How to produce and dispatch customer letters with an AI agent

The agent prepares the run and a named person releases it. It selects the approved template, merges data from the system of record, applies suppressions for deceased, gone-away and accessibility needs, builds a proof sample and an exceptions list. Compliance approves the template wording, an operations lead releases each bulk run, and every letter traces to its data.

What is the letters process, and who owns it?

Regulated firms send large numbers of letters: annual statements, notices of rate or fee changes, arrears letters, account closure notices, renewal invitations, responses to requests. Many carry legal weight. A notice of a rate change must reach customers a set time before it takes effect. An arrears letter may be a required step before further action. A wrong figure in a statement can become a complaint.

The customer communications team usually owns production and dispatch. Compliance or legal owns template wording. The product team owns the trigger and timing. A print and mail supplier, or an email platform, handles physical or electronic delivery. The agent sits between the systems of record and the dispatch channel.

What does the agent prepare, and what does a person approve?

1. Agent, depth 1: identifies the population for the letter type from the system of record: customers whose rate changes, accounts in arrears at a stage, policies due for renewal.
2. Agent, depth 1: selects the approved template by letter type, product and language. Only template versions compliance has approved are eligible.
3. Agent, depth 3: merges data into each letter from sealed sources: balances, rates, dates and amounts come from the ledger or policy system, never from a spreadsheet someone prepared.
4. Agent, depth 3: applies suppressions and preferences: deceased customers, gone-away addresses, customers who asked for large print, braille or a specific channel, customers with a recorded vulnerability that changes how they should be contacted.
5. Agent, depth 1: builds a proof sample, chosen to cover every template variant, the extremes of each merged figure, and a random set.
6. Agent, depth 1: builds an exceptions list: missing data, figures outside expected ranges, addresses that fail validation, customers in more than one population.
7. Person, communications lead: reviews the proof sample and exceptions and releases the run, or stops it.
8. Agent, depth 3: dispatches the released run to the print supplier or email platform and records each letter's dispatch.
9. Agent, depth 2: drafts bespoke letters that fall outside templates, for a person to approve and send.

Why does release stay with a person?

A bulk run is many letters, sent at once, that cannot be recalled. A merge error that shifts one column puts the wrong balance on every letter. An outdated rate table sends notices with the wrong rate to everyone. One person looking at a well-chosen proof sample and a short exceptions list catches most of these, at little cost.

The proof sample should be designed to catch errors, not to confirm success. Include the highest and lowest values, every template variant, a joint account, a customer with a long address, a customer in a different language. If the reviewer only sees typical letters, they will only catch typical errors.

Which controls matter, and what evidence do they leave?

Approved template versions only. Evidence: the template version on every letter.

Merge data from sealed sources. Evidence: for each letter, the source and extraction time of each merged figure.

Suppressions applied before proofs. Evidence: the suppression list and the rule that removed each customer.

Release by a named person. Evidence: who released, when, and what proofs and exceptions they saw.

Dispatch confirmed per letter. Evidence: the supplier's or platform's receipt for each item.

What about letters that answer complaints?

Final responses to complaints are letters, but not routine ones. Each is written for one customer, carries the firm's decision and often an offer, and sets out the customer's rights to escalate. They follow their own process, set out in drafting final response letters to complaints, and should not be routed through bulk production.

What should you measure?

• Runs stopped at release, and why. A run stopped is a control working.
• Letters reissued after dispatch, the most important quality measure.
• Returned mail, which shows address quality.
• Complaints about letters: wrong figures, wrong recipient, wrong format.
• Notices sent inside their required period, for time-bound letters.

Dispatch volumes and failures belong in the daily figures the team already reviews; see producing daily operations reports.

Where does it usually go wrong?

Data comes from a spreadsheet. Someone extracts the population into a spreadsheet, edits it, and gives it to the agent. The letters are now only as good as that spreadsheet. Merge from the system of record.

Deceased customers are written to. Bereavement flags are set in one system and letters are produced from another. Check every system that records a death.

Notice periods are counted wrongly. The letter says the change takes effect on a date, but dispatch slips by two days and the notice period is short. The agent should check the effective date against the dispatch date and the required period, and stop the run if it no longer fits.

Duplicates go out. A customer with two accounts gets two copies, or a rerun after a failure sends the same letters again. Check dispatch records before release.

Template sprawl. Teams add variants for every edge case until no one knows which applies. Keep the set small and owned, and certify the agent on each one, following train, shadow, certify, widen.