AI email triage with a review step

It reads each customer email that reaches a shared inbox and fills in a fixed record: who sent it, for which organisation, what they are asking for, by when, and what is missing before anyone can act. Every value points back to the phrase in the email it came from, which is what the highlights on the bench show.

One model call produces the record, and code checks it against the schema before anything else happens. The email is then routed by rules your team can read: a clear request gets a drafted reply, a complaint or anything unclear goes to a person's queue without a draft, and mail that is not from a customer is set aside. A record that fails the schema goes to a person.

Nothing is sent by the tool. A drafted reply waits in a queue until a person reads it, edits it if needed and sends it from the inbox they already use.

An email routed to a person arrives with the record filled in as far as it got and the reason it stopped. The routing rules are agreed with the people who own the inbox, and only they change them.

It reads from and writes to the systems the business already uses: the inbox the emails arrive in, and wherever orders, bookings or customer records are kept, so a request can be checked against them. Which systems those are is the first thing we establish on the first call.

If one of them has no way to connect, we say so on that call.

• It cannot judge tone or urgency reliably. A polite email from an angry customer, or a calm one about something urgent, can read the wrong way, so anything ambiguous goes to a person instead of getting a draft.
• It records what the customer wrote, not what they meant. A wrong order number in the email is a wrong order number in the record, marked as written by the customer.
• Long threads with several people and several requests are harder than single emails. It reads the latest message and sends the thread to a person when it finds more than one request.
• Its drafts sound generic until your own past replies and house rules are written into its instructions, and they still need a person to read them after that.

• One inbox and one kind of email, chosen with you on the first call.
• The record and the routing rules, written down and agreed before we build.
• 30 to 50 of your past emails with the correct record for each, used to score the tool before it reads live mail.
• A pilot on exported mail or read-only access, then live use with every draft approved by your team.
• A log of every email read, the record produced, the route taken and who approved what.