Assistants that answer from your own material
The useful ones answer from what you actually know, admit when they do not, and pass the question to a person before the customer gets angry.
Your team answers the same forty questions every week. The answers exist, in documents nobody outside the building can search.
What changes
What you can expect it to do.
Answer from your material
Your documentation, your policies, your product data. Grounded in sources it can cite rather than in what the model half remembers.
Hand over cleanly
When it does not know, it says so and passes the conversation to a person with the history attached. Confident nonsense is worse than no assistant.
Tell you what it could not answer
The log of unanswered questions is the most valuable thing this produces. It is a list of what your customers need and you have not written down.
What is included
What building one actually involves.
- A defined scope: the questions it is allowed to answer, and the ones it must not
- Retrieval over your own content, so answers are grounded and traceable
- A handover path to a person, with the conversation carried across
- An evaluation set written before launch, so a change can be proved better rather than felt better
- Cost per conversation measured and capped, because this is the line that surprises people
- A log of what it refused or failed, reviewed monthly with you
The difference
A demo, and something you can leave running.
Almost anyone can produce the first column in an afternoon. The second is the work, and it is where the value is.
A demo
- Answers anything, confidently, including things that are not true
- Tested by the person who built it asking questions they know it handles
- No path to a human, so a stuck customer is a lost one
- Cost per conversation unknown until the first invoice
- Changes shipped on the basis that they feel better
Something you can leave running
- Answers within a defined scope and says so when a question falls outside it
- Tested against a written set of real questions, including the awkward ones
- Hands over to a person with the history attached
- Cost measured per conversation and capped before it runs anywhere near customers
- Changes measured against the evaluation set before they ship
The most valuable output is not the answers. It is the list of questions it could not answer, because that is your customers telling you what you have never written down.
What we look at first in the monthly review
Questions people ask
The ones that actually arrive.
Will it say something wrong and embarrass us?
It can, which is why scope, grounding and the handover path are built before anything faces a customer. We also start it somewhere low risk, usually internally, so the first month of mistakes happens in front of your team.
What does it cost to run?
Per conversation, and it depends on how much of your material each answer needs. We measure it during the build and give you a figure with a cap, because an unbounded bill is the most common unpleasant surprise with this work.
Does our data go into training a model?
No. We use the providers' business terms, where content is not used for training, and we tell you exactly what leaves your systems and where it goes.
Can it replace our support team?
No, and we would not build it on that promise. It removes the repetitive half so your people can spend their time on the half that actually needs a person.
Read next
Your turn
Bring one question you cannot currently answer.
30 minutes. We will tell you whether it is measurable, roughly what it would take, and whether we are the right people for it.