Assistants and agents
Assistants that know your business.
Assistants that answer from your own documents and records, act only with the user’s permissions, and say so when they don’t know. Agents that draft the work and wait for a person to approve it.
01What it is
A general chatbot knows a lot about the world and nothing about your business. A useful assistant is the opposite. It answers from your documents, your standards and your records, cites where the answer came from, and declines when the answer isn’t there, because a confident wrong answer is worse than no answer.
Agents go a step further and do things: draft the day’s paperwork, prepare an entry, take a booking. We build them on the principle the rest of our work follows. The assistant does the typing and a person makes the decision. It can only see and do what the signed-in user already could.
02Where it pays
- Technical questions answered from standards and internal documents, with the clause cited
- Questions about hours, costs or approvals asked in plain English
- Routine paperwork drafted for one-tap approval
- Product and installation questions answered for the exact product
- Enquiries and bookings handled after hours by voice and chat
- Marking and feedback to an expert’s own standard
03How we build it
Grounded
Answers come from your sources, with the source cited, and the assistant declines when they don’t cover the question.
Permissioned
It works with the signed-in user’s own access, so it can never see or do more than they could.
Draft first
Nothing is saved until a person approves it.
Evaluated
It’s tested against expert answers before release, so you know how often it’s right before anyone relies on it.
Systems we’ve built
Questions owners ask
Will it make things up?
Any language model can, which is why we design against it. Answers are grounded in your own documents with the source cited, the assistant declines when the documents don’t cover a question, and it’s evaluated against expert answers before release. Our standards assistant agreed with senior engineers on 92% of a fifty-question evaluation set.
Can it see information a user should not?
No. It works with the signed-in user’s own access, so it can only see and do what that person already could.
Does it act on its own?
Only as far as you decide. By default it drafts and a person approves. Low-risk work, such as taking a booking through to confirmation, can be left to it.
Is our data safe with an AI assistant?
The assistant runs in accounts you own, works with each user’s own access, and reads only the sources you choose to give it. Before anything is built we go through where the data goes, including which model provider processes it.
What happens when it can’t answer?
It says so, rather than guessing. Where it’s set up to, it passes the question to a person with the conversation attached.
Which AI models do you use?
Whichever suits the job. Some of our systems use Claude, from Anthropic. The model matters less than the grounding, the permissions and the evaluation around it.
Can it answer customers after hours?
Yes. Voice and chat assistants can answer enquiries outside business hours and take a booking from first contact through to confirmation, in the business’s own tone.
Got work your team shouldn’t be doing by hand?
Tell us how the business runs and where the time goes. We’ll tell you honestly what’s worth building, and what isn’t.