--- title: "Testing the tone of an AI reminder: a review grid for sensitive cases" canonical: https://www.billabex.com/en/blog/ai-reminder-tone-review-grid/ lang: en alternate: https://www.billabex.com/fr/blog/ton-relance-ia-grille-evaluation.md updated: 2026-10-17 index: https://www.billabex.com/llms.txt --- # Testing the tone of an AI reminder: a review grid for sensitive cases A reminder can open with “we hope you are well” and still be unsuitable to send: the amount is wrong, a known dispute is ignored or a judicial consequence has been invented. A short, direct message can respect the relationship when it acknowledges the facts and asks for something the recipient can actually do. Testing an AI reminder's tone therefore requires more than choosing a favourite formulation. The question is whether the message remains accurate, appropriate to the context and understandable in a sensitive situation. A shared review grid allows finance and sales to explain disagreements without reducing the assessment to “too cold” or “not firm enough”. For French and European customer relationships, the review should also respect the language and contractual context of the exchange. A score from another setting does not establish that a message is appropriate for your portfolio. ## Review the facts before discussing politeness Microsoft explicitly distinguishes the writing quality measured by its coherence and fluency evaluators from factual correctness. That separation is useful in collections: a strong style score does not establish that the balance or history is accurate. The consulted documentation also states that those evaluators currently support English. [Microsoft Foundry general-purpose evaluators, 2 June 2026](https://learn.microsoft.com/en-us/azure/foundry/concepts/evaluation-evaluators/general-purpose-evaluators). Start by comparing the request with the case record. Does the amount reflect payments and credit notes? Is the recipient associated with the right customer? Is the invoice overdue or scheduled for a later date? Resolve those questions before debating the greeting or sign-off. Diplomacy does not require weakening a verified fact. It means avoiding an accusation or certainty unsupported by the available evidence. “We have not identified the payment in our records” leaves room for reconciliation. “You refuse to pay” attributes an intention that needs evidence and is often unnecessary to obtain the next useful piece of information. ## Choose situations that require adaptation A simple case with no previous response or payment exposes different problems from a long-running exchange. Build a small collection inspired by actual work, using properly anonymised information or wholly synthetic cases. Retain enough context for the writer to determine what to request. | Situation to examine | Expected behaviour | | --------------------------------------------- | -------------------------------------------------------------- | | A partial receipt has just been reconciled | Acknowledge it and address only the relevant remaining balance | | The supplier has not sent a promised document | Address that action before blaming the absence of payment | | One invoice line is disputed | Distinguish the known reservation from undisputed elements | | A proposed payment still has a condition | Preserve the condition without presenting a firm agreement | | The recipient has left the role | Find the correct route without attributing non-cooperation | NIST recommends evaluating systems under conditions resembling their intended use and documenting generalisation limits. It advises against extrapolating from narrow or anecdotal assessments. This US reference is a voluntary method, not certification of your tool. [NIST AI 600-1, July 2024, printed pages 30–31](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf). Keep some cases separate from those used to refine instructions. Otherwise, improvement may reflect familiarity with the reviewed examples. For conditional language, the article on [AI call notes and commitments](https://www.billabex.com/en/blog/ai-call-notes-payment-commitments/) helps identify passages whose qualifications must survive the drafting process. ## Use a short grid with errors that prevent sending The proposed grid has four dimensions. Accuracy checks amounts, dates and identities. Context checks whether the message accounts for previous exchanges and actions. The request dimension checks whether the next step is identifiable. Tone checks for judgemental language, inappropriate familiarity or unjustified pressure. For each dimension, record an observation and identify the relevant passage. “Aggressive tone” does not explain the correction. “This sentence announces seizure when the case contains no established decision or corresponding action” identifies the problem. Some failures should prevent sending even when other dimensions are satisfactory. These concerns have documented precedents. In its targeted French investigation conducted in 2020 and published in December 2022, DGCCRF found anomalies at 18 of 68 inspected establishments, or 26.5%. The investigation included confusion between amicable and enforced collection. This inspection sample does not measure all providers or current AI reminders. [DGCCRF commercial-debt investigation](https://www.economie.gouv.fr/dgccrf/laction-de-la-dgccrf/les-enquetes/controle-du-recouvrement-de-creances-commerciales-pendant). Adding “kind regards” cannot repair an unsupported threat. Treat it as a content and decision failure, then rebuild the message from the available facts. Likewise, a friendly sentence cannot make an incorrect amount acceptable. ## How a 95% average can hide four unusable messages Consider a simulation with twenty messages, each checked against the four dimensions. That produces 80 observations. Suppose 76 pass and four fail, each in a different message: an incorrect balance, an ignored full payment, an unsupported threat and a contradiction of an existing promise. The aggregate is 76 divided by 80, or 95% of criteria satisfied. Yet only 16 of 20 messages pass without those blocking errors, or 80%. Reporting 95% as the proportion of correct messages changes the unit of measurement and hides the actual decision. These invented figures explain the method; they describe no model or customer. Separate quality by dimension from the proportion of usable messages. Also retain the nature of the failures. Four amount errors require different corrections from a vague request or an unnecessarily distant formulation. The useful result identifies which data, rule or review step needs to change. A small successful test is a reason to continue a controlled evaluation, not proof that every future message will be correct. Keep the boundaries of the tested cases visible in the review record. ## Resolve reviewer disagreements using evidence Ask someone from finance and someone close to the customer relationship to review the same cases independently first. Then compare disagreements against the criteria. If one person finds the request too direct, establish whether context is missing or whether it simply states a known obligation clearly. A simulated case starts with a €3,600 invoice, a €1,200 receipt and an allocated €600 credit note. The remainder is €1,800. “Please pay €3,600 despite our previous requests” is incorrect regardless of style. A usable message recognises the receipt and credit note, identifies the €1,800 remaining and asks a precise question about payment or the blockage. Do not force a more empathetic phrase to speculate about the customer's personal circumstances. The useful context concerns the transaction and the action needed. A courteous question can acknowledge uncertainty without inventing an explanation for the delay. Handling [one disputed line](https://www.billabex.com/en/blog/disputed-invoice-line-partial-payment/) and identifying [exceptions that require human review](https://www.billabex.com/en/blog/ai-collections-agent-human-exceptions/) complement the assessment. The right correction may be clarification or pausing the message, rather than producing a more elegant sentence. After changing instructions or a model, rerun sensitive cases and examine some new situations. Record the team's decisions so the criteria remain understandable. To assess [the Billabex agent](https://www.billabex.com/en/product/agent/), use your demonstration cases and examine how context changes the proposed reminder. Validation concerns observed behaviour in those cases, without promising universal perfection or inferring collection gains from a writing score. ## Sources - Microsoft Foundry, _General purpose evaluators_, 2 June 2026. - NIST AI 600-1, July 2024, MEASURE 2.3 and 2.5. - DGCCRF investigation of commercial debt collection in 2020, published 26 December 2022. Sources consulted on 7 September 2026.