Set AI adoption goals around work that people can complete and check, at the pace that work actually occurs. Name the task, the opportunities to try it, the conditions for accepting the result and the decision you will make at review. A daily login requirement cannot supply those details.
For a sponsor or department lead, the practical question is what you want colleagues to be able to do next. Someone learning an unfamiliar method needs a different commitment from a team operating an established workflow. Neither should have to invent extra AI activity to fill days when the agreed task does not arise.
Decide what the goal is asking people to achieve
“Use AI more” leaves employees to decide what more means and managers to invent a standard afterward. Replacing it with “save time” can be just as vague if nobody has established the current effort or the checks the finished work needs.
Choose the type of commitment first:
| Commitment | What colleagues are being asked to do | Evidence for the next decision |
|---|---|---|
Learn a suitable method | Complete supported practice, explain the important checks and identify a limit | The attempted work, checking decisions and remaining support need |
Establish a working practice | Use the agreed method on relevant work and deliver an acceptable result | Ordinary cases, exceptions, corrections and the recipient's acceptance |
Improve an outcome | Change a defined result without losing the required quality | A credible comparison of complete work, effort and other relevant changes |
These commitments can connect, but they are not interchangeable. A successful practice attempt does not establish a saving. An improvement in one person's drafting does not establish that the receiving team can use the result.
Microsoft's current AI strategy guidance recommends naming an outcome, choosing the relevant workflows and setting a baseline, target, owner and review date. It also asks leaders to decide what happens when a target is missed. That is vendor planning advice, rather than evidence that a particular target will improve your company. Its useful discipline is to make the expected change explicit before asking for a result.
If you cannot yet say what acceptable work looks like, begin with a learning or investigation commitment. Do not conceal that uncertainty inside a promised productivity percentage.
Match the goal to a real work cycle
Suppose Jamie Lee coordinates routes for a commercial laundry serving hotels. At a shift change, Jamie prepares a handover from approved collection records and driver notes. It must distinguish a missed collection from one completed later, preserve uncertain return quantities and identify who will follow up.
An approved assistant may help draft that handover. Jamie still checks it before the next coordinator relies on it. The draft does not dispatch a vehicle, change a service agreement or promise a remedy to a hotel.
Claire Bennett, an account manager in the same company, prepares a monthly service review. She brings together confirmed collection exceptions and their resolutions for the hotel account owner. Her preparation has a different cycle. She might have no relevant monthly review to prepare during a particular week.
The company can set clear expectations for both roles without demanding the same number of active days:
| Work | Suitable opportunity | A goal that can be examined |
|---|---|---|
Jamie's shift handover | A scheduled handover with the required records available | Prepare and check the agreed exception handovers during the trial, retaining corrections and reasons for using the existing method |
Claire's account review | The next completed monthly reporting cycle | Prepare a review from the approved records, explain unresolved cases and compare the complete preparation effort with the existing method |
A colleague awaiting permitted access | The prerequisite has not been supplied | Resolve the access decision with its owner and arrange a suitable attempt afterward |
The third row is not an alternative way to claim completed adoption. It records work the organization must do before it can reasonably expect the attempt. Keep the unmet prerequisite visible rather than recording either success or employee failure.
The Cabinet Office's 2025 account of scaling Assist makes this contextual choice explicit. Its team used a weekly-login definition, aggregated monthly, while acknowledging variation in members' roles and tasks. That is one programme's monitoring definition, not a frequency to impose on every job. The transferable question is why the chosen interval fits the work.
For deciding when someone has had a fair opportunity to return, use the guide to AI activation and retention. Define that opportunity before interpreting a quiet usage chart.
Write a goal the receiving colleague can check
Jamie and the next-shift coordinator should be able to read the same goal and agree what counts as a usable handover. “Create ten AI summaries” would leave them arguing about summaries that were generated but never checked, or that needed substantial repair.
A more useful agreement connects the method to the result:
That is a starting point to complete locally. Agree the actual dates, the ordinary handovers expected, who provides review time and where the evidence belongs. Avoid setting the volume by guessing what a persuasive adoption target would look like. Examine the workload and the current method first.
The result also has a physical reality outside the assistant. If a returned sack has not been reconciled with the collection record, a fluent handover must not silently turn an uncertain quantity into a confirmed count.
Agree the treatment of ordinary difficulties alongside the success condition:
- An incorrect draft needs correction. Count the review and repair work. Generation alone does not complete the task.
- An input is unsuitable for the approved tool. Keep it out and use the permitted route while the responsible owner resolves the question.
