An assistant is only useful once someone has decided what it may answer, which information it is allowed to see, and what happens when it reaches the edge of that. We write those rules with you first, then build to them, then test them before anything goes live.
The assistant your customers talk to, the AI steps your own team uses internally, and the written material that keeps both running after we hand over.
An assistant on your website or on WhatsApp that answers common questions, captures enquiry details and passes anything outside its remit to a named person. We agree the opening message, the tone, the questions it is allowed to ask and the point at which it stops and hands over.
AI steps inside work your team already does: summarising long threads, drafting a first reply, sorting incoming requests, or retrieving the right paragraph from an approved document. A person reviews the output before it reaches a customer, unless you decide otherwise in writing.
A written library of the prompts your team actually uses, each with the input it expects and the output it should produce, plus short procedures for the steps around them. It lives in your own account so your team can edit it without waiting on us.
Most of the risk in an AI system sits in what it is allowed to say and what it is allowed to see. We settle both on paper before any build work starts.
We list the questions the assistant should handle, the ones it must decline, and the ones it should route to a person. Escalation triggers are explicit: an unhappy customer, a pricing or contractual question, a regulated topic, or simply a question it cannot answer with confidence.
Retrieval runs over approved company knowledge only: the documents, pages and records you have signed off. We agree what is deliberately excluded, who is allowed to see what, how long conversations are retained and where the logs are stored.
We build a test set from your real questions, including the awkward ones and the ones that fall outside scope, and run it before launch. You read the transcripts and sign off on the answers rather than taking the behaviour on trust.
Handover is part of the build, not a favour arranged afterwards.
Everything sits in your accounts: the model provider, the automation platform, the knowledge base and the conversation logs. We document how the assistant is configured, where each answer comes from, and how to change one without breaking the rest. Your team gets a working session on updating the knowledge base, adjusting the escalation rules and reading the logs, so a wrong answer is something you can correct the same day.
Typical delivery for a first assistant runs in weeks rather than months, depending on how much of your knowledge is already written down. If it is not written down yet, that clean-up is usually the honest first step, and we will say so during scoping rather than three weeks into a build.
The assistant creates work. Automation moves it: enquiries into your CRM, alerts to the right person, follow-ups on a schedule.
The site or app the assistant sits on, built so the page and the conversation tell a customer the same thing.
A live view of what the assistant is being asked, what it handles and what it escalates, alongside the rest of your numbers.
Some of what looks like an AI problem is a process problem or a missing document. We would rather establish that in the first conversation than after the build has started.
Start a project enquiry