SurveyCo* is a construction consultancy in the North of England that manages the commercial side of projects for its clients. Its teams build and update project programmes, prepare monthly applications for payment, and price claims for extra time and money when the scope of a job changes. Most of that work falls to junior staff, who turn handwritten site records and contract schedules into figures a senior colleague can sign off. Leadership believed AI could take on some of it but had no way to judge where it would repay the cost. A founder put the question to us at our first meeting: “Is it something that you can buy off the shelf? Is it something that needs [a] million pounds worth of development?”
How we ran the week
We spent one week on that question, two days of it on site. Instead of asking people to describe their processes, we sat with them while they worked through a live matter using its own files. We opened each person’s timesheet with them and observed the firm’s finance, planning and spreadsheet tools in ordinary use. We then sorted every step into one of three groups: work that could be automated, work where AI could produce a first draft for a person to check, and work that has to remain a matter of professional judgement. Seven sessions with six people covered all three of the firm’s disciplines.
What the payments work showed
The clearest finding came from the work on applications for payment. Allocation sheets, which record who worked where and on what, arrive from site largely handwritten and used to be retyped into spreadsheets by hand. One team had automated that step and the retyping had stopped, but the hours it used to take were now spent checking what the automation produced: about a day a week, according to the staff who do it. Part of the reason was the automation itself, which broke without warning and had on occasion dropped whole days of data. The other reason was that retyping a sheet by hand had been catching errors as it went, a check nobody had ever recorded as a task. Once the typing stopped, that check had to be done separately. Several people described the same pattern in separate sessions.
None of this could be costed. SurveyCo’s time system recorded that an hour had been worked, but not what it was spent on. Without that, there was no reliable basis for putting a pound figure against any of the fifteen points where we found time being lost. The firm’s licence already included a feature that builds timesheets from calendar and email activity, but it had never been switched on.
The diagnosis also raised a question that only the firm’s leadership could settle: where checking numbers ends and charging for professional judgement begins. Seven of the improvements we identified depended on where that line is drawn.
What we recommended
We aimed the recommendations at the checking that the automation had created, and used tools SurveyCo already paid for wherever we could. Three first steps needed no build:
- Record time by task, presented to staff as a way to evidence their work and not as monitoring.
- Set up a company AI account governed by a written use policy, with client material kept out of it.
- Check quoted rates against the contract with an AI prompt, replacing a check staff had been doing by hand on every claim.
The longer-term work is to give that AI account the firm’s own contracts, rates and methods to draw on, so that its answers reflect how SurveyCo works.
SurveyCo has since engaged us to set up the AI account and to run a ten-person leadership trial starting in October 2026.
*Name changed to protect client confidentiality.