Knowing how to ask ChatGPT for an email with an appropriate tone does not demonstrate that a person can decide whether that text can be sent to a client. An apparently correct draft may include an unconfirmed date, a non-existent commercial condition, an operational promise beyond the available scope, or information that should not be shared. The skill worth observing is not writing alone: it is human review before using the output.
ChatGPT training can assess this skill through a brief, safe test based on fictional cases. Each participant receives a simulated enquiry, an internal source created for the exercise, and a deliberately imperfect draft. They must then correct it, explain their decisions, and identify which aspects require validation. This focuses the assessment on judgement applied to the role, rather than on the number of prompts used.
The issue: fluent text can still be unsuitable
Generative AI can help structure a response, summarise provided information, or suggest an initial tone. However, ChatGPT documentation recommends reviewing, correcting, and verifying responses before using or sharing them. In commercial or customer service communications, that review should check that every relevant statement is supported by the available information and that an assumption is not turned into a commitment to the client. [1]
For this reason, a test should not reward an email merely for being longer, more persuasive, or more sophisticated. It should show whether the person removes unsupported claims, asks for missing information, and distinguishes between what they can communicate and what they must escalate. For example, if the draft states that an issue will be resolved by Friday, the expected response is not to retain the sentence because it sounds helpful, but to check whether the case sheet confirms that timeframe.
Choose one low-risk case for each role
Designing an exercise for each function makes the review observable. For sales, a fictional quote request may be used; for customer service, an update on a simulated issue; and for administration, a request for incomplete documentation. It is advisable not to combine all three roles in a single test: each works with different authorisations, language, and decisions.
The case should remain within low-risk communications. It is preferable to exclude contractual decisions, financial commitments, complex complaints, or messages that require specialist validation. A hypothetical example: a fictional client asks whether a service includes a specific integration. The exercise sheet states that this must be confirmed with the technical lead. If ChatGPT writes that the integration is included, the participant must remove that statement and set out the next validation step.
- Define a fictional recipient, objective, and context.
- Indicate which person or role can validate prices, deadlines, scope, or conditions.
- Limit the material to simulated data and avoid connecting it to real inboxes, CRM systems, or automations.
Prepare a test pack with an authorised source
The pack may include four elements: the fictional client's initial message, a current internal sheet for the exercise, the AI-generated draft, and a work instruction. The internal sheet is the authorised source for the test. If a piece of information does not appear there, the person should not fill it in based on intuition or ask the tool to invent it: they should record that confirmation is needed.
Include between two and four deliberate issues in the draft so that the task requires review rather than simply rewriting. These may include an unconfirmed date, an included service that does not appear in the sheet, a request for unnecessary data, or an unsuitable tone. ChatGPT features may vary depending on the plan, platform, region, and workspace configuration; therefore, the test should assess review behaviour that does not depend on a specific tool feature. [1]
- Include an explicit instruction not to copy real client, operational, pricing, credential, or internal conversation data.
- Visually separate the authorised sheet from the draft to be reviewed.
- Clarify that the aim is to identify limits and correct the text, not to produce an apparently perfect response.
Assess review decisions with a simple rubric
A shared rubric reduces subjectivity and makes it possible to compare exercises. It can use three levels for each criterion: meets expectations, requires correction, and does not meet expectations. Assess accuracy, identification of missing information, prudent handling of data, and communication usefulness. This final dimension considers whether the response has an appropriate tone, explains the next step, and addresses the recipient correctly.
The NIST AI RMF proposes defining measures linked to the context of use and documenting test sets, methods, and evaluation results. Applied to this training, it is advisable to retain the case used, the initial version, the reviewed version, the rubric, and a brief note explaining the reason for each change. The aim is not to turn the exercise into a complex audit, but to have useful evidence for improving practice. [3]
- Accuracy: corrects or qualifies anything the sheet does not support.
- Missing information: identifies what is needed before confirming a price, deadline, scope, or solution.
- Confidentiality and limits: removes unnecessary data and indicates when an authorised person must be consulted.
Compare versions and repeat with a new case
Comparing the first draft with the reviewed version reveals learning better than assessing the final text in isolation. Ask each participant to mark the changes and explain, in one or two sentences, why they made them. This makes it possible to distinguish a stylistic correction from a relevant decision: removing an unauthorised commitment, requesting additional information, or directing the enquiry to the role that must approve it.
Conduct a later second test with a different scenario, rather than with the same memorised exercise. A sales, support, or quality lead can review a sample together with the trainer, always using unsent drafts or simulated cases. NIST includes the participation of domain specialists and people who did not directly develop the solution in evaluation processes. Repeated errors can be used to adjust the examples and practices in the next training session. [3]
Frequently asked questions
What errors should be included in a fictional draft generated with ChatGPT?
Include errors that the person can identify using the case sheet: an unconfirmed date, an unstated condition, a promise beyond the available scope, a request for unnecessary information, or an incorrectly assigned next step. Avoid high-risk scenarios or those requiring legal, financial, or specialist assessment.
How can you measure whether ChatGPT training has transferred to the role?
You can repeat the assessment using a new fictional case and apply the same rubric. Observe whether the person identifies unsupported statements, requests information before responding, and refers validations to the appropriate person in charge. The comparison should be used to guide further training support, not to promise automatic results.
Is it necessary to use real client data to assess human review?
No. Simulated data reduces unnecessary exposure of information and makes it possible to control which response is supported. In addition, in managed workspaces, access to tools and data depends on configured members, roles, permissions, and policies; authentication alone does not replace those controls.