Krisp
An AI that answers what it knows - and says when it doesn’t
Autopilot replies from what you teach it, shows the confidence and the sources behind each answer, hands the conversation to a person when unsure, and shows what every reply cost. The built-in engine is free; a hosted model is metered to the cent and capped where you say. Or switch it off, and visitors talk straight to your team.
Grounded, honest, and never in the way of a person
From what you taught it
Your help articles, your website crawled page by page, documents, questions you answered, and the replies your team gave in resolved conversations. The best match is quoted, and the inbox names the source under each reply. The built-in engine quotes an article only when it matches the question well, answers thanks and small talk with a short line of its own, and answers a question about prices from your own page about them or leaves it to your team.
With a confidence it shows
Every reply carries the confidence the engine reported. Below the threshold you set, the AI asks for more detail or brings a teammate in - in your own words - instead of guessing.
Hands over, and stays out
Someone asks for a person, the budget runs out, or an automation escalates: the thread goes to the team, the visitor is told once, and no bot speaks in it again until it is handed back. Once a teammate has replied in a thread, nothing automatic answers it until it is resolved, on every channel. Choose how readily it hands over - or turn the AI off, and every message goes straight to your team (order lookup, if you switch it on, still answers order IDs).
Replies to specific questions
Write the phrases a visitor might use and the answer; it is sent at once, before the AI gets a turn, whole-word matching if you want it. Each is also an automation with its run history.
Answers by email too
From the Starter plan up, a visitor who left their address and leaves your website without seeing the AI’s or your replies gets them in one email, and can answer from their inbox. Emails sent to your workspace address are answered the same way.
Your words, not ours
The greeting, the “tell me more” line, the hand-off notice and the line under a waiting thread are all yours to write, per workspace.
A price tag on every reply
Hosted AI is logged call by call: model, tokens in and out, latency, cost, and the conversation it was spent on. The usage page adds it up by day, model and purpose, with a projection to month end; amounts under a cent are shown in full, never rounded to zero.
- A monthly cap and an hourly cap, with alerts at the percentage you choose
- A hard stop that hands chats to people instead of going silent - or running up a bill
- The built-in engine costs nothing and books nothing
Built-in, hosted, or your own key
The built-in engine answers by rule from your knowledge, free, with nothing leaving Krisp. Switch to the model this server provides, or bring your own account with OpenAI, Anthropic, Google, Mistral, Groq, DeepSeek, OpenRouter or any OpenAI-compatible API - tried before it is saved, booked at your prices, the key never shown again.
- A key that stops working is reported until a call succeeds, and the built-in engine answers meanwhile
- Drafts and rewrites in the composer use the same model, and the same ledger
It gets better with what teams share
Workspaces that choose to share send masked copies of the answers their teams gave - no addresses, phone numbers, links or long numbers - into a pool that trains Krisp’s own classifier. It learns the words of each kind of question, so an unusual phrasing is still understood. Nothing is shared unless you switch it on, and switching it off takes your copies out again.
Let it answer the routine
Feed it your FAQ page, then watch what it answers and what it costs. Free to start.