An HR answer bot without an IT project or a new HR system

You already fixed the handbook. The questions kept coming.

The advice is always the same. Restructure the handbook. Add a contents page. Move it into Notion, tag the pages, pin the link in the all-staff channel, mention it again at induction. You did all of that. It's a good handbook. The policies are current, the process docs are written up, and someone spent a fortnight in March reformatting the whole thing.

Then last Tuesday: "Sorry, quick one — if it's a hospital appointment for my kid, do I book that as annual leave or is there something else?"

Nobody reads a handbook. They ask a person, because asking a person takes eleven seconds and reading a 60-page PDF does not. Writing it down was never the bit that failed. Retrieval was.

What closes that gap is an AI knowledge base for HR: the same documents you already have, except a colleague types the question in their own words and gets a straight answer back with a link to the paragraph it came from. Which brings you to the choice you're normally offered, and both options are bad.

Option one is the cheap toy. Somebody in the team starts pasting policy PDFs into a general AI tool and sharing screenshots of the answers. Option two is the enterprise build: rip out your HR system, migrate to a platform with a knowledge module, and book six months of an IT project you don't have the budget or the people for. Most HR and ops leads look at those two and quietly decide to keep answering the questions themselves.

Why both options fail, specifically

Start with the toy. It answers, but nothing about it holds. Nobody else can see the answer. There's no record of which version of the sickness policy it read. It'll answer confidently on a document that was superseded in 2024, and the person asking has no way to check because there's no link back to the source. Worst of all, whoever pasted the handbook in has to be online and available, which is the exact bottleneck you were trying to remove.

It also doesn't stay in one place. Glean's Work AI Index 2026 found 77% of workers juggle multiple AI tools every week and 60% rerun the same prompt across more than one of them, hunting for a version they trust. [1] That's your policy questions getting asked three times, in three tools, with three answers, none of them cited.

The enterprise build fails differently. It fails on the same things that stall most SME technology projects: nobody is certain which use case justifies the spend, the cost is hard to swallow, and there's no one in-house with the expertise to run it. None of those goes away because you signed a bigger contract. A migration means owning a new system of record, retraining everyone on it, and paying per seat for 140 people when maybe 30 of them ask a question in a given month.

The part that should genuinely bother you is what happens after all that effort. Glean asked workers whether the information they need is actually reachable through the AI systems their employer has bought.

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

Glean Work AI Institute

Work AI Index 2026 (survey of 6,000 full-time digital workers)

What an AI knowledge base for HR actually has to do

Strip out the vendor framing and the requirements are short. Five things, and you can hold a supplier to all of them.

It reads the documents where they already live. Your handbook in Google Drive, the process docs in Notion, the PDFs from your pension provider, the benefits pages on your intranet as a URL. No migration, no new system of record, no asking anyone to move house.

It answers in the words the question was asked in. Someone types "can I carry over holiday if I've been on maternity leave" and gets the answer, not four keyword matches on the word "leave".

It cites. Every answer links to the document and section it came from, so the person reading it can check, and so you can spot when it's quoting the old policy.

Everyone gets in without you administering anything. Anyone with an @company.com address joins, full stop. No seat list to maintain, no SSO project, no ticket to IT every time someone starts.

It costs by usage, not by headcount. Charging per employee to answer occasional policy questions is a pricing model designed against the way the tool is actually used.

The ten-minute build, done properly

Before you touch any tool, do the useful bit. Open your inbox and your HR channel and pull out the last 30 questions you were actually asked. Paste them into a plain document. That's your test set, and it's the only honest way to judge whether any of this works, including tools that aren't ours.

Second unglamorous tactic: check what your handbook looks like as text. Word documents full of nested tables and text boxes come through as mush, and a language model reads mush badly. Open the file, and if the policy grid is a table inside a table, ask ChatGPT or Claude to convert it into clean headed sections or Markdown before you import it. Ten minutes of that beats an hour of wondering why the answers are vague.

Then the build. Create a project. Point it at the Drive folder with the handbook and policy docs in it, drop in the pension and benefits PDFs, add the intranet URL. Upload your logo, set your colours. Set access to your domain suffix so anyone with an @company.com address is in. Publish. You now have a link you can paste into the all-staff channel.

Here's the part vendors skip. Ten minutes gets it live. It doesn't get it good. Getting it good takes a couple of afternoons, and those afternoons are spent running your 30 real questions through it and finding out that two documents contradict each other on notice periods, and that the 2021 expenses policy is still sitting in the folder. The assistant doesn't create that problem, it just makes it visible in a way a shared drive never did. Budget for the tidying, not the build.

Where to be careful

Scope it to "what does the policy say". Not "should Priya get the enhanced rate", not anything that shades into a decision about a person. The EU AI Act treats AI in employment as a high-risk area, with obligations for those systems applying from 2 December 2027 following the AI Omnibus, which entered into force on 27 July 2026. [4] A cited assistant that quotes your own published policy to your own staff sits a long way from that line, and it's worth keeping it there deliberately rather than by accident.

Second: import is not a live feed. When you rewrite the sickness policy, re-import it. Put it in the same step of your policy update process as "tell the team".

Third: nobody should be treating the output as gospel, and mostly they don't. Microsoft's 2026 Work Trend Index found 86% of AI users already treat AI output as a starting point rather than a final answer. [3] That instinct is exactly why citations matter more than confidence. An answer with a link to page 14 of the handbook is checkable in five seconds. An answer without one is a rumour with better grammar.

“86% of AI users treat AI output as a starting point, not a final answer; asked which human skills are becoming more important, 50% cite quality control of AI output and 46% cite critical thinking.”

Microsoft WorkLab

2026 Work Trend Index, published 5 May 2026

Measure it the way you'd measure a hire

For the next two weeks, tally the repeat questions. Not a study, just a note in a spreadsheet each time you answer something you've answered before. Holiday carry-over, probation length, expenses limits, who signs off what. Then build the AI knowledge base for HR, share the link, and count the same fortnight again a month later.

If the number doesn't move, the documents are the problem and no software fixes that. If it halves, you've just handed your own week back to yourself for less than the cost of a morning of consultancy, without migrating a single system.

Your next new starter will ask somewhere between ten and twenty questions in their first fortnight. The only real decision is whether they ask you or ask the handbook.

Sources

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

[3] Microsoft WorkLab, 2026 Work Trend Index, published 5 May 2026 — https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization

[4] European Commission, Shaping Europe's digital future — Regulatory framework for AI, 2026 — https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai