How to build an internal assistant that answers staff questions

TLDR: The HR team spends hours answering the same questions: how many days do I have left, how do I request leave, when does the bonus land. The internal assistant answers instantly by reading your own documentation: Tidio if you want the assembled conversational product, FastGPT if you prefer the cheap self-hosted knowledge base. The condition is the usual one: concrete documentation to read, and escalation to a person for everything delicate.

This guide is for people teams acting as a service window: the same question twenty times a month, each one interrupting another task. Also for companies without formal HR where those questions land on admin or the owner.

The rule presiding over the system: the assistant answers the repeated and documented, and people answer the delicate. The payslip that does not add up, the conflict and the personal situation always escalate to a human.

1. Inventory the questions HR answers on loop

Two weeks of noting (or one afternoon reviewing email) produce the real list, which usually concentrates in four families, plus the fifth the assistant never answers:

An assistant trying to resolve a payroll dispute does more harm than service.

That inventory is also an uncomfortable mirror: most questions exist because the information is not where staff look, and that is half the fix.

2. Write the documentation the assistant will read

The assistant answers with what it finds written, and vagueness is inherited: if your leave policy says “with reasonable notice”, the assistant will repeat that fog to everyone. The rewriting pass turns each policy into concrete answers: exact days, deadlines with numbers, steps with the button’s name, and the link to where it is done.

Claude speeds that rewriting by turning your notes into clear answers, from $17 a month and with a permanent free tier, with your final review because these are company commitments. What the task needs is precision when rewriting a policy, not a choice of models: the ambiguous answer gets multiplied across the whole workforce.

The format that works best: one page per question family, short answer first and detail after. That structure serves the assistant and the person who prefers to read equally.

3. Build the assistant with Tidio or FastGPT

With documentation ready, two routes by profile:

Tidio has a free tier and applies the same mechanics as customer service, here pointed inward. FastGPT also has a permanent free tier, bills in its own currency, and covers the case’s core for companies with someone technical and a tight budget.

On both, the configuration that matters is escalation: the inventory’s delicate families go straight to a person, and the assistant says so plainly (“the people team handles this, I have just notified them”).

4. Close the cycle with the unanswered questions

The system only improves if someone looks at what failed: the weekly review of questions the assistant could not answer (fifteen minutes) produces the missing-documentation list, and every gap closed lowers next month’s queries.

The three-month accumulated effect is the real goal: the HR window stops being constant interruption, staff get instant answers at any hour (the night shift exists too), and the written policies improve for everyone, including those who never ask and suffer silently.

The system connects with onboarding (newcomers’ questions are its first customer) and with internal communications, which feeds the documentation of changes. The whole sector lives in AI for human resources.

Common mistakes

Building the assistant on vague policies. It multiplies the fog at scale. First the policies get rewritten with figures, then it gets connected.

Letting it answer the delicate. A personal situation or conflict answered by a machine creates the grievance HR exists to prevent. Direct escalation is not a system failure: it is the system.

Not telling staff what it is. People must know they are talking to an assistant and can request a person. Transparency here is internal trust.

Abandoning the improvement cycle. Without the weekly failure review, the assistant freezes while policies change, and answers confidently things that are no longer true, which is worse than not answering.

Frequently asked questions

How much does it actually save?

The typical math: if HR spends an hour daily on repeat questions, the system frees most of that hour from month one. The invisible improvement is bigger: instant answers remove the waits nobody currently measures.

Tidio or FastGPT?

Assembled product versus own base: Tidio gives the chat, escalation and management without touching code, from $24.17. FastGPT gives the core from 99 yuan, billed in its own currency, with technical hands alongside. The rewritten documentation serves both, so you can start cheap and migrate.

Which questions should the assistant never answer?

Personal situations, conflict, health, and any “my specific case” on payroll or contracts. The practical rule: if the correct answer depends on who is asking, escalate to a person.

Will staff actually use it?

If it answers well and fast, yes: the alternative is waiting for HR to answer the email. The adoption keys: announce it well, make the first answers good (hence documentation first) and keep requesting a person easy and frictionless.