Recorded hours tell you where a job has been. The only number that tells you where it is going is the one almost nobody maintains between billing cycles.
Key takeaways
A job is over budget when the cost already committed plus the effort still required to finish exceeds the fee — not when time recorded passes the estimate. Most firms discover overruns late because they watch the second number, which only turns red once the money is already gone. AI is useful here in an unglamorous way: it can recompute the forward number every night from records the firm already keeps, and flag the job in the week the crossover happens rather than the month after.
The decision that follows is not the machine’s. Whether to absorb the overrun, re-scope, or ask the client for a variation is a professional and commercial judgement, and in a regulated practice the person who signs the advice owns it.
Three inputs, all of which a working practice already has in some form:
Add the three and compare to the fee. That is an estimate at completion, and it is the only figure worth putting in front of a principal weekly.
It adds currency and reach. A model can read the deliverable register, the time narratives and the subconsultant correspondence, and produce a nightly estimate at completion for every open job with the reasoning attached. It can flag the crossover, draft the internal note, and draft the client-facing variation request. On a practice with sixty open jobs that is the difference between reviewing the four that matter and reviewing none.
What it must not do is present a projection as a finding. The Engineers Canada guideline on the code of ethics — itself described as a synthesis of the individual regulators’ codes, intended as a general guide regardless of jurisdiction — states that registrants should provide professional statements that distinguish between facts, assumptions and opinions, and should act as faithful agents of their clients or employers. A remaining-effort estimate produced by a model is an assumption built on other assumptions, and a variance report that does not label it as one is a professional statement that fails that test.
The code that actually binds you is your provincial regulator’s — Professional Engineers Ontario, APEGA in Alberta, Engineers and Geoscientists British Columbia, and their equivalents elsewhere. Read yours before you circulate an automated report to clients.
An overrun discovered internally and absorbed silently is a margin problem. An overrun discovered internally, worked through, and invoiced afterwards is a dispute. Treadstone Law’s article on change-order disputes sets out how change-order disputes run when the extra work was done before the paperwork caught up, and its article on quantum meruit claims explains what a claim for the value of work performed without a clear contractual entitlement actually involves — which is a good deal more effort than sending a variation request would have been.
So build the workflow around the notification, not the report. The moment the forward number crosses the fee, the system should produce a draft variation request naming the change, the reason, the additional fee and the effect on programme, and put it in front of the person entitled to send it. If your process ends at a dashboard, the dashboard will be admired and the work will be done anyway.
One drafting point for the invoice that follows: if you intend to charge interest on a late-paid variation, section 4 of the Interest Act caps recoverable interest at five per cent a year, outside mortgages on real property and hypothecs on immovables, where a written contract expresses a rate for a period of less than a year without stating the equivalent yearly rate. Treadstone Law on late-payment interest clauses covers the drafting.
A twelve-person Ontario engineering practice runs job costing monthly, from timesheets, and reviews it at a partners’ meeting. Two jobs in a year go badly over and both are discovered at the same point: when the subconsultant’s invoice arrives.
The change is to capture commitments at the moment a subconsultant is instructed rather than when they invoice, and to hold remaining effort as a deliverable list rather than a percentage. A nightly job recomputes the estimate at completion and writes a one-line movement note per job. A model drafts the variance narrative for any job whose forward number moved more than a set amount, and the project lead either accepts the narrative or rewrites it.
Two things become countable that were not before: the number of days between a commitment being made and appearing in the job cost, and the number of jobs whose first recorded overrun signal came from an external invoice rather than from the firm’s own numbers. Those are the honest measures of whether the change worked.
Should the client see the forward number?
Not raw. Clients should see a variation request with a reason and an amount. An internal projection shared without context becomes a negotiating position you did not intend to take.
Can AI estimate the remaining effort?
It can propose one from comparable closed jobs and say which ones it used. Treat it as a prompt for the project lead’s own estimate, not a replacement for it, and record which of the two went into the report.
What if the overrun is our own error?
Then the forward number still needs to be right, because the firm has to decide what to do about a loss-making job. Suppressing the signal to avoid the conversation is how one bad job becomes a bad year.
A 30-minute call is enough to tell you whether AI pays for itself here.
