How do you use AI with your email?

Email is the highest-volume text in most small businesses and the place where AI help is most immediately felt. It is also the system with the least forgiving error mode: everything else can be corrected before anyone sees it, and a sent message cannot. That asymmetry should shape which parts of the job get delegated, and it points at the same conclusion in most businesses — automate everything up to the send, and treat the send as its own decision.

Triage is the first and least risky use. Classifying what arrived into categories that matter to you — new enquiry, existing customer, supplier, administrative, ignore — and surfacing the ones that need attention today. The error cost is low because a misfiled message is recoverable, the volume makes the saving real, and the classification is checkable by looking at the folder. This is also the use that reveals what your inbox actually contains, which is frequently a surprise.

Extraction is the second and is underrated. Enquiries arrive as prose and businesses need fields: what they want, where, when, what budget was mentioned, what deadline. Pulling those into a row means the enquiry is countable, sortable and answerable without reading the thread again. The check is trivial, because the extracted fields sit next to the original text, and this single step is what makes it possible to answer questions about response times and enquiry volume later.

Drafting is the third and is where judgement enters. A drafted reply to a routine enquiry saves the blank page and the composition time, which is most of the effort on a standard message. The failure mode is not a badly written draft, which is obvious and discarded. It is the confident draft containing a specific commitment nobody made — a price, a date, an availability — which reads well and creates an obligation. Drafts should therefore be reviewed for specifics rather than for style, which is the opposite of how people naturally read them.

Automatic sending sits behind a genuinely different bar, and the honest case for it is narrow: acknowledgements that commit to nothing, confirmations of something that already happened, and messages whose entire content is drawn from a record rather than composed. The moment content is generated rather than assembled, the message becomes a statement your business made, and the case for a human read is the same case as for anything else you sign.

One boundary worth stating explicitly. Connecting an assistant to a mailbox connects it to the whole mailbox, including correspondence with your accountant, your solicitor and your staff. If triage is the job, a connection scoped to a single label or folder does the job with a fraction of the exposure, and it is worth the extra ten minutes of setup that scoping requires.