What the AI in a Workforce System Actually Does, Function by Function
“AI-powered” has become a category label rather than a description, which is a problem for a buyer, because in a workforce platform it covers at least four different technologies with four different failure modes. Evaluating them as one thing is how a company ends up satisfied with a demo and surprised in month three.
Here is what is actually under the word, what each part can honestly promise, and the one question worth asking about each.
1. Face matching — a comparison with a threshold
This is the most visible and the least mysterious. A face presented at check-in is turned into a numeric representation and compared with the one stored when the employee enrolled. The comparison produces a similarity score. A threshold turns that score into a yes or a no.
Everything that matters follows from that last sentence. The threshold is a business decision wearing a technical costume: set it high and genuine staff get rejected in bad light or after a haircut; set it low and the verification stops meaning much. There is no setting that avoids both, only a choice about which error you would rather have.
Ask: where does the threshold sit, who is allowed to move it, and — more important — what happens on a failed match? A rejection that strands someone at a gate and a rejection that queues them for supervisor confirmation are the same accuracy and entirely different products. How verification works at check-in is worth reading for the second half of that answer specifically.
2. Roster suggestion — an optimiser with your rules in it
When a system “suggests a roster”, it is searching for an arrangement that satisfies constraints you gave it: coverage per site, rest requirements, contracted hours, skills, stated preferences. It is doing quickly what a good scheduler does slowly.
Its limitation is not intelligence but knowledge. It does not know that these two people should not work the same site, that this client dislikes rotation, or that someone is going through something difficult at home. A scheduler holds dozens of those facts and writes none of them down.
Which makes the right posture clear: the suggestion is a first draft that removes the blank page, and the scheduler still edits it. A system that presents the output as final is overselling; one that makes editing awkward is worse than no suggestion at all.
Ask: how long does it take to override a suggestion, and does the system learn from the override or discard it?
3. Anomaly flags — statistics, and a tone problem
Flagging unusual patterns is ordinary statistics: this person’s check-ins have moved, this site’s hours drifted, this correction rate is above its own baseline. Useful, cheap, and the part most likely to cause harm — not because the maths is wrong but because of how the output is framed.
A pattern that differs from its history is usually explained by something dull: a shift change nobody recorded, a site that moved its start time, a supervisor on leave. Presenting that as suspected misconduct puts a manager in an accusatory conversation on the strength of an outlier, and the damage to the relationship outlasts the correction.
Ask: what does a flag look like to the manager who receives it? “Worth a look” and “possible fraud” are the same computation and different companies.
4. Language assistance — the newest and the most oversold
The fourth kind is a language model: draft this announcement, summarise this month, explain this policy, set up this leave type by describing it. Genuinely useful for drafting and navigation — the tasks where a good-enough first version saves real time and a person reads it before it goes anywhere.
The line worth drawing is between drafting and deciding. A model that helps write an announcement is doing what it is good at. A model that approves leave, or decides an attendance dispute, is being asked to be accountable, and it cannot be. Someone has to be able to say why a decision was made, and “the assistant suggested it” is not an answer that survives being challenged.
Ask: what can it do without a human confirming? If that list is longer than drafting and summarising, ask who is accountable when it is wrong.
The two questions that cover all four
Rather than evaluating “the AI”, evaluate each function against the same two questions:
- What does it decide on its own, and what does it only suggest? The boundary matters far more than the accuracy, because the accuracy is roughly comparable across serious products and the boundary is not.
- What happens when it is wrong? Every one of these is wrong sometimes. A product designed around that fact behaves visibly differently from one that assumes it will not be.
Those two also happen to be the questions a data-protection review will ask, in different words, so answering them early is not wasted work. What runs where, and what it is allowed to decide sets out our answers; the full feature list shows which modules each function touches.
None of this is an argument against AI in workforce software. It is an argument for asking which of the four you are being sold, because the word on its own has stopped narrowing anything down.
Frequently asked questions
Is face recognition in an attendance system the same kind of AI as a chat assistant?+
No. Face matching is a classifier comparing one face against one stored template and returning a similarity score, which a threshold turns into yes or no. A chat assistant is a language model generating text. They fail in completely different ways, and conflating them makes both harder to evaluate.
What should we ask a vendor about face matching?+
Ask where the threshold sits, who can move it, and what happens on a failed match. A system that silently rejects a person at the gate and a system that queues them for supervisor confirmation have the same accuracy and completely different consequences on a wet morning.
Can AI build our roster for us?+
It can propose one against the rules and preferences you encode. It cannot know that two people should not be on the same site, unless someone told it. Treat the output as a first draft that saves the blank page, not as a decision.
Is an anomaly flag an accusation?+
It should never be presented as one. A flag says a pattern differs from its own history — which is often explained by a schedule change nobody recorded. Systems that present statistical outliers as suspected misconduct create more damage than the misconduct they find.