AI for the shared inbox: triage, drafting and answering for a UK small firm, with the costs.

In most small firms the shared inbox, info@, sales@, bookings@, is read by two or three people who each open everything, decide who owns it, and answer the same twenty questions in slightly different words. The measurable costs are time to first response, which decides whether an enquiry becomes a customer, and the hours of reading and retyping between the inbox and the system where the answer lives. Both are easy to baseline for two weeks, which is why the inbox is often the best first process for a measured 90-day return.

Every message read on arrival, classified (enquiry, order, complaint, supplier, spam), prioritised by urgency and value, tagged with what the sender is asking for, and routed to the right person or queue, with a one-line summary. Nothing is sent. It needs a service account that can read the mailbox and apply labels or move messages, and a definition of the categories the firm actually uses. It is the safest automation there is, and the return is visible in the first week: the urgent enquiry is answered first, and nobody reads the newsletter three times.

For each message that needs a reply, a draft prepared from the thread, the firm’s standard answers and, where connected, the relevant record in your system: the order, the booking, the account. The person reads, edits and sends. The quality comes from grounding: an assistant that reads your own answers and your own data drafts well; one that writes from general knowledge drafts plausibly and wrongly. The controls are the checking habits, sources shown, and a rule that the person is responsible for what they send. This is the tier most firms should stop at for anything involving money, complaints or individuals.

Routine enquiries answered without a person: where is my order, what are your hours, can I book Thursday, what does this cost. This is an agent: it reads, decides, acts in your systems and replies, and it needs the agent controls. Every incoming email is untrusted input, because a message can contain instructions aimed at the assistant, which the NCSC has warned cannot be reliably prevented; so the system has the least access that does the job, cannot send money, change access or delete, escalates whenever it is unsure, says it is an AI, and logs everything. Done that way it answers most routine mail within a minute at any hour and hands the rest to a person with the context attached.

Triage and drafting together are a first implementation at £4,000 to £15,000 fixed at UK rates, integrated with the mailbox and the systems the answers live in, with the running cost stated before go-live; reading email is among the cheapest things a model does, a few thousand tokens per message on a small model, so the running cost is dominated by any messaging or telephony the firm adds, not by the model. Unattended answering is a custom system at £8,000 to £25,000, because the controls and the integrations are the work. Off-the-shelf help-desk tools cover parts of tier one and two for a subscription and are worth checking first; our guide on buying against building sets out how.

Prices from our published ladder. Running cost stated per feature before go-live.

Two weeks of baseline: messages per day, time to first response, hours spent by everyone who reads the inbox, and how many replies needed a second attempt. Two weeks of tier one and two live. Compare. Most firms see time to first response fall from hours to minutes for the urgent category and the reading hours halve, which is the number that pays for the build. Then decide whether tier three is worth its controls, category by category.