The first call is on time and everything after it slides

The afternoon runs late because every job in the morning ran slightly long and nothing in the day was built to absorb it. Six to twelve minutes per stop is invisible on any single job and completely determines where the fourth stop starts. This is one cause repeated, not four unrelated bad days, and the two have different repairs — so prove which one you have first.

The tell is that the failure is ordered. Genuine chaos scatters lateness across positions and technicians. This does not: first stops land, second stops wobble, third and fourth are consistently late by a growing amount. That ordering is already sitting in the timestamps your field service platform writes on every job.

Plot arrival delta against stop position for a fortnight

One export, one pivot, one afternoon. Arrival delta is actual arrival minus the start of the promised window, in signed minutes, so early shows as negative. Stop position is that job's ordinal in that technician's day, which you may have to derive rather than read.

1. Export ten working days of completed jobs with technician, date, promised window start, actual arrival, actual completion, job type and site identifier.
2. Derive stop position by sorting each technician-day by actual arrival and numbering from 1. Do not use the planned sequence — the point is to compare what happened against what was planned.
3. Compute median delta per stop position. Report the median, not the mean: one two-hour disaster drags a mean and tells you nothing about a typical day.
4. Repeat the pivot split by job type. A diagnostic call and a full changeout do not overrun by the same amount, and that split is what turns the finding into a fix.
5. Add a third pivot on day of week. Monday and Friday behave differently from the middle of the week, and mixing them hides both.

Read the shape before the numbers. Four shapes cover almost every result, and three of the four are not this page's problem at all.

If your plot is spiky rather than rising, stop here: the fix is reserved capacity and bump rules, and it is set out in one emergency call and the rest of the day. If position 1 is already late, the question is what the board thinks the day starts with, which is settled in does the drive from home count as the first stop.

Nine minutes a job is the whole afternoon

This survives so long unfixed because no individual number in it looks wrong. Nobody defends a nine-minute overrun, and nobody can point at the job that ruined the day, because there is not one.

There is no late job. There is a booked duration that was never measured, repeated four times, and an afternoon customer standing at the window.

Where the missing minutes actually go

In rough order of how much time they account for. Each is separately measurable, which is the point: you can size them rather than argue about them.

Two rows carry a caveat. Travel is the front end of a larger problem — a matrix computed once at plan time diverges further as the day goes on, and it needs its own measurement, in the drive times are wrong by eleven o'clock. And a job waiting on a part is not an overrun at all: it is a job that should not have been on the board, argued in a job waiting on a part.

Wrap-up is the row most operators underrate. In several trades a stop is not finished when the work is — there is a record to complete, a regulated form to file, an inspection to book — and those minutes are consumed on site. When the day is behind they are also the first thing dropped, which is how you get closed jobs whose permits nobody followed through. Slippage and paperwork quality are one problem seen from two ends.

Recalibrate duration by job type, not by technician

The instinct after this measurement is to adjust estimates for the technicians who run over. Resist it. Duration varies far more by what the job is and where it is than by who does it, and per-technician adjustments become a performance conversation that stops the recalibration dead.

1. Pull 90 days of completed jobs and compute p50 and p90 of the en-route-to-complete span, grouped by job type. Ninety days covers the tail of a job type you run twice a week.
2. Split any job type whose p90 is more than double its p50. That spread means two different jobs are wearing one label, and no single number will schedule both.
3. Add a site dimension where it matters. A rooftop unit on a commercial building and the same unit in a bungalow are the same work and not the same duration.
4. Plan on p50 and hold the p50-to-p90 difference as buffer. Planning on p90 sounds safe and quietly sells a third of your capacity to nothing.
5. Replay the last fortnight of boards against the new numbers before publishing them. If yesterday's board could not have been run, that is the finding, and it belongs in front of whoever sets daily job targets.

One consequence is why recalibrations get reversed. Honest durations reduce the jobs a technician-day holds. That is not a loss — those jobs were never completed on time — but it reads as one on a utilisation report, so argue it before the change lands, using the dispatch board as a capacity model as the frame.

Put the slack somewhere the board can see it

Recalibrated durations remove the drift. They do not create room to recover from the one job that genuinely overruns, and undeclared slack gets sold twice — once by the schedule and once by whoever books over the top of it.

• A per-stop allowance inside the duration. Access, parking and wrap-up folded into the booked figure so the board packs the real span, not the working span.
• A held block, not a habit. Reserved minutes on the technician-day with a name and an expiry, which the board enforces and an intake agent can read before promising anything.
• Slack at the tail, weighted late. Buffer in front of the last two stops of a day is worth more than buffer at 09:00, because that is where the accumulated error lands.
• A trigger, not a hope. When cumulative slippage passes a stated number — 20 minutes is a reasonable start — the day stops absorbing silently and the affected customers get told, a separate discipline covered in telling the customers who slipped.

That trigger is the piece most worth automating first. Watching a running board against its plan, detecting the crossing and firing the right message to the right customers is narrow, rule-shaped work with a clear owner — the kind of thing we scope as AI agents and automation rather than a platform replacement. The same live-decision pattern, borrowed from routed delivery work, is in AI in logistics operations.

What better durations will not fix

Recalibration removes systematic drift. It does nothing about an emergency landing on a full board, or a travel model that was wrong before the day started. Fix durations first anyway: both are easier to see once the baseline stops lying, and both have their own measurement in dispatch, scheduling and route sequencing, inside our field service and trades work.

It also will not fix a duration measured against the wrong clock. Which of a job's four timestamps the board should pack — dispatched, en route, on site, complete — is worth settling before recalibrating, and it is answered in four clocks on one job.

• dispatch
• scheduling
• arrival windows
• job duration