AI chatbot for employees: a checklist before you launch

1. What it should handle

Start with repeat questions that have one correct answer: time off, expenses, tools and access, where a document lives, who to ask. Leave decisions, pay, hiring and anything legal with a person. If a question has no single approved answer, the bot should say so, not choose one.

2. Where its answers come from

• Only from answers a named person approved, not from old chat threads or the model's general knowledge. Anthropic's guide on reducing hallucinations lists restricting the model to provided documents as a technique.
• Every answer shows its source and the date it was last confirmed.
• Old and conflicting documents are removed. Helpdocs notes that two conflicting answers leave both people and an AI assistant unable to tell which is right.

3. What it does when it does not know

Ask the vendor, and then test it yourself, what happens when no approved answer exists. The right behavior is a fixed, plain line such as: I do not have an approved answer to that, and I have asked the owner. The question should then go to a named person, and the answer, once written, should be available next time.

4. Questions to ask a vendor

• Does it answer only from my documents, or can it use general knowledge?
• Can it show the source and date for each answer?
• What does it do when it has no answer? Can I see an example?
• Who can see which answers, and can I limit an answer to one team?
• Where is my data stored, who can read it, and how do I export or delete it?
• How do I see which answers it gave, and flag a wrong one?

5. A one week pilot with five people

• Day 1: load ten approved answers and tell five colleagues what the bot is for and what it is not for.
• Days 2 to 4: they ask real questions. Everything the bot cannot answer goes to one person.
• Day 5: read every answer it gave next to the source. Mark each right, wrong or unsupported.
• Day 5: write the missing answers, fix the wrong ones, and decide whether to widen the pilot.

6. Test it before the pilot

Ask it twelve questions that no document answers, such as the Wi-Fi password, last quarter's revenue or what a colleague earns, three tries each. A bot that makes things up here will do it in front of your team. You can run this test on any bot with the free honesty test at askmystartup.com/honesty-test.

7. Who owns a wrong answer

You do. In 2024 a Canadian tribunal rejected Air Canada's argument that its chatbot was a separate entity responsible for its own actions, found the airline had not taken reasonable care to make the chatbot accurate, and ordered it to pay. Decide before launch who reads the flags, who fixes the source, and who tells the person who got the wrong answer.

Our bot, as one example

AskMyStartup is built around this checklist: the founder approves each answer, the clone answers only from approved ones, shows the source and date, and says it does not know otherwise. We ran our honesty test once on a made-up bakery workspace: it said it did not know in 36 of 36 tries. It is in private beta.

Related guides

• How to test an AI chatbot before your team relies on it
• How to reduce AI chatbot hallucinations in a team bot
• If your team bot gives a wrong answer, your company owns it

Sources

• Anthropic: Reduce hallucinations (Claude documentation): platform.claude.com
• Helpdocs: How to create a knowledge base for employees: helpdocs.io
• Manatt: AI Gone Wild, Airline Has to Honor a Refund Policy Its Chatbot Fabricated: manatt.com

Each page was opened and read on October 20, 2026. Our own numbers come from the honesty test page.