AI and the daily production schedule that people follow

5 days ago
3 min read

Introduction

Sequencing work is a genuinely hard mathematical problem and a genuinely simple organisational one, and small manufacturers usually fail at the second. A schedule is produced, the first urgent phone call arrives, the sequence is abandoned, and by mid-morning the floor is working from the same instinct it used before there was a schedule.

The optimisation is worth having: better sequencing reduces changeover time, work in progress and late deliveries at once, and the gain is real. But a mathematically superior sequence that nobody follows is worth less than a crude one everyone does, so the ordering of effort should be trust first and precision second.

1. AI and the daily production schedule are limited by whether it is followed

Adherence before optimisation.

Measure schedule adherence for a week — how much of the planned sequence actually ran in order. If it is below half, the problem is not the algorithm and a better one will change nothing.

2. Group by changeover, which is the largest hidden cost

The most common available gain.

Sequencing jobs by material, colour, tooling or setup can remove hours of changeover a week. This is the gain that arrives before any sophistication and the one people can see the logic of.

3. Get the real constraints into the plan

Otherwise the plan is fiction.

Tooling availability, operator skills, curing and drying times, shift boundaries, material arrival dates, and machine maintenance windows. A sequence that violates any of these is discarded by the person on the floor, correctly.

4. Use recorded run times, not standard times

Standards drift.

Standard times set years ago no longer reflect the machines, the materials or the people. A schedule built on them is optimistic by a consistent margin, which is why it fails every afternoon.

5. Leave a slot for the urgent job

Because there will be one.

Reserving capacity for the interruption you know is coming means the plan survives it. Schedules with no slack are broken by the first exception, and once broken they are abandoned entirely for the day.

6. Publish the sequence where people work

Visibility is most of adherence.

A printed or displayed sequence at the machine, updated once, with a clear order. Schedules that live in an office system are not schedules; they are intentions held by one person.

7. Give the supervisor authority to resequence

With a recorded reason.

They will know about a machine running badly, an operator who is faster on a particular job, or a material issue. Recording the overrides gives you the list of constraints your plan is missing, which is the most useful improvement data available.

8. Measure late deliveries and work in progress, not utilisation

Utilisation misleads.

Keeping every machine busy produces high utilisation, large queues and late deliveries. On-time delivery and total work in progress are the outcomes that matter, and they frequently improve when utilisation falls.

9. Reschedule at a fixed frequency, not continuously

Stability has value.

A sequence that changes every hour cannot be worked to. Rescheduling once a day, or once a shift, with exceptions handled by the reserved slot, is more effective than continuous reoptimisation.

Watch what happens to quality when the sequence tightens. Removing slack sometimes removes the time in which people were checking their work, and the defects appear a fortnight later.

Conclusion

Fix adherence before optimisation, because an unfollowed sequence gains nothing.

Group jobs by changeover to capture the largest visible saving, get tooling, skills, curing times and shift boundaries into the plan so it is workable, build it on recorded run times rather than old standards, reserve a slot for the urgent job you know is coming, publish the sequence where the work happens, let the supervisor resequence with a recorded reason, measure on-time delivery and work in progress rather than utilisation, and reschedule on a fixed cadence rather than continuously.