of this would actually start.
Book the AI Discovery Meeting →30 minutes · no pitch.
About the author
Ian Griffiths. 30 years keeping people safe from high-voltage power and lightning. Chartered engineer (CEng). Sits on multiple British Standards committees. Runs Greycell Technical, building and running AI systems inside engineering firms, including his own.
© 2026 Greycell Technical Limited. Share with attribution, not for resale.