- A relevant task does not occur. Record that there was no opportunity in the period, rather than manufacturing an extra exercise and calling it routine use.
- The existing method is the better choice for a case. Retain the reason and examine it at review. Do not make employees hide that choice to protect a usage target.
An agreed practice exercise can still be useful when ordinary work is infrequent. Label its purpose as learning and assess the checks demonstrated. It does not become evidence of operating performance merely because the employee used the same tool.
Treat a usage benchmark as information to interpret
Frequency measures have a legitimate role. They can reveal whether a supported practice is returning, whether access is being used and where to ask questions. The problem is treating the measure's threshold as the work requirement without examining its purpose.
For example, Microsoft's AI adoption score documentation, updated in August 2026, describes a target of 12 active days in the previous 28 for licensed Copilot users, equivalent to an average of three days per week. That is a product-specific frequency score. It does not establish the quality of Jamie's handover or the right cadence for Claire's monthly review.
A sponsor can inspect that dashboard without copying its threshold into every person's objectives. Ask:
- Which tasks account for the recorded use?
- Which relevant opportunities were missed, and why?
- What does the accepted work show about capability?
If the objective is useful capability, a high count with weak checking deserves attention too.
In a September 2026 account of his first months leading Precisely, CEO Walid Abu-Hadba acknowledged that some lighter AI users demonstrated the judgment and skills the company needed, while some heavier use had not translated into greater capability or better outcomes. This is his company account, not an independently evaluated comparison. Its useful detail is the admission that the two observations diverged.
usage and capability aren't the same thing.
His account distinguishes employee capability from frequency of use and says the company still has work to do on assessing development.
The answer is not to abandon expectations. Jamie can be expected to learn an agreed method, perform the necessary checks and explain an exception. Those are clearer expectations than opening a chatbot before starting an otherwise unrelated day.
Review the commitment before raising the target
At the agreed review, compare what happened with what the goal actually asked for. Keep a failed output, a missing opportunity and an unprovided prerequisite separate. They call for different decisions.
If Jamie produces a handover that omits a missed collection, investigate the checking method before widening the practice. If the handovers are accepted but take more total effort, the operating commitment may have been met while the improvement claim remains unsupported. If Claire has not yet reached a monthly cycle, her result is still awaiting observation.
Bring the workflow owner into the next decision:
- Inspect the completed work. Include important exceptions, corrections and the recipient's view, rather than selecting only the most attractive drafts.
- Check the conditions. Establish which records, permissions, support and task opportunities were actually available.
- Choose what changes. Continue the bounded commitment, repair a specific problem, revise the scope or stop an unsuitable method. Name the owner and next review.
Keep a material goal revision visible. If the company removes access or the task changes, record the earlier commitment and why the new one differs. Conversely, “the conditions changed” should lead to evidence and a decision, not an indefinite exemption from reviewing the work.
Use the AI value framework for the separate question of whether accepted work improves complete effort or business results. If payment depends on the goal, the usage-based bonus guide covers influence, dependencies and reward assessment.
The next step is to bring one proposed objective to the person doing the task and the colleague receiving it. Rewrite it until both can explain the expected work, its acceptance condition and what they will do when the conditions do not hold.
Questions about setting AI adoption goals
Can a company expect employees to use AI every day?
A company can establish expectations for an approved working method, but a daily-use goal should have a reason grounded in the role's actual work. A coordinator preparing suitable handovers each shift has different opportunities from someone producing a monthly review. Define the task, quality checks and treatment of exceptions before choosing a frequency. The work-cycle comparison shows how to make that distinction without dropping accountability.
What percentage of work should use AI?
There is no universal percentage established here for work that should use AI. First identify the relevant tasks, permitted inputs, required quality and available support. If you choose a coverage target, explain which opportunities belong in it and keep accepted results separate from attempts. Review a bounded sample of ordinary work before raising the target; more AI involvement is not itself evidence of a better result.
How do you set a goal for work that happens monthly?
Set the goal around the next suitable reporting cycle and allow time for source preparation, checking and acceptance. A practice attempt before that cycle can establish learning, but not the performance of a monthly operation. Record a cycle that has not occurred as awaiting observation, then review the completed work. For repeat-use reporting, give people a comparable opportunity to return.
Can choosing not to use AI meet an adoption goal?
A reasoned decision not to use AI can be evidence of the judgment a learning or evaluation goal requires. It does not automatically complete an operating or improvement commitment. Record why the task, input or result was unsuitable, retain the permitted existing route and ask the workflow owner to continue, change or stop the proposed method. Agree this treatment when writing the goal, so employees do not have to conceal a useful finding.
