A team bot has a different job from a general chatbot. When a colleague asks it about the refund rule, the leave policy or the discount limit, the answer turns into an action. A wrong answer that sounds sure of itself costs real money, and it costs trust too, because people stop asking the bot and go back to asking you.
This guide is for the person who owns that bot: a founder, an operations lead or a manager. It lists seven fixes in the order that matters most, says what each one can and cannot do, and ends with what we saw when we tested our own.
Why a team bot makes things up
In September 2025, researchers at OpenAI and Georgia Tech argued that language models produce confident false statements partly because of how they are trained and scored. Most scoreboards reward a correct guess and give nothing for an honest 'I don't know', so a model that always guesses looks better than one that admits doubt. The default behavior is to produce an answer.
Inside a company that default is the problem, and it has already cost organizations real money. In 2024 a Canadian tribunal ruled against Air Canada after its website chatbot told a grieving customer he could apply for a bereavement fare after buying a full-price ticket. The airline argued the chatbot was a separate entity responsible for its own actions. The tribunal rejected that, found the airline had not taken reasonable care to make the chatbot accurate, and ordered a partial refund of 650.88 Canadian dollars plus interest and fees.
A public example from the same year: The Markup tested New York City's business chatbot and found it said things that were wrong, such as that bosses can take workers' tips. It was also inconsistent. It told one reporter that landlords must accept housing vouchers and told ten other staffers they did not have to.
Fix 1: Answer only from approved sources
Tell the bot to use only the documents you give it and to ignore what it learned in general training. Anthropic's guide on reducing hallucinations lists this as one of its techniques, called external knowledge restriction. For a team bot, go one step further: the documents should be answers a person approved, not a pile of old chat threads. A chat thread is not an official answer.
Fix 2: Give it permission to say 'I don't know', and write the line
Anthropic's guide puts permission to admit uncertainty first among the basic techniques and says that simple step can sharply reduce false information. Do not leave the wording to chance. Write the sentence you want people to see, for example: I do not have an approved answer to that, and I have asked the founder. A fixed line is easy to recognize, easy to test for, and easy to count.
Fix 3: Make every claim carry a quote or a source
The same guide recommends making answers auditable: ask for a supporting quote for each claim, and if none can be found, retract the claim. In a team bot this shows up as a source and a date under every answer. If you cannot see where an answer came from, you cannot tell a good answer from a lucky guess.
Fix 4: Date and own every source
Every approved answer needs a date it was last confirmed and a person who owns it. Without both, the bot will repeat an old rule with full confidence. When the owner changes the rule, the old answer should be replaced, not left beside the new one.
Fix 5: Send unknowns to a named person, and write the answer once
A refusal with no next step is a dead end, and people will route around it. Each question the bot cannot answer should land with one named person. When they answer it, it becomes an approved answer, and the same question is covered next time. Over a few weeks the list of unknowns becomes your to-do list.
Fix 6: Test with questions that have no answer, then test again
Write a dozen questions the bot should not be able to answer, such as the office Wi-Fi password, last quarter's revenue or what a colleague earns. Add a few traps that ask it to guess, to ignore its rules, or to treat a claimed approval as real. Ask each one in a fresh chat. Then ask each one again, because the same model can answer differently on a second try. Anthropic's guide suggests running the same prompt several times and comparing the outputs, since inconsistency can signal a hallucination.
Fix 7: Read the flagged answers every week
Give your team a one-tap way to say 'this looks wrong', and send every flag to one person. Fix the source first, then the bot. A weekly ten-minute read of the flags and the unknowns is the cheapest quality control you have.
What these fixes do not do
Anthropic is plain about this in its own guide: these techniques significantly reduce hallucinations but do not eliminate them, so critical information should still be validated. For anything that moves money, signs a promise or affects a person's job, keep a human in the loop. A bot that answers from approved notes and says when it does not know is a strong helper, not a decision maker.
What we saw when we tested ours
For a test of our own product, we set up a made-up company, a small bakery with 8 approved notes about its hours, deliveries and baking, none of which answers these twelve questions. We asked our clone the twelve questions as a teammate, three tries each (36 tries), through the same code a real teammate's question goes through. In that one run it said it did not know in 36 of 36 tries and made nothing up. It is one run on a made-up workspace, so your result with your bot will differ. You can run the same twelve questions on your own bot in a few minutes with the free honesty test at askmystartup.com/honesty-test, and read all 36 of our replies on that page.
Related guides
Common questions
No. Techniques such as approved sources, quotes and a fixed 'I don't know' line reduce false answers a lot, but vendors that publish guidance on this say they do not remove them. Keep a person responsible for anything that costs money or commits the company.
Restrict the bot to a short list of approved answers, write the exact 'I don't know' sentence you want, and test it with a dozen questions that have no answer. Those three steps catch most of the damage.
A language model can answer the same question differently each time. Our own earlier runs moved by a try or two between runs. Asking three times shows whether a good result is stable.
One named person per answer, with the date it was last confirmed. If the owner is unclear, the answer will go out of date and the bot will keep repeating it.
Sources
- Anthropic: Reduce hallucinations (Claude documentation): platform.claude.com
- Kalai, Nachum, Vempala and Zhang: Why Language Models Hallucinate (arXiv, September 2025): arxiv.org
- Manatt: AI Gone Wild, Airline Has to Honor a Refund Policy Its Chatbot Fabricated: manatt.com
- The Markup: NYC's AI Chatbot Tells Businesses to Break the Law (March 29, 2024): themarkup.org
Each page was opened and read on October 20, 2026. Our own numbers come from the honesty test page.