Separate drafting and tone review for payment messages

Automating payment reminders has an uneven balance of benefit and risk. A good reminder may speed up payment, while a threat, excessive pressure, or an unsupported promise can damage a long-standing relationship and create legal risk. The system therefore needs a reliable way to stop an occasional harmful draft.

Why self-review can repeat the drafting error

A common design asks the model to review its own draft. The text and its assessment then come from the same context and understanding of the goal, so an error made during drafting may be repeated during review.

In the receivables monitoring we built, the assessment is independent. The first model receives the case, contact history, and tone policy and writes a draft. The second receives only that text, with no goal or history. It checks for threats, pressure that is excessive for the debt's age, unsupported promises, and language that could damage the relationship.

The two models work in separate contexts and have different tasks. When their assessments differ, a rule in code stops the draft and records the reason.

A boundary enforced by capability

The system deliberately has no sending function, mail integration, button, or alternative path. A specialist must send an approved draft outside the system.

The system can create a draft but cannot contact a client. The absence of a send capability is easy to verify, while disputed cases and clients on hold are excluded from drafting altogether.

A contact-risk radar

Around this mechanism sits the working view: receivables on a contact-risk radar, where a case's position comes from the debt's age, the conversation history, and open disputes. The specialist starts with cases where contact could cause harm. Relationship risk sets the order of work.

Contact policy comes first

Two models cannot replace a collections policy. Before implementation, specify when the system may prepare a reminder, which facts it may use, and which states exclude drafting. Common exclusions include an open dispute, an unreconciled payment, a complaint, an active promise to pay, and a case already handed to legal counsel.

Tone policy needs concrete allowed and prohibited examples for each stage of delay. “Be polite” is too vague for consistent review. Useful guidance covers permitted calls to action, forbidden threats, how consequences may be described, and language that requires legal review.

The source state matters as much as the wording. A courteous reminder based on stale payment data is still harmful. Deterministic checks for payment, dispute, and hold status should run before either model.

Evaluating the reviewer

The errors have different costs. Releasing a harmful draft matters more than holding a safe one, but an overcautious reviewer can send the whole queue back to people. Report harmful drafts released, safe drafts held, and the reasons for both outcomes separately.

Build the evaluation set from the client's risk categories after removing personal data: disputes, different ages of debt, partial payments, earlier commitments, and relationship segments. Collections, customer-service, and legal owners should label the material. Tone is company policy, so the technical team should not define the reference alone.

Review held drafts regularly after launch. A recurring reason may indicate a weak drafting instruction, missing case data, or a change in commercial policy. The reviewer should not rewrite its own policy from previous decisions.

Cost and operating value

The prepared run cost USD 0.015652 per draft with both model checks. It excludes receivables integration, policy maintenance, and specialist time. Two model calls will cost more than one, so independent review belongs where the consequence of bad communication justifies it.

Measure preparation time, correctly excluded cases, specialist edits, and contact outcomes such as a reply or agreed payment date. Cash collected is a weak short-pilot measure because many factors beyond reminder wording influence it.

Pre-launch checklist

1. Which case states prevent drafting altogether?
2. Are payment, dispute, and previous-contact data current?
3. Does the tone policy contain examples for each stage of delay?
4. Does the reviewer receive only the text needed for assessment?
5. Is sending technically absent from the system?
6. Who reviews harmful releases and excessive holds?
7. How does a specialist record an edit and its reason?
8. When does the case leave automation for legal or account management?

The measurement

The labeled set contained five harmful drafts and three safe ones. The controller rejected all five harmful texts, cleared all three safe texts, and released none incorrectly. Across the prepared workflows, recorded cost was USD 0.015652 per draft with both checks and the 95th-percentile time was 21.26 seconds. This verifies the gate on prepared correspondence; production safety still needs testing against the client's real tone, disputes, and contact history.

The same design can apply wherever communication affects money, contracts, or reputation. Separate drafting from review and do not give the system a function whose misuse would create the greatest harm.

Details on the case page, and a free process scan will help assess where such double control makes sense in your communication.