An internal chatbot is an assistant your team asks what they used to ask the coworker at the next desk, and it answers with the company's own documents. You set one up in three moves: gather what's already written, give it to the chatbot to read, and test it with real questions. And you judge it less by what it knows than by how it behaves when the document is old or missing.
The example in this guide is a made-up, twenty-person HVAC parts distributor.
Ten days in your first year, accrued monthly. From your second year on, fifteen. That's from this year's employee handbook.
An internal chatbot is a chat only your team can get into. It answers with the company's documents: procedures, price lists, terms, policies. The one on your website is something else: it deals with customers, anyone can see it, and it should only say what's public.
The headlines about internal AI chatbots tend to come from very large companies. You don't need that scale. A twenty-person company has the same problem in miniature: the same HR and IT questions, asked again every week, answered by whoever happens to be at their desk.
The word is a bit of a trap. Strictly speaking, a chatbot follows a script written ahead of time. What people look for today as an “internal AI chatbot” is closer to an internal assistant: it understands the question even if nobody planned for it, and it writes the answer from what it finds. Here we use the everyday word.
Fewer than you think. An internal chatbot starts well with what gets asked the most and what does the most damage when it's answered wrong:
That set is called a knowledge base. Short and current beats long with three versions of everything.
If the documents exist, the effort is days, not months. If they don't, the job is writing them, and no tool fixes that. What to upload and how to prepare it is in the guide to training AI on your own data.
If there are four of you at the same table, asking out loud is faster. If almost nothing is written down and nobody is going to write it, it doesn't pay off either. And if what you want is to handle customers from your website, you need a different kind of tool; our comparison with CustomGPT looks at one.
Someone in the distributor's warehouse asks about returning pallets. Change the state of the document and see what a good internal chatbot should answer in each case.
What's our pallet return policy?
Pallets go back to the warehouse within fifteen days of delivery. After that, the deposit is charged.
1 source consultedWarehouse rules 2026.pdf
It knows. It answers and shows where the answer comes from, so you can open the document and check.
I can only find a fact from 2023: back then the window was thirty days. I can't confirm it still applies. I've noted it so someone can fill it in.
Old factWarehouse memo 2023.pdf
It hesitates. It gives the fact it has, warns that it's old, and doesn't present it as current.
I don't have that. There's nothing in the documents about returning pallets. I'm logging it so someone can fill it in.
I was asked and didn't knowPallet returns · 1 time
It says “I don't know.” It doesn't invent a deadline, and the gap is logged where the person who can close it will see it.
This is a mockup with sample data. All three behaviors are Flipa's; we have the old-fact one recorded in a real case, but we don't promise it will always detect an old document.
Flipa is an internal assistant: it has no chat window for your website and doesn't talk to your customers. Here's how it answers. Someone asks how much the workshops cost, it replies with the prices, and opening the sources shows the document they come from. The animation recreates the real screen and the answer Flipa gave in a test account. The app's interface is in Spanish today, and English isn't available yet; the assistant answers in the language you write in.
Six tests you can run in ten minutes, with your own documents. The last column says where Flipa stands on each.
| What to ask for | How to check it | In Flipa, today |
|---|---|---|
| It shows the source | Ask something that's in a document and see whether you can open it. | Yes: “sources consulted” under each answer. It doesn't point to the exact sentence inside the document. |
| It admits what it doesn't know | Ask something about the company that isn't written anywhere. | It says so plainly. It's instructed not to fill the gap with general knowledge. |
| It logs the gaps | See whether that question gets recorded somewhere. | Yes: the topic lands under “I was asked and didn't know,” with how many times it came up. |
| It tells roles apart | Sign in as a regular team member and try to delete a document. | Whoever administers the account uploads and deletes; everyone else asks. Everyone consults the same knowledge: you can't reserve a document for a few people. |
| It can be corrected without a technician | Tell it a fact has changed. | It proposes the change with the before and after, and saves it when you approve. That works for notes and brand documents in text, not for a PDF. |
| It keeps your company apart from the others | Ask the vendor how they verify it. | Each company has its own separate space, with dedicated tests that try to get around it. It's on the security page. |
They ask what they'd be embarrassed to ask a third time. More on this in Flipa for HR.
How do I submit an expense report?
They pick up the phone with the fact in front of them, knowing which document it comes from.
How long does a customer have to report a part that arrived damaged?
They put a quote together without waiting for the office to confirm the terms.
What's the minimum order for a new contractor account?
They stop being the switchboard for every IT and paperwork question.
How do I get VPN access for someone working from home?
A chat only the team can get into, which answers with the company's documents. It's there so the usual questions (deadlines, terms, procedures) don't depend on the one person who knows them being in the office.
Gather what you already have in writing, upload it to a tool that looks it up before answering, and test it with last month's real questions. Then open it to three people and, once it works, to everyone else. No coding needed.
The internal one is for the team: it sits behind a login and can use documents that aren't public. The one on the website talks to customers and should only say what anyone can read. They're different tools, with different risks.
With the documents already written, an afternoon to gather and upload, and a week or two of testing with real questions. What stretches the timeline is almost always finding out that something important was never written down.
A named person, not “the team.” Usually it's whoever already answers the most questions: operations, the office manager or the owner. In Flipa, uploading and deleting documents is up to whoever administers the account; everyone else asks.
With caveats you should know up front. The app's interface is in Spanish for now: English isn't available yet. The assistant answers in the language you write in, so questions and answers can be in English.
Start with that one. If it's already written in some document, your internal chatbot is half built.
Start free