Answers from your actual documents, not invented ones.
The case against putting AI in front of customers is a good one: it will confidently make something up, and the screenshot will end up somewhere you don't want it. That risk is real and it is the reason most support automation is either useless or reckless.
This can only repeat what you have already written down. It searches your documents, your past tickets and your policies, and answers with the source shown. When the answer isn't in your material it says so and fetches a person, rather than assembling something that sounds right.
How a question gets answered only when the answer really exists.
A question is read and searched against your own documents, past tickets and policies. If the answer is genuinely in your material, it is given with the source shown. If it is not, the system says so and fetches a person rather than inventing something.
Good fit, and not.
The second list is the useful one. If you recognise yourself in it, say so on the call and we’ll tell you that rather than sell you something.
Good fit if
- Your answers exist in writing somewhere, even if that somewhere is messy
- A meaningful share of tickets are questions you've answered many times before
- You'd rather it escalate too often than guess once
Not a fit if
- Your answers live in one person's head. Fix that first — it's worth doing anyway
- Every case is genuinely different and needs judgement
- You need it to make commitments on your behalf. It won't, by design
What a build usually looks like.
Descriptions of what we build, not case studies. We have no clients yet, and these are illustrations rather than work delivered for anyone.
A first-line build
Answers the repeat questions with the source attached and passes everything else straight to a person.
A triage build
Doesn't answer at all. Reads, categorises, sets priority, and routes to the right queue with a summary.
A drafting build
Writes the reply and leaves it for an agent to send, which is often the right place to start.
For whoever asks the hard question.
Answers are generated only from retrieved passages of your own material, and the passage is shown alongside the answer so anyone can check it in a second. If retrieval returns nothing above the confidence you set, the model is not asked to answer at all — it isn't asked to be careful, it simply isn't given the opportunity. Every escalation is logged with the question, which over time is a rather good map of the gaps in your documentation.
- Grounded retrieval
- It can only answer from passages found in your own material.
- Source shown
- The passage behind the answer is displayed, so it can be checked.
- Confidence threshold
- Below the line you set, it doesn't answer — it fetches someone.
Thirty minutes to find out whether this is worth building.
Usually three to five weeks. You’ll get a written scope and a fixed price before anything is built, and if it isn’t a fit we’ll say so on the call.
Explore related work