Why HR answers the same policy questions fifty times a week

You already wrote it down. They still asked you.

Open your sent items and search "holiday carry over". Count the replies. Then count how many contain the same three sentences, pasted from the same handbook page you wrote in January and updated in March.

You have the policy. You have the intranet page. Somebody built the New Starter FAQ and it now runs to fourteen sections with a contents list at the top. None of it has stopped the questions arriving one at a time — in Teams, in the corridor, at 4pm on a Friday from someone who wants to book a flight.

The question this post is here to settle is whether an AI knowledge base for HR actually fixes that, and what it costs you to find out. The thing to accept first is less comfortable: the repetition is not evidence that your documentation is bad.

The same five questions, everywhere, quarter after quarter

If repeat questions were a writing problem, you'd expect the mix to shift as guidance improves. It doesn't. Acas ran roughly 130,300 helpline calls in the April to June 2026 quarter, and their own bulletin says the percentages by topic "have remained stable over the last 12 months" [1].

The concentration is the interesting part. Discipline, dismissal and grievance accounted for around 38% of calls. Contracts took 15.9%. Diversity and discrimination 13.6% [1]. A national service, staffed by employment specialists, with published guidance available free to anyone, and the demand still clusters into a handful of topics that never go away.

Your inbox is the same shape at a smaller scale. Sick pay. Notice. Probation. Carry over. Expenses. The questions don't change because the people asking them keep changing, and each one arrives at their moment of need, not yours.

LESSON ONE: Stop treating a repeat question as a documentation defect and start treating it as normal traffic that needs somewhere to go.

“The percentage of calls about each call topic have remained stable over the last 12 months. This suggests the issues that have been affecting employees and employers over the last year have not changed much. The most common call topic this quarter was discipline, dismissal and grievance, with around 38% of calls related to this topic.”

Acas

Acas helpline data, April to June 2026

It's a retrieval problem, and retrieval has a different fix

Here's what actually happens. An employee types into the intranet search box: "do I get paid if I'm off sick Monday and Tuesday". Your handbook section is called "Absence Management: Statutory Sick Pay and qualifying days". Search matches on the word "sick" and returns the absence policy, the return-to-work form and an annexe from 2021 nobody deleted. Three results, none of which answer the question in the shape it was asked.

So they ask you, because you translate. That translation step is the whole job, and it's the bit no wiki has ever done.

An AI knowledge base for HR is the version that does it: it reads the question for what the person means rather than which words they used (the technical term is hybrid search — intention matching mixed with keyword matching), answers in a sentence or two, and shows which document and section the answer came from. That last part is not decoration. An answer about notice periods with no source attached is a rumour, and you'll end up re-checking it anyway.

The category word doing the work is AI. You already own an employee knowledge base. What you don't own is one that answers.

The people who stop asking are the expensive ones

The interruptions are the visible cost. They're also the cheap half.

SurveyMonkey's State of Curiosity 2026, run across 1,925 US workers in April 2026, found four in ten admit to pretending they understood something rather than asking the question, and half of workers have had to redo work because questions weren't asked at the start [2]. It's a US sample. The behaviour is not remotely American.

You know the internal monologue: "I'm not asking Priya again, she already explained it once." So they guess the notice period, guess the expenses limit, guess whether the probation extension needed to be in writing. HR finds out later, usually when it's a grievance rather than a question.

The same study puts the gap starkly — 94% feel safe asking questions in high-capacity cultures, against 41% in low-capacity ones [2]. Self-service doesn't win because it's efficient. It wins because it's anonymous, and a fair share of your workforce would rather ask a search box than a person.

LESSON TWO: Count the questions you never received, because those are the ones that turn into rework and claims.

“Four in ten admit to pretending they understood something rather than ask the question. Half of workers have had to redo work because questions weren't asked at the start.”

SurveyMonkey

State of Curiosity 2026

Where this genuinely falls over

An AI knowledge base answers from what you give it. Point it at a handbook that contradicts three contract templates and you get confident answers to a question your organisation hasn't actually decided. It won't reconcile that for you. It will surface the mess faster, which is useful and uncomfortable in roughly equal measure.

Scope is where most of these deployments die. Glean's Work AI Index 2026, covering 6,000 full-time digital workers, found more than half of workers say critical information they need to do their jobs is not accessible through their AI systems [3]. If the answer isn't in the sources you imported, the assistant is a very fast way of saying "I don't know".

There's also a sequencing problem specific to HR. Culture Amp's 2026 AI in HR study found HR use of AI concentrated in content creation (86%) and information synthesis (81%), with only 39% having moved AI into HR operations automation [4]. Drafting is where everyone starts. Answering employees is a different risk profile, and it deserves the caution.

One thing worth doing before you buy anything: export the last 90 days of your HR inbox, list the top twenty questions, and mark each one A if the answer already exists in a document, B if it doesn't. If most are A, you have a retrieval problem and this technology applies. If most are B, write the policies first. No tool rescues you from B.

“More than half (53%) of workers say critical information they need to do their jobs is not accessible through their AI systems.”

Glean, Work AI Institute

The Work AI Index 2026: Global

The reason this is worth revisiting in 2026

The last time you looked at company knowledge base software it probably meant a migration, a seat count and a conversation with whoever handles your IT. That's the bit that changed.

The UK Business Data Survey 2026, covering 4,450 UK businesses, found the most common reason businesses gave for using AI was to research information — 28%, described in the survey as "in place of a traditional search engine such as Google" — followed by summarising or collecting in-house information at 21% [5]. Adoption sits at 51% among small businesses and 41% among micro businesses [5]. Retrieval is already the default use. The missing piece was pointing it at your documents and nobody else's.

What that looks like now: import the handbook PDFs, the Drive folder of policies and the Notion process pages; brand it; publish it to a URL; and let anyone with a matching @company.com address join automatically. No seat admin, no SSO project, no IT ticket. Ten minutes, not a quarter.

The honest test of an AI knowledge base for HR isn't whether the answers impress you in a demo. It's whether, six weeks in, you can search your sent items for "holiday carry over" and find nothing new.

Sources

[1] Acas, Acas helpline data April to June 2026 (quarterly statistical bulletin), 2026 — https://www.acas.org.uk/page-download/2175

[2] SurveyMonkey, State of Curiosity 2026, 2026 — https://www.surveymonkey.com/curiosity/state-of-curiosity-report/

[3] Glean, Work AI Institute — The Work AI Index 2026: Global, 2026 — https://www.glean.com/work-ai-institute/reports/work-ai-index

[4] Culture Amp, 2026 AI in HR Study, 2026 — https://www.prnewswire.com/news-releases/culture-amps-2026-ai-in-hr-study-reveals-transformation-gap-task-level-tinkering-masks-opportunity-302831663.html

[5] UK Department for Science, Innovation and Technology, UK Business Data Survey 2026 (Official Statistics), 2025 to 2026 — https://www.gov.uk/government/statistics/uk-business-data-survey-2026/uk-business-data-survey-2026