What needs my attention now?
Consider a fictional example: sales wants better proposal drafts, support wants more reliable product answers, and a third team wants better meeting notes. You need to decide where to spend your team’s limited time. That starts with understanding what goes wrong today and the result each team needs.
We suggest walking through a recent task with each application owner. Which results reach the person who needs them, what needs reworking, and where is an application acting beyond its approval? Bring urgent problems to the responsible person immediately. You can then compare the remaining improvements.[S1] NIST also connects an AI inventory to risk priorities and the question of whether a system meets its intended purpose.[S2]
Real work reveals the benefit and the effort.
Use this sheet to review the applications on a comparable basis. Record specific examples and the period you examined. Missing information gives you something to investigate before deciding.
| My question for each application | What I need to record |
|---|---|
| What result does my team need? | The task, who uses its result and how the work would get done without this application. |
| How well does it work today? | Results used, errors and corrections from a normal work cycle. |
| What does it take to keep it running? | Fees and time spent on review, rework, maintenance and exceptions; mark estimates. |
| Who can decide and intervene? | The responsible person, permitted actions and who takes over if it fails. |
| What do other applications also depend on? | Shared sources, integrations, permissions or checks. |
| What do I need to decide next? | One improvement or investigation, its owner, the time needed and the expected result. |
In our meeting-notes example, nobody has recorded correction time. Measure it during normal use before deciding to expand the tool. The license fee covers only part of the effort; our operating-cost example shows how to include people’s time. Choose comparable periods: a monthly report and a daily tool have different opportunities to demonstrate value.
One shared data source can hold back two improvements.
Sales and support find the same outdated product details in our example. Both applications read the same old export. Have that source checked before starting two separate projects for better prompts or new models. AWS’s AI portfolio guidance likewise recommends considering dependencies between initiatives.[S3]
In this example, first agree how affected drafts will be checked or held back while the issue is investigated. Give the person maintaining the product export the correction task. Sales and support then test their own outputs. Those checks establish which errors are resolved and what work remains.
Plan the shared source work once and the two application tests separately. If the same rework is removed, count that saving only once too. The applications themselves can remain separate: a proposal needs different approvals from a support reply.
Which improvement should I take on next?
Once urgent problems and dependencies are understood, compare the expected benefit with the full effort of making and running each improvement. Include whether your team can use it and has time for the work. Microsoft names impact, feasibility and user desirability as criteria for agent ideas.[S4] Keep the basis for each assessment beside it.
Our fictional example leads to the following work plan:
| Decision | Next step | How I can check the result |
|---|---|---|
| Check the shared cause first | Contain affected drafts, investigate and correct the product export, then test both applications. | Both applications use the checked source version; each team reviews its new drafts. |
| Measure the missing effort | Record correction time and useful meeting notes during normal use. | Effort and results can be compared with the team’s previous way of working. |
| Defer the model change for now | Check which errors remain after the source review. A separately evidenced model problem may warrant an earlier test. | A specific remaining error gives the next test a purpose. |
Agree an owner and a review date for each action. If one person owns several tasks, their time also needs an order of priority. If the review supports retiring an application, our pause, repair and retirement guide helps you plan the handback of its work.
Basis for this decision guide
The example is fictional. The tables present our proposed approach, informed by public guidance, rather than measured customer results or a validated ranking. The right priority depends on your evidence, obligations and available people.
For your next review, choose two applications that compete for the same people’s time or depend on the same data. Fill in the sheet with their owners and agree the first action whose result you can check at the next meeting.
