Search for "what is agentic ai" and most answers stay abstract: autonomy, reasoning, multi-step planning. None of that tells an operations director whether it is safe to point at a real job. So take one job, the weekly timesheet, and walk through what an AI worker would actually have to do to run it, step by step, without something that only answers questions on the side and without a rules engine that breaks the moment a rota changes.
MSD's term for this is an AI worker: software that performs one operational job end to end inside the systems already in place. The search phrase above is how people describe the same idea from outside; the rest of this article uses the working term instead, because that is the one that actually describes the mechanism.
Answering a question versus doing the job
Plain AI, in most people's daily experience, answers a question and stops. Ask it to summarise a document, draft an email, or explain a policy, and it hands back text. A person reads that text and decides what to do with it. Nothing happens in any system unless a human acts next.
An AI worker is built on the same underlying technology, pointed at a different task: instead of answering, it acts. It reads something, checks it against a rule, and then does something inside a system of record — approves a line, flags an exception, writes a record, sends it onward. That is the whole difference: producing an answer, or producing an outcome.
That distinction only means anything against a real job. So here is the timesheet.
The timesheet, run as a mechanism
A weekly timesheet, for most operations teams, is not one task. It is four: read what came in, decide whether it is right, act on that decision inside payroll, and escalate whatever cannot be resolved without a person. An AI worker built for this job does all four, in that order, every week.
Read
The worker opens the submitted timesheet (a spreadsheet, a portal export, a scanned form, whatever format the client actually uses) and extracts the entries: who worked, which shifts, which rate band, which site. This is not a chatbot summarising a document for a human to check later. The extraction is the first step in a chain that continues without a person reading it first.
Decide against written rules
Every extracted entry gets checked against the rules that already govern payroll at that organisation: the rota for that week, the agreed rate card, the client's own timesheet policy, any overtime threshold. The AI worker's judgement shows up here: it is not matching a template, it is applying a set of written rules to a specific case and reaching a decision about that case. A shift that matches the rota and the rate card passes. A shift that does not match anything on file does not.
This is the part that separates an AI worker from either extreme people usually imagine. It is not blind automation, because the decision depends on the specific entry, not a fixed template. It is not open-ended judgement either, because the rules are written down in advance by the organisation, not invented by the AI worker as it goes. The worker follows the rules that exist. It does not set them.
Act inside payroll
A passed entry gets keyed into the payroll system the client already runs — no new platform, no separate dashboard to check. The worker writes into the same system a person would have typed into by hand, in the same fields, following the same process the organisation already trusts.
This is where an AI worker earns its keep against something that only talks. A tool like that can draft the entry and hand it to someone to paste in. An AI worker built to run the timesheet keys it in itself, because keying it in is the job, not a suggestion about the job.
Escalate what it cannot resolve
Every entry that fails a check, such as no rota match, a rate outside the agreed band or a site the worker does not recognise, goes to a named person, not a queue nobody owns. The escalation carries the specific reason it stopped: which rule the entry failed to satisfy, not a generic "needs review" flag. Someone can act on that in seconds rather than starting the investigation from nothing.
Escalation is not a fallback bolted on afterwards. It is the fourth step of the same mechanism as reading, deciding and acting — the point where the worker's own written rules run out and a person's judgement has to start.
An AI worker compared with a question-answering tool
A tool that only answers questions waits to be asked. It has no job of its own, no rota to check, and no system to write into. Point it at a stack of timesheets and the best it will do is describe them back to you, or answer a question about one entry if you ask it directly. The read-decide-act-escalate chain still runs — through a person, by hand, at the pace a person can sustain it.
The difference is not intelligence. Today's question-answering tools and today's AI workers can draw on comparable underlying models. The difference is what the software is pointed at: a conversation, or a job with a defined outcome inside a system of record.
An AI worker compared with a rules-only automation tool
A conventional automation tool — a macro, a script, robotic process automation — can also read a timesheet and write to payroll, provided the format never changes and every case fits a template drawn up in advance. The moment a shift pattern is unusual, or a field is filled in slightly differently than expected, a rules-only tool either fails silently or stops the whole run.
That gap is what an AI worker is built to close. An AI worker built the same way still checks every entry against the same written rules a rules-only tool would use, but it can read a case that does not match the template exactly and still work out whether the rule applies, then escalate cleanly when it genuinely cannot decide, rather than breaking or guessing. The written rules stay in charge either way. What changes is whether an unexpected case gets handled or gets stuck.
Why the timesheet is a fair test case
Timesheets are a reasonable job to reason about publicly because the stakes of getting it wrong are visible and immediate: someone gets paid the wrong amount, or a client gets billed incorrectly, and it surfaces within a week. There is nowhere to hide a mechanism that does not actually check its work.
To be direct about where MSD stands on this: no AI worker has run for a client yet. We scope it and discuss it; we do not claim it is running for a client. The mechanism described here — read, decide against written rules, act inside payroll, escalate the rest — is how we build the timesheet job when we run it, not a record of a live deployment.
The same read-decide-act-escalate shape applies to work well beyond payroll. An AI worker can be scoped against onboarding packs, compliance chases, or invoice matching using the same four steps. Where a business runs several such jobs together, that sits alongside bespoke back-office systems rather than replacing them.
In recruitment and staffing specifically, where timesheets, compliance chasing and payroll runs repeat every single week, the case for handing one job to a worker built this way is worth seeing set out against that sector's own administration in full. See recruitment and staffing for the sector-specific detail.
The scoping conversation
None of the above is a promise that an AI worker is simple to point at a given timesheet process. Every organisation's rota, rate card and payroll system are different, and the written rules an AI worker checks against have to be the organisation's actual rules, not a generic template. Getting that right is a scoping exercise before it is a build.
That is the honest next step: a conversation about one process, in enough detail to write down its rules properly, before anything gets built against it. Get in touch to talk through what that would look like for a specific timesheet, rota or payroll run.
No AI worker has run for a client yet. We scope it and discuss it; we do not claim it is running for a client.
A few useful questions.
Is agentic AI the same as a chat assistant?
What makes AI 'agentic' rather than just automated?
Does agentic AI make decisions without a human?
