Email Triage for Small UK Firms: A Classify, Draft and Approve Pattern

#email #llm #automation #python

Disclosure: I run NxFlowAI, an automation agency serving UK businesses remotely from Mumbai. The pattern is provider-neutral.

Many small UK firms still run on one shared inbox: info@ or hello@. New enquiries, supplier invoices, complaints, spam and job applications all mix together, and the owner sorts them at night. Here is a small pattern I use as an AI automation agency building inbox triage for UK firms: classify, draft, approve. Nothing is sent without a person.

1. Classify into a short, fixed list

CATEGORIES = ["new_enquiry", "existing_customer", "invoice_or_supplier",
              "complaint", "job_application", "spam_or_marketing", "other"]

def classify(email):
    result = llm.classify(email.subject, email.body_text, CATEGORIES)
    if result.confidence < THRESHOLD:
        return "other"            # unsure goes to a person
    return result.category

Keep the list short. Eight categories the team understands beat thirty nobody uses.

2. Route by category

new_enquiry:         {label: "Enquiry", assign: sales, draft_reply: true}
existing_customer:   {label: "Customer", assign: account_owner, draft_reply: true}
invoice_or_supplier: {label: "Accounts", assign: bookkeeper, draft_reply: false}
complaint:           {label: "Complaint", assign: owner, draft_reply: false, alert: true}
job_application:     {label: "Jobs", assign: owner, draft_reply: false}
spam_or_marketing:   {label: "Low priority", assign: none, draft_reply: false}
other:               {label: "Check", assign: owner, draft_reply: false}

Complaints never get an AI draft. They get an alert.

3. Draft, do not send

For enquiries, the system writes a draft reply into the thread using your approved snippets (hours, services, next steps). A person reads, edits and sends.

4. Log every decision

Store message id, category, confidence, who approved, and whether the draft was edited. Heavily edited drafts tell you which snippets to improve.

5. Test with real mail

Anonymise a few hundred real emails and label them by hand before trusting any classifier. Include forwarded chains, replies with no subject and emails with attachments only.

We build this kind of triage after a 72-hour audit of the inbox workflow. If you are weighing a ready-made helpdesk bot instead, our write-up on custom AI versus packaged chatbots covers the trade-offs.

DEV Community