AI for Employee Onboarding in Manufacturing

It is a quarter past six on a Monday and the line is one operator short. The supervisor looks at the people who came in through the agency that morning, picks the one who seems most confident, and puts him on the palletiser. Nobody in the building can say whether that person has ever been signed off to run a palletiser, or whether he has worked here before under a different supplier reference. The line runs. The only record is a name on a sheet in the supervisor's office.

That is manufacturing onboarding compressed into one decision: a specific person, a specific machine, and a question about authorisation that has to be answerable now and provable later.

What ready to work means on a production site

In an office, onboarding is finished when the accounts work. On a production site it is finished when someone with the authority has certified that a named person is safe and competent on a named task, and that certification exists as a record you can produce later. Until then the starter is a supervision burden and an exposure.

The certificate is not paperwork attached to onboarding, it is the output of onboarding. Everything an agent does here serves getting that record created sooner, keeping it current, and making it answerable when asked.

Site induction and the queue that forms

Most sites run induction as a scheduled session. Starters do not arrive on a schedule. Someone starting midweek waits and spends days doing nothing useful, or gets a rushed corridor version nobody records properly, which is worse because the gap is invisible.

An agent breaks the session into what needs a room and what does not. Reading, acknowledgements and the site rules a person can absorb alone move ahead of arrival, and the training lead sees the queue building rather than discovering it.

The same run handles provisioning: protective equipment in stores, a locker on another sheet, tooling in a third log, an access card in the security system, none of which share an identity. The agent assembles the set from role and area, raises each request with its owner, and tracks one status against the start date.

What it must never do is mark an induction as delivered. It tracks completion against your records, it does not create it.

The competency and authorisation matrix as a live record

Almost every site has one, usually a spreadsheet, owned by one person, and behind. Every row has to answer four things: who is authorised, on what task or machine, who signed them off, and when that authorisation expires.

An agent reads the sources that already hold fragments of this: assessment forms, training completions, scanned sign off sheets, the learning system, the area's own tracker. It reconciles them into one current view and, where two sources disagree, flags the contradiction to the competency lead rather than quietly choosing a winner.

Expiry is the quiet killer. Authorisations lapse unnoticed and the person keeps doing the job. Surface lapses far enough ahead that reassessment can be booked on a shift the person is actually rostered to work, a harder constraint than it sounds on a rotating pattern.

The moment the shift runs short

The check that matters is not the one at induction. It is the one at deployment, made by a supervisor with a line stopped and no time.

So the useful mechanic is an answer, not a reminder. Is this person authorised for this task on this machine today. Yes or no, what is missing if no, and who else on shift is authorised, on a phone in the channel the supervisor already uses.

The agent cannot stop a supervisor moving a body to a line. What it can do is make the position known before the decision rather than after the incident, and ensure that when the decision is taken anyway a record exists with a named person attached. That beats a verbal choice nobody can reconstruct.

Safety induction and the refresher that lapses quietly

Initial briefings usually happen. It is the refresher cycle that decays. Your isolation procedures, your permits to work, the hazards of an area, all briefed once at induction and then assumed.

An agent holds the evidence that each briefing happened: who attended, who delivered it, on what date, against which version of the procedure. That last field is the one most sites do not capture and the one that matters most. When a procedure is revised, the agent identifies everyone briefed against the superseded version and produces the rebriefing list.

It cannot tell you whether the briefing was understood. An agent treating attendance as evidence of competence produces a comfortable fiction.

Agency, contract and seasonal labour

Here the record is split in half. The supplier holds identity, prior placements and some training history. The site holds induction and authorisation. Neither sees the other. The same person leaves in autumn and returns in spring under a different reference, and the site inducts them from zero or, worse, assumes they are already inducted because a supervisor recognises the face.

An agent should look for a prior record before anything else happens, match on several signals rather than a name, and put a candidate match to a human instead of merging automatically. Once confirmed, it requests from the supplier only what is missing or lapsed and maintains one answer to a question a plant manager cannot answer today: of the agency workers on site right now, which hold a valid induction.

Where that data can sit and what the supplier must hand over is a commercial question before it is a technical one.

Contractors and visiting engineers

Contractors are inducted to a different and usually shorter standard, and the record still has to exist. An agent can run the pre arrival pack, check the documents your site requires are present and in date, and produce the day's list of who is on site, in which area, under whose supervision. That list gets asked for after an incident.

The first week across a rotating pattern

A new starter's first week frequently lands across nights and days. Their supervisor changes midweek. The only handover of how this person is doing is a remark at shift change, and it is gone.

An agent can carry a short structured record forward with the person: what was covered, what was not, what the incoming supervisor should watch. Prompted out of the outgoing supervisor in under a minute. Small mechanic, disproportionate effect, because a starter quietly struggling on nights is invisible until they stop coming in.

Training material and answering the starter from your own documents

Most procedures are written to satisfy an auditor. They are accurate and close to useless for a person doing the task for the first time.

An agent can draft a task level version from your own documents: your procedure, your machine manual, your recorded settings, your deviation records. Never invented, never published as a controlled document, and reviewed by a competent person before anyone learns from it. The gain is a draft to edit rather than a blank page.

The same corpus answers the starter's own questions, with the source document named, and with a tested answer of not knowing when the question falls outside what the site has written down.

Early leaver signals in the first weeks

In a high churn plant this is where the money is. Most of the people who leave do it early, and the signals are already in your systems: shifts not turned up for, training not completed, no authorisation gained after several weeks, repeated moves between areas.

An agent can surface that pattern with context and draft a short conversation opener for the supervisor. It should never send anything itself: a message about someone's attendance reads badly when it is obviously machine written.

The honest limit: it can show you the pattern, not fix the reason. If people leave because the area is hot, the training was rushed and the supervisor has no time for them, a better alert changes nothing.

Producing the evidence when someone asks for it

An auditor names a person and a date and asks what they were authorised to do, what training supported it, who signed it off and against which version of the procedure. Today that is folder archaeology across several systems and a filing cabinet.

Retrieval is what agents are genuinely good at. If onboarding produces the evidence as a by product, with sources linked rather than retyped, the pack assembles in minutes. Keep the agent's own logs under the retention rules of the records they describe.

Where AI fails here

It cannot certify competence. Only a qualified assessor watching a person do the task can, and no number of completed modules substitutes.

It cannot observe whether a safety briefing was understood. It can prove the briefing happened. Those are different claims and must not be blurred. Nor can it replace a supervisor watching a new starter work: the judgement that someone is fine on the machine but rushing, or capable but nervous, comes from a person on the floor.

A confidently wrong authorisation record is worse than no record, because people stop checking. The agent must answer unknown rather than infer, and must never derive an authorisation from a training completion. Those are separate facts, and confusing them is what puts someone on a machine they should not be on.

If your matrix has never been written down, or your training records are a folder per area with no shared identity, the agent has nothing to reconcile. That is the usual reason this work stalls, and the fix is yours before it is anyone's software.

And if operators read it as surveillance rather than the thing that gets them authorised on more machines, it will be ignored inside a fortnight.

Where to start

Take one area and one shift pattern. Start read only: reconcile the authorisation matrix from every source holding a fragment of it, and let the competency lead see the contradictions and the lapses. That alone is a fair test of whether your data supports anything more.

Then add the deployment check, the question with real consequences. Test it on your own contested cases before launch, not the easy ones, and measure the failure that matters: how often it says authorised when the person is not. Run it in shadow while your team works as they always have.

If you run a plant and want a straight read on whether your records could support this, write to Adopt AI Partners at hello@adoptaipartners.com.