Training a team to use ChatGPT, Microsoft Copilot or another AI tool often begins with familiar tasks: summarising an email, completing a CRM record or classifying an incident. The problem arises when, to make the exercise more realistic, a real conversation, an internal export or a client document is copied. The practice then stops being just a training activity and starts to depend on data, permissions and configurations that may not have been reviewed for that purpose.
The alternative is to create a training test bank: entirely fictional cases that reproduce the structure, language and decisions of each role without reusing recognisable information from the organisation. In this way, AI training by role can test something more important than ease in writing prompts: whether each person knows how to identify limits, ask for clarification, review an output and stop an action when data is missing.
The mistake of practising with real information “just to test it”
A real email may contain personal data, confidential information, commercial references or context that is not obvious at first glance. The Spanish Data Protection Agency recommends not entering confidential information from the organisation, employees or customers into AI tools, and notes that processing may also involve metadata in addition to the text entered. For a learning session, it is advisable to start with material invented from the outset, rather than partially redacted documents. [4]
It is also important to separate training from operations. A practice exercise seeks to observe how a person responds to a defined task; it should not become implicit authorisation to update systems, send communications or feed automations. In Microsoft Copilot environments, responses are based on the user's existing permissions. Therefore, teaching people how to use the tool does not replace reviewing what content is shared or who can access it. [5]
How to design a fictional sample that resembles the work
A good sample imitates the task, not the real data. It can retain columns, incident types, process steps and usual vocabulary, but it must use invented names, companies, products, amounts, dates, addresses and identifiers. If a data point makes it possible to recognise a person, an operation or an internal source, it is not a suitable sample for a general training exercise.
Each case should contain deliberate friction. Include at least one ambiguous piece of information, one missing item and one contradiction. The goal is not for the AI to find an elegant answer at all costs, but to check whether the participant distinguishes a fact from an assumption and knows when to stop. For example, a delivery date may not match the date stated in an order; the valid response should flag the discrepancy, not choose a date without justification.
- Define the role: sales, administration, operations, support, management or IT.
- Describe a single observable task and an expected outcome.
- Indicate which fictional data may be used and what information is deliberately missing.
The minimum case template and three exercises
The template can be simple: target role, objective, fictional material, instructions, output format, conflicting information, escalation reasons and review criteria. Add a version for the case so that similar exercises can be compared over time. NIST frames AI risk management as governance, mapping, measurement and management activities, and includes scope definition, the skills of those operating the system and human oversight as relevant elements. [2]
Hypothetical example for sales: provide an invented conversation with a client and request a draft CRM update. The person must separate confirmed fields from those pending validation. Hypothetical example for administration: use three simulated orders in which the reference, date and quantity do not match; the outcome must record the exception. Hypothetical example for support: present fictional incidents and request a priority table indicating what evidence is missing before proposing an action.
- Do not assess only whether the response is well written.
- Ask for facts, assumptions, uncertainties and the next human review to be identified.
- Avoid having the exercise include real sending, changes or actions.
Assessing human review, not just the prompt
A useful rubric observes verifiable behaviours: whether the person identifies contradictions, retains the requested format, avoids inventing fields, distinguishes proposals from facts and escalates a decision when appropriate. You can also ask them to identify the fictional source of each relevant data point. This prevents a fluent response from being mistaken for a response suitable for entry into a CRM, an email or an automated workflow.
Testing should take place before expanding a use case and continue during operations, comparing results with references and documenting what is observed. The NIST framework states that measurement includes evaluating AI systems and recording results that help manage identified risks. In training, this means retaining only the necessary evidence: assigned task, case version, response, correction and need for further reinforcement. [3]
Moving from simulation to a controlled internal test
Passing fictional exercises does not demonstrate lasting competence or authorise automations without review. If the company decides to test an internal case, the next step should be gradual: limited scope, expressly authorised information, reviewed permissions, a person responsible for validation and a clear criterion for stopping the pilot. The review should consider both the AI output and the quality and exposure of the source material.
Available capabilities may change depending on licences, connected services, data retention and workspace configuration. In the case of ChatGPT for workplace environments, administrative controls and the behaviour of files or connectors depend on the applicable configuration. Before turning a practice exercise into a routine process, it is advisable to confirm these limits with the people responsible for the technical environment and the information. [1]
- Start with a task with limited impact and an output that can be reviewed.
- Establish who validates the result and where the correction is recorded.
- Use observed failures to adjust training, test data or controls.
Frequently asked questions
Why is it not enough to anonymise a real document to train the team?
Hiding some fields may not remove references, context or metadata that make the information recognisable. Creating a fully fictional case makes it possible to practise the same task with a clearer separation between learning and operations.
What should be scored in practical AI training?
In addition to text quality, it is advisable to assess whether the person detects missing or contradictory data, follows the format, distinguishes facts from assumptions, avoids inventing information and requests human review when appropriate.
Can a company move directly from a course to automating a task?
It is not advisable to assume so. Training helps observe skills in defined cases, but automation requires separate assessment of scope, permissions, data quality, human validation and operational limits.