The Assistant Read Every Policy. That Was the Problem.

Part I

The Demo

The demo took eleven minutes and nobody in the room looked at their phone.

The vendor typed a question into a clean white box: How much parental leave am I entitled to? The answer appeared before he had finished turning to face the board. It was correct, polite, and footnoted. It cited the policy by name.

The COO asked whether it could also tell a manager how to run a disciplinary hearing. It could. Whether it could explain the bonus scheme. It could. Whether it would be available at three in the morning, in four languages, without a ticket. It would.

Lena, the HR director, was the only person in the room who did not smile. She was not against it. She had simply spent nine years in that company's shared drives, and she knew what the assistant was about to read.

The board approved the pilot before lunch.

Part II

The Third Week

For two weeks it was the most popular thing HR had ever launched. Ticket volumes fell. Managers stopped forwarding the same four questions. Someone in finance sent a thank-you note, which had never happened before.

In the third week, an engineer asked the assistant how much parental leave he would get, and planned the next six months of his family's life around the answer.

The answer was quoted accurately, cited correctly, and formatted beautifully. It came from a policy that had been replaced two years earlier.

When Lena went looking, she found three parental leave policies. The old one, still sitting in a folder called Policies FINAL. The current one, approved by the board, stored in a different folder called Policies (new). And a regional draft that nobody had ever approved, which a local manager had been quietly using as if it were real.

The assistant had not been wrong about any of them. It had read every policy. It simply had no way of knowing which one was in force, because nobody had ever told the company that either.

Then the other questions came in. Asked who owned the leaver process, it named a manager who had left the year before. Asked for the grievance procedure, it summarised two documents that contradicted each other on the deadline for an appeal, and politely reconciled them into a third answer that matched neither.

Part III

The Wrong Lesson

The COO drew the obvious conclusion. AI was not ready for HR. Switch it off, write a careful email, try again in a year.

Lena drew the uncomfortable one. The assistant was not broken. It was the most accurate audit her function had ever had. It had done, at machine speed, exactly what her own team did every day: answer from whatever it could find, and sound sure of it. The only difference was that the machine left a log.

There is a line we have written before, in the story of the CHRO who was asked how mature her function was: governance is not the alternative to AI in HR, it is the precondition. Lena agreed with it. She also noticed something it leaves out.

Precondition does not mean wait. The work of governing an HR function is overwhelmingly a particular kind of work. Reading hundreds of documents. Comparing versions. Sorting things into categories. Checking who owns what. Reminding people that a review date has passed. Drafting the first version of something so that an expert can tear it apart.

That is why governance never gets done. Not because anybody disagrees with it, but because it is mostly reading, and nobody has the time to read.

It is also, almost exactly, the list of things a language model does well.

Part IV

She Turned It Around

So she did not switch the assistant off. She turned it around.

For six weeks it answered no employees at all. Instead of pointing it at the workforce, she pointed it at the function, and gave it the jobs her team had been postponing for years. It got eight of them.

She gave it the whole estate: every policy, procedure, form, template, and folder name, plus a year of HR tickets. Its job was to propose where each one belonged in the taxonomy of process groups, families and processes.

It found that induction, onboarding, new joiner programme and Day One were one process wearing four names. It found documents that belonged to nothing, and processes that had no documents at all. It decided nothing. Every classification was a proposal, and a person accepted or rejected it. The work of months became the work of a few weeks of deciding.

Next it compared every version of every document against every other. It produced a register of what was actually in force, and a list of what was not: documents with no recorded approver, documents past their review date, and documents that contradicted each other.

The most useful finding was not the duplicate policies. It was the gap between policies and the forms that carried them out. One policy granted a benefit that the form people actually filled in made impossible to claim. Nobody had noticed, because nobody had ever read both on the same day.

It cross-checked every process owner against the list of current employees. Eleven processes were owned, on paper, by people who no longer worked there. Several more had a responsible person and nobody accountable.

Where a process had never been mapped, it drafted a swimlane map from the existing procedures and a recorded conversation with the expert. The experts, who had refused for years to author a map from a blank page, were merciless when correcting a draft. Drafts had become cheap. Their judgement had not.

When a new pay transparency requirement landed, the old method was a fortnight of meetings to work out what it affected. This time she asked the assistant to list every policy, procedure, form, process and job architecture document the change would touch, with the clause and the reason for each.

A specialist checked the list and removed two false positives in an afternoon. The impact assessment was finished before the first meeting would have been scheduled.

Her effectiveness assessment asked four groups to score the same processes: HR itself, the executives, managers, and employees. The scores were easy. The free text was not, so for years nobody had read it properly.

The assistant clustered every comment into themes, by group, and showed where the groups disagreed. HR rated its own recruitment process highly. The executives' comments were almost entirely about the time it took. It drafted the readout. It never saw a name: the data it was given was aggregated before it arrived, which is the only reason people had answered honestly in the first place.

Every quarter, her HR strategy was locked and reviewed on the record. The assistant compared this quarter's version with the last one and wrote the difference in plain language: which objectives had changed, which initiatives had no key result, which key results had no owner, and which had quietly disappeared without a decision.

The board pack used to take her a week. Now it took her a morning, most of which she spent arguing with the draft.

Who approved this policy, and when? Which version was in force on the date of this claim? Who reviewed this process last? These used to be excavations. With the register in place, they became lookups, and the assistant answered them with a citation to the record, never from memory.

It had read-only access. It could find the answer. It could not change it.

Six weeks later she switched it back on for the workforce, with one change that mattered more than every other.

It now read only the register of approved, in-force documents. Every answer named the version it came from and the person who owned it. And when there was no governed answer, it no longer improvised one. It said so, and told the employee who to ask.

I don't know, but this is the person who does turned out to be the most valuable sentence the assistant ever produced.

Part V

The Line She Drew

There was a ninth job the vendor offered, and she refused it.

The platform could score candidates, suggest performance ratings, flag employees as flight risks, and recommend who should be in the next round of redundancies. Lena allowed none of it. Not because the machine would necessarily be worse at it than a tired manager, but because a decision about a person needs a person who can be held to account for it. A model cannot be asked to explain itself at an employment tribunal.

She was not alone in that view. Under the EU AI Act, AI used to recruit, promote, dismiss, allocate work to, or monitor employees is classed as high-risk, with obligations to match. Other jurisdictions are heading the same way.

So she wrote the line down, on one page.

Under the table, five rules. The assistant reads, it never writes to the record. It sees only what the person asking is allowed to see. Survey data reaches it as aggregates only. Every answer cites its source, or it does not give one. And anything that changes the record is signed by a named human being.

Part VI

The Moral, Stated Plainly

Because fables should not be coy.

AI does not ease HR governance by being clever. It eases it by being tireless at the parts of governance that are reading, comparing and reminding, which turn out to be most of it.

Point an assistant at an ungoverned function and it will read the mess back to your employees, quickly, politely and with a citation. Point it at the work of governing, and the mess starts to shrink. You do not need a governed function to start using AI. You need one before you can trust its answers.

The order matters. Machine on the governing first. Machine on the people second. Decisions about people, never.

Three questions, before you switch yours on.

Most HR functions will not discover how governed they are from an audit. They will discover it in the third week after the demo.