Training a team on Microsoft Copilot is not just about explaining features or gathering attendees for a demonstration. The problem arises when, back at their desks, people do not know what information they can use, how to frame a reviewable request, or how to spot an incomplete response. A task-based assessment makes it possible to prepare for that situation before training and observe specific behaviours, not merely interest or attendance.
The approach is straightforward: choose a common, low-risk task, create fictional or anonymised materials, define a rubric, and repeat the exercise after training. The goal is not to turn the assessment into a prompt-writing exam or to attribute unmeasured business results to it. It is to check whether each role can request, verify, and correct a Copilot output within defined boundaries.
Start with the Copilot experience the team will use
Before drafting an exercise, confirm the actual configuration available to each participant. Microsoft distinguishes between experiences and requirements that may vary according to the application, identity, licence, and environment configuration. It also indicates that its product naming has evolved, so it is advisable to avoid materials based on assumed names or capabilities and to verify the tenant before scheduling the practical session. [1] [3]
Prepare a brief profile for each group: work account, Microsoft Entra ID sign-in access, the application to be used, available practice documents, and, if the task depends on email, calendars, or meetings, the required mailbox conditions. This check prevents the design of an activity for an experience the team does not have enabled. If technical conditions are missing, the useful conclusion is not to fail the person, but to adjust the environment or redesign the task. [3]
- Define the specific application and experience to be assessed.
- Check that everyone uses an authorised work account.
- Limit the assessment to sources and permissions reviewed in advance.
Choose a recognisable task and prepare safe data
A good assessment is short, repeatable, and relevant to the role. For administration, it could involve turning three fictional emails into a follow-up summary. For sales, it could involve drafting an initial email after a simulated meeting. For customer service, it could involve summarising an incident and identifying missing information. For a manager, it could involve turning fictional notes into agreements, owners, and outstanding questions. These are hypothetical examples: each organisation should adapt the scenario to its processes and authorisations.
Use fictional or anonymised documents, messages, and records. Include controlled ambiguities, one missing data point, and one or two deliberate inconsistencies. This allows you to observe whether the person identifies the limitation, asks for clarification, or flags uncertainty, rather than accepting a convincing response without checking it. Microsoft states that Copilot respects existing permissions and accesses only organisational content to which the user already has access; precisely for that reason, responsible practice requires reviewing permissions and content before involving internal information. [2]
- Include fictional names, dates, and figures that need to be checked.
- Avoid personal data, confidential information, and unnecessary real cases.
- Specify which sources the person may consult during the exercise.
Turn the assessment into an observable rubric
The rubric should assess behaviours that a reviewer can identify, rather than impressions such as “uses AI well”. You can assign a simple scale to understanding of the task, clarity of instructions, checking of figures, names, dates, and recipients, identification of uncertain information, relevance of included data, and the decision to correct, discard, or escalate the result. Define in advance what evidence will be retained: the request, the response obtained, the corrected version, and the reviewer’s observations.
Human review is a central part of the exercise. The NIST framework recommends selecting metrics related to relevant risks, documenting human oversight, and retaining signals such as reported errors, overrides, and decisions on whether or not to proceed. Applied to this practice, it is not enough to record that someone obtained a response: it is important to know what they detected, what they corrected, and when they decided that the output was not suitable to proceed. [5]
- Create consistent criteria for people carrying out the same task.
- Describe what constitutes a critical error in that simulated scenario.
- Record necessary corrections without using the rubric to penalise prudent experimentation.
Repeat, interpret, and improve the training
Carry out an initial assessment before training to establish the starting point, followed by a second one using an equivalent, rather than identical, task after practice. You may add a review weeks later to observe whether the behaviour is sustained in a similar simulated situation. Separate three indicators: participation, meaning who completed the practice; applied learning, meaning who met the rubric criteria; and transfer to the role, meaning who repeats the appropriate process later on. This distinction prevents attendance from being confused with competence.
Microsoft adoption and usage reports can provide context on activity by application or feature, but they do not replace a review of the quality of a specific task. Use them as a complementary signal, not as the sole proof of learning or work improvement. Close each round with operational decisions: reinforce data checking, clarify which sources are permitted, modify the samples, review permissions, or reserve certain tasks for mandatory human validation. [4]
- Compare evidence and criteria, not only the number of uses.
- Identify error patterns that should be addressed in the next session.
- Tell the team that the purpose is to improve the work process and its controls.
Frequently asked questions
Should the task-based assessment be carried out before and after training?
It is advisable to carry out both phases if you want to observe change. The first helps design training suited to the starting point; the second makes it possible to check performance using an equivalent task. A later review may provide information about transfer, provided that the rubric and human review remain consistent.
Can real documents be used in a Microsoft Copilot assessment?
It is advisable to prioritise fictional or anonymised materials. If an organisation considers it necessary to use internal information, it should define the scope in advance, check permissions, and limit access to authorised people. The assessment should not expand access or expose data that are not necessary for learning.
Do Copilot usage reports demonstrate that the training has worked?
Not on their own. They can show activity and adoption, but they do not make it possible to determine automatically whether a response was correct, whether it was reviewed appropriately, or whether the person identified a limitation. Combine them with a task-based assessment, review evidence, and improvement decisions.