Your mailbox already knows your FAQ. It also knows every rule that changed since. What to do before an AI assistant reads it.
October 2026
6 min read
You can build an AI HR chatbot from email, but not straight from the mailbox. At Gemeente Medemblik HR judged the answers we mined and about 7 in 10 needed a correction or were outdated. Review every answer, write one document per theme and audience, merge duplicates, and give the assistant a clear fallback.
Yes, you can build an AI HR chatbot from email. Your HR mailbox is the best source of real questions you will ever find. But don't feed it to the bot as it is. When I did this for Gemeente Medemblik, HR reviewed the answers we pulled from years of mail, and about 7 in 10 of the judged answers needed a correction or were outdated. An assistant trained straight on that mailbox would have been confidently wrong most of the time.
So the order matters. Privacy first, then review, then structure, then a fallback. Only then the bot.
Every HR team has a second handbook. It isn't written down anywhere. It lives in the mailbox: can I take a bridge day, how does travel reimbursement work, who signs my employer's statement.
One thing before anything else, because an HR mailbox holds the most sensitive data an organisation has: privacy comes first in the order of steps, not as an afterthought. At Medemblik the raw mail stayed in local storage and never went to a cloud tool. Names, email addresses, phone numbers, BSN numbers and IBANs were removed before any text reached an AI model. The analysis ran through the commercial Claude API, where customer data is not used to train models, and the final knowledge base holds no personal data at all.
With that in place, I connected to the HR mailboxes through Microsoft Graph and pulled roughly 245,000 messages. Claude went through the anonymised threads looking for questions that keep coming back, with the answer HR gave. That produced 1,271 candidate question-and-answer pairs.
That sounds like a finished HR knowledge base. It isn't. A mail from three years ago is still in the mailbox, with the same confidence as a mail from last week. AI is good at finding patterns. It has no idea which rule changed last year.
So every answer went into a shared review sheet, and seven people from HR marked each one:
That is the number in the title. Not because HR did a bad job back then, but because rules, terms and limits move, and nobody goes back to fix old mail.
One more thing that saved a lot of work: we looked at how often each question was actually asked. 884 of the 1,180 questions had come up only once. That long tail goes to a central HR address instead of the bot. It cut the remaining review work from 633 rows to about 135.
This is the part most people skip, and it decides whether your bot is any good.
Medemblik uses Visma Youforce, and the Youforce assistant is the HR chatbot inside the app employees already open. So the question was never "should we build a chatbot". The question was what to put in it, and in what shape.
Here is how it searches, which I learned from the vendor and then tested myself:
The surprise came in testing. I set up a separate test assistant in the Youforce sandbox, with debug on and a small parser that showed which fragments were retrieved for each question. Our short themed documents were pushing the CAO and the staff handbook out of the top 3. The fix was two contexts: one for our reviewed documents, one for the CAO and handbook. Now the assistant pulls 3 plus 3, so 6 fragments instead of 3. Ask about bridge days and you now also get the salary rule from the handbook.
Other assistants are built differently, but the lesson travels. Whether it is Copilot, a custom GPT or a helpdesk bot, it does not read your whole library. It grabs a handful of pieces. Your job is to make sure the right pieces win.
A good answer for HR is not always a good answer for everyone. Some content is for managers only. Some is for administrators.
In Youforce, access is set per file. So a theme gets split when the audience differs. At Medemblik a script builds one Word document per theme plus audience. 737 reviewed answers ended up in 127 uploaded documents:
Three more were held back because their audience (council members, aldermen) couldn't be mapped to an access group yet. I'd rather leave something out than show it to the wrong people.
And the documents are plain, editable Word files. HR maintains them without me: edit, upload, done. If your knowledge base needs a developer for every change, it will be out of date within a year. Then you're back to the mailbox problem.
The bot will not know everything. It shouldn't pretend to.
At Medemblik the assistant lists the topics it covers and shows the central HR address as a clickable link when it doesn't know. That is where the long tail of one-off questions goes too. A clear "ask HR here" beats a confident guess, every time.
Before anything goes to all staff, there's a test with colleagues: 15 hard questions, including checks that HR-only content never reaches other employees. Right now the content is live in the Youforce environment, and staff rollout is in testing. I'll say that plainly, because I have no measured time saved or accuracy numbers yet.
The pattern isn't specific to a municipality or to Youforce. Any team that answers the same questions by email can do this.
Where the questions come from:
Where the answers can live:
The steps stay the same: anonymize, extract with AI, deduplicate, let people review, write one document per theme and audience, set access, add a fallback, test with real questions. You can see the full build in the Medemblik case, and more builds like it on the work page.
The hard part is never the bot. It's deciding what goes in, having people check it, and writing it in the shape the assistant can find. If you're sitting on years of HR or support mail and wondering what's in there, let's talk for 30 minutes.
Yes, and your mailbox is a great source of real questions. But don't train or feed the bot on raw mail. At Gemeente Medemblik about 7 in 10 of the answers HR judged needed a correction or were outdated. Extract the questions with AI, then have people review every answer before it goes live.
It shouldn't be, and at Medemblik it wasn't. Use a commercial AI API where customer data is not used for training, never paste mail into consumer chat tools, and strip personal data before any analysis.
Write one document per theme and per audience, with a clear file name, because assistants often use the file name as the topic. Merge duplicates, since each one takes a retrieval slot. Keep long sources like a CAO or handbook in a separate context so they don't get pushed out.
Yes. The Youforce assistant searches documents you upload, with access set per document. At Medemblik that meant 127 Word documents split over four audiences, plus the CAO and staff handbook in their own context.
Only in the right order. Keep raw mail local, remove personal data before analysis, let HR review every answer, and set access per audience so HR-only content stays with HR. Test that last part explicitly before rollout.
Send people to a human. A clickable link to a central HR address, plus a list of topics the bot covers, works better than a guess.