What decision should the ROI model support?
Decide whether you are evaluating an initial trial, additional licences, training or a wider workflow change. Each decision has different costs and evidence needs. A model built to justify a purchase after the decision has already been made is less useful than one that can support continuing, revising or stopping the proposal.
Choose a task and a group whose work can be observed. An accounts exception review or an administrative handover gives the reviewer a concrete output to inspect. Record what acceptable work looks like and who decides whether the result meets that standard. Counting generated answers without checking their usefulness will overstate what the pilot has established.
Keep the scope limited enough to understand. If the trial changes the source process, account configuration, staffing and reporting format at the same time, it becomes difficult to attribute the result. Record those changes as part of the evidence instead of presenting every difference as an effect of Copilot alone.
Which licence cost should we use?
Use the quote or invoice for the account and billing term the organisation is considering. Microsoft’s guidance distinguishes included Copilot Chat experiences from features associated with an add-on licence1. Confirm the task’s access requirements with the administrator before assuming every participant needs a new paid subscription.
Record currency, tax presentation, billing commitment, seat count and any usage-related conditions in the cost sheet. Avoid copying a promotional figure from an old article. The actual terms may differ from a general reference price, and a monthly equivalent can hide a longer cash commitment. Ask purchasing staff to confirm the basis used in the model.
Separate existing costs from incremental costs, while showing both where the decision needs them. An account already paid for may still have a renewal or expansion consequence. Make that assumption visible. Do not treat existing expenditure as free in one option while charging it fully to another without explaining the comparison.
How do training and preparation enter the calculation?
The published private workshop starting point is $4,000 + GST for up to ten people, with $200 + GST for each additional participant2. Use the actual agreed proposal for a decision model. Scope, delivery and other agreed terms must be known before the published starting figure is treated as the final cost.
Include the organisation’s preparation and participation effort. Someone will select and clear examples, confirm access, attend the session and review later practice. Those activities may be worthwhile, but they are part of the investment. Keep trainer fees, internal time and software costs in separate rows so a reviewer can see what has been included.
If the pilot needs ongoing prompt or source maintenance, record who will do it and how the estimate was made. A workflow that works well only while a specialist continually repairs it has a different cost from a method the team can maintain. Use observations from the trial to refine that assumption.
How should task effort be measured?
Record the normal process on an example of comparable difficulty, then record the assisted process. Split the latter into preparing input, drafting, reviewing and correcting the result. Include the effort of the final approver. A fast answer that requires extensive checking may not reduce the total work needed to complete the task.
Keep unsuccessful attempts in the log. If a draft is abandoned and the task is completed manually, record both the attempted AI work and the fallback effort. Excluding those attempts makes the model look better than the process participants actually experienced. Note why the draft failed and whether the issue was access, source quality or output quality.
Use several permitted examples and describe their differences. A tiny sample cannot establish a stable result for every workload. Retain the raw observations so the reviewer can inspect the spread, not just an average. Where timing is estimated rather than observed, label it as an assumption and test how much the decision depends on it.
How do we turn observations into a value estimate?
First calculate the difference in total effort for accepted work. Then ask how often the same task occurs and whether the observed method can be repeated under ordinary conditions. Use the organisation’s own workload evidence. Avoid multiplying a demonstration’s best result across every employee and every working day.
If the business wants a financial model, use a cost basis approved by its finance owner and explain what that value represents. Time released may support other work without reducing payroll or external expenditure. Show capacity value separately from a cash saving that requires an actual change in spending. This distinction prevents an attractive estimate from being mistaken for money already saved.
For a return percentage, subtract the included costs from the supported benefit estimate, divide by the included costs and express the result as a percentage. Keep the time period consistent and do not calculate a ratio when the denominator is absent or meaningless. The formula cannot make weak assumptions into reliable evidence.
What should a sensitivity review show?
Vary the assumptions that are least certain: task frequency, acceptance rate, review effort and the extent to which useful capacity can be used. Show whether the decision changes under a less favourable case. A result that depends on every assumption being optimistic should be described as fragile rather than presented as the expected outcome.
Consider quality consequences alongside time. A serious error may require specialist review or make a task unsuitable for the proposed process. Do not force every consequence into a speculative dollar amount. Record the risk and the oversight needed, and let the responsible owner decide whether the workflow remains acceptable.
Make clear which inputs came from invoices, which from observation and which from judgment. A reviewer should be able to challenge an assumption without rebuilding the entire sheet. Keep the cost model understandable enough that the business can update it when the licence quote, source process or review standard changes.
| Input | Preferred evidence | Uncertainty to test |
|---|---|---|
| Licence and training costs | Current quote and approved proposal | Scope or billing changes |
| Task frequency | Observed workload | Seasonal or unusual demand |
| Total effort | Baseline, draft, review and repair log | Different task difficulty |
| Useful capacity | Manager’s plan for the released time | Whether other valuable work is available |
| Cash effect | A confirmed spending change | Capacity may not reduce expenditure |
How should the pilot decision be recorded?
Write a short decision note that states the task, evidence, limitations and next action. Include examples of accepted and rejected outputs and explain what the team learnt about checking. The decision might be to continue a narrow workflow, change the training plan, resolve access or stop. Each can be reasonable when supported by the observations.
Do not turn the pilot into a customer success claim without permission and an accurate account of what happened. A fictional training workbook is not evidence of commercial results. Likewise, an estimated return should not be described as achieved. Keep the distinction between measured effort, modelled value and actual financial change in the final note.
Assign an owner for reviewing the result after the next relevant change. If the task becomes more complex or a new group joins, the original estimate may no longer apply. The model should support an ongoing decision, not become a static justification that nobody revisits once the licences are purchased.
What can a workshop contribute to this work?
A workshop can help the team define a bounded task, practise output checking and agree how to record effort. Participants can compare a flawed draft with a corrected result and identify which review steps are essential. These activities prepare a better pilot; they do not themselves prove a return on investment.
Bring a fictional example, the intended acceptance rule and the roles involved in approval. Use the session to test whether another person can repeat the method and whether the review sheet catches the deliberate errors. Keep the resulting materials so the pilot can be evaluated on work done after the workshop.
The Copilot rollout and manager courses are relevant to planning and reviewing a trial. Tell thecolab.ai which decision the pilot should inform, what accounts are available and who owns the task. We can discuss the training scope while leaving purchasing and financial assumptions with the organisation’s authorised reviewers.