AI for job duration estimates that match what actually happens

Introduction

Almost every unprofitable job traces back to a time estimate. The price was calculated from an assumed number of hours, the work took longer, and the difference came out of margin. It is not a pricing failure in the usual sense; the pricing was arithmetic performed on a wrong input.

What makes this correctable is that the correct input is already recorded. Most businesses know how long jobs actually took, in timesheets, job cards or scheduling records, and almost none compare that against what was estimated. The comparison is straightforward, the pattern is consistent, and correcting it improves both profitability and delivery reliability at once.

1. AI for job duration estimates starts with the estimate-versus-actual comparison

The single most valuable table you can build.

Estimated hours against actual hours, per completed job, grouped by job type. Firms doing this for the first time typically find a consistent overrun in specific categories, which is a correctable arithmetic error rather than a mystery.

2. Look for consistent bias before variability

Bias is the easy win.

If a job type overruns by twenty per cent on average, the estimate is twenty per cent low and can be corrected immediately. Variability around a correct average is a harder problem and a smaller one.

3. Include the time that surrounds the work

The commonest omission.

Travel, setup, access delays, waiting for materials, clearing up, paperwork, and the conversation with the customer at the end. These are real hours that estimates habitually exclude, and together they frequently explain the whole gap.

4. Find the factors that actually drive duration

This is where analysis adds something.

Site access, floor level, building age, equipment type, customer preparedness, distance, weather, whether you have worked there before. A model that includes two or three real drivers outperforms a flat rate substantially.

5. Record actual time honestly, which is the hard part

The data problem behind everything.

If people record what was estimated rather than what happened, the whole exercise is circular. This requires a stated purpose — better estimates, not performance monitoring — and it will not work if the figures are used to criticise individuals.

6. Estimate a range, and price the upper part

Practical protection.

A likely duration and a realistic worst case. Pricing near the upper end, or including a contingency line, converts variability from a margin problem into a stated commercial term.

7. Feed the estimate into scheduling, not just pricing

Two uses, one input.

Better duration estimates make the schedule achievable, which reduces the overruns that cascade through a day. This is frequently the larger benefit and it is available from the same work.

8. Handle the unusual jobs separately

Automation should decline to answer.

A job unlike anything in your history needs an experienced person and a bigger contingency, not a generated estimate. A system that flags "insufficient comparable data" is more valuable than one that always produces a number.

9. Review the estimates every quarter

Conditions change.

New equipment, new methods, a different team, changed materials. An estimating table that is not revisited drifts out of date silently, and the drift shows up as declining margin nobody can explain.

Watch what happens to your win rate as estimates rise. Correcting a systematic underestimate raises prices, and some of the work you were winning was work you were winning because you were cheap.

Conclusion

Compare estimated against actual hours by job type, because that comparison explains most unprofitable work.

Correct any consistent bias first because it is simple arithmetic, include travel, setup, waiting and clearing up in the estimate, identify the two or three factors that genuinely drive duration, make time recording honest by using it for estimating rather than for monitoring people, quote a range and price towards the upper end, use the improved estimates in the schedule as well as the price, flag unusual jobs for a person instead of generating a number, and review the table quarterly.

Related reading

• AI for supplier price comparison across what you actually buy

• AI and job roles that change rather than go, handled openly

• Explaining veterinary costs to owners at the worst possible moment

• Before and after photos for trades cost ten seconds a job

• How buyers actually first hear about you, checked against evidence