What it is
Most businesses do not have an AI problem. They have an inbox: the same twenty questions, a quote that has to be written by hand, an enquiry that waits until somebody notices it. AI is genuinely useful there: it reads, sorts, looks up and drafts, and a person still decides.
That is the shape we build: one channel, one or two kinds of enquiry - an inbox, a contact form, or the phone if missed calls are the thing costing you work - with your own documents as its memory and a person approving anything that commits you to something. It is unglamorous, it is the version that survives contact with a real Tuesday, and it is the only kind worth building.
What you get
- Enquiries that arrive sorted. Classified by what they are about and how urgent they are, routed to the right person, and acknowledged the same hour rather than the next day.
- Answers grounded in your own material. Prices, policies, procedures and opening hours read out of the documents you already maintain, with the passage they came from shown beside the draft, so whoever sends it can check it at a glance.
- Drafts, not decisions. A reply written to be edited rather than sent. Prices, refunds, delivery dates and anything about a customer's rights stay yours, held behind your rules and a person's approval.
- The dull half of the work, done. An enquiry turned into a ticket, a history summarised, a field filled in, a follow-up created - the jobs nobody counts and everybody loses an afternoon to.
- A number you can defend. Resolved enquiries, confirmed by the customer rather than assumed, set against repeat contact, complaints and the minutes each one took. Deflection is not a result; it is a word for a conversation that stopped.
Where the law lands
Handling customer contact with AI means processing personal data from the first message, so it is ordinary data-protection work with a new supplier in the middle: a lawful basis recorded for each purpose, privacy information that mentions the assistant, a data-protection assessment where the risk asks for one, a contract that says what your provider may keep and train on, and a retention period for transcripts that somebody chose rather than inherited. Consumer law does not change because a machine is answering - you are still responsible for what it says, so it cannot misstate a refund right or stand between a customer and a person. And since August 2026 a conversational AI handling customers in the EU has to say what it is at the start. We do that work with you, and you keep the record of it.
How it starts
Two to four weeks, one inbox, one or two kinds of enquiry. First a baseline, taken from your own records: how many arrive, how long they wait, how long each one takes a person. Then the system itself, seeded with twenty to fifty real questions and answers from your archive and run in the background while the team works as it does now, so its drafts can be compared with the replies you actually send. You get a written scope, a price and a date before anything starts, and a review at the end that says whether it paid for itself - including when the honest answer is that it did not.