Skip to main content

Automation and Agents

How to Diagnose AI Adoption with OpenAI Admin Analytics

Turn activity patterns into a specific support decision with a team owner and a real work sample.

A work sample sits between an owner’s open hand and a separate support choice beneath the OpenAI publisher mark.
On this page
  1. 1. Establish what the chart actually covers
  2. 2. Turn a category into a recognizable workflow
  3. 3. Investigate quiet usage without assigning blame
  4. 4. Inspect high consumption before calling it success or waste
  5. 5. Choose training that changes a specific result
  6. 6. Keep engineering activity connected to engineering outcomes
  7. 7. Automate reporting after its meaning is stable
  8. 8. Leave the review with a decision and an owner

A busy AI dashboard can tell you where to look. It cannot tell you which team needs another training session, which workflow deserves more capacity, or whether a quiet user is doing valuable work elsewhere. Those decisions need a conversation about the actual work.

OpenAI's September 16, 2026 business-value walkthrough connects ChatGPT Work and Codex analytics with questions about adoption, spending and results. The useful opportunity is to bring administrators and business owners into the same discussion. An admin can spot a pattern; the owner can explain the task behind it.

The practical approach is to choose one workflow, check what the available measures mean, inspect a representative output with its owner, and try a specific improvement. Start small enough that you can tell whether the change helped. The steps below are proposals informed by the cited sources.

Key takeaways

  • Keep reporting scope and metric units intact before comparing teams.
  • Treat low activity or high consumption as a question to investigate, not a staff performance verdict.
  • Choose support that addresses a concrete obstacle in a real deliverable.
  • Review completed work and quality before expanding a workflow or reducing access.

1. Establish what the chart actually covers

Before interpreting a trend, write down which workspace, period, population and measure it represents. That small step protects the whole discussion from a deceptively simple mistake: comparing numbers that answer different questions.

OpenAI's current Usage Insights documentation distinguishes messages, credits and active users. It also says sampled counts must not be extrapolated into workspace totals. A credit share is not a share of finished work.

Suppose, hypothetically, a department's activity falls between two reporting periods. The meaningful comparison depends on who was included and what was happening. A quieter week during an annual planning cycle differs from a decline during steady recurring work. If group membership changes, the chart may be comparing different people. Mark a missing period as missing. Do not read a flat line as proof of no use.

Choose the dates and team that match your question. If you are checking whether training helped a particular team complete a weekly planning task, use that team's work and a relevant interval. A company-wide trend can supply context, but it should not replace the narrower evidence. The wider denominator may hide the change you care about.

A team’s work sits within a chosen reporting period, separate from a broader company view. View image detail

Choose Actual size to read the graphic closely.

Describe missing data clearly. “No recorded activity in this view” is more precise than “the team does not use AI.” The first statement leaves room to check coverage. The second already assigns a cause. Ask the owner whether this view includes the relevant surface and whether the work happened during the chosen dates.

Compare teams only when their work makes the comparison useful. Comparing a research-heavy planning role with a frequent support task can reward frequency rather than usefulness. Choose comparisons that share a meaningful workload, then keep any difference in task complexity visible. If that is impossible, describe the teams separately.

The result of this first check should be a short statement you can reuse: which team, what time window, which activity, and what the number measures. It is the caption your report needs before its first chart.

2. Turn a category into a recognizable workflow

A task category becomes useful when someone can connect it to a deliverable. “Sales activity” is too broad to tell you whether the team needs better account research, follow-up messages or proposal preparation. Ask which recurring job the team recognizes in the pattern.

The September walkthrough describes task classification as grouping a sample of messages into use cases and tasks. Classification offers a starting point for investigation. It does not certify what was delivered to a customer.

Consider a hypothetical agency preparing proposals. A category might point toward writing or research, but the owner cares about a specific result: a proposal that reflects the client's brief, uses current pricing and can be reviewed without rebuilding it. That gives the diagnosis a concrete object. You can inspect a proposal and ask where time or attention went.

Separate the stages of that job. Gathering inputs, creating the first draft, checking facts, obtaining approval and sending the proposal are different steps. AI may improve one while the slowest step remains elsewhere. If the missing client brief causes repeated rewrites, teaching a more elaborate prompt may simply produce a better-looking draft of the wrong proposal.

A proposal’s stages remain distinct, with an unresolved client brief before the draft and review steps. View image detail

Choose Actual size to read the graphic closely.

Ask the team to show a representative example they are authorized to discuss. The point is to understand the workflow, not collect everyone's conversations. Use the business record that normally shows completion, and respect the team's existing permissions when reviewing it. Do not infer that an aggregate analytics view makes every underlying document available to an admin.

Then give the workflow a plain name, an owner and a result check. “Prepare the weekly account brief, owned by the sales manager, checked against the agreed account facts” is usable. “Increase advanced feature saturation” tells people very little about what to improve.

This also creates a useful boundary for sharing. A well-understood workflow can be taught to a neighboring team. A broad category is an invitation to ask better questions, not yet a reusable operating method.

3. Investigate quiet usage without assigning blame

Low recorded use can point toward several different decisions. The team may need access, a suitable task, better source material, help judging results, or no change at all. The chart alone cannot choose between those explanations.

Begin with the team's purpose. Ask which work they expected the tool to help with and whether that work occurred during the reporting period. A specialist with infrequent but substantial assignments should not be judged against someone handling a daily queue. Frequency is a feature of the job as well as a feature of tool use.

Next, ask about the last attempt. What input did the person have? What result did they need? Where did they stop? Someone who did not know which file to provide needs a different intervention from someone who produced a draft but could not check its accuracy. A person who already has a faster existing method may need a fair comparison rather than more encouragement.

Use neutral language. “What made this task difficult?” invites information. “Why aren't you adopting AI?” makes the answer a defense of the employee. If the goal is useful work, the discussion should make it easier to surface unsuitable tasks and disappointing results. Those are valuable findings.

Two people discuss a difficult task without assigning an adoption score to either person. View image detail

Choose Actual size to read the graphic closely.

A hypothetical operations team might say its weekly summary is not the problem. Collecting consistent updates from several owners takes most of the effort. A drafting clinic would miss that obstacle. A clearer intake template or a designated source owner might make the eventual AI assistance more useful, without requiring higher consumption as the success measure.

Independent research also cautions against a single adoption story. The 2025 Generative AI at Work study examined a staggered rollout to customer-support agents at one firm and found different effects across workers. That is context for asking about different needs, not a forecast for your team or a test of this Admin Console.

End the conversation with a support hypothesis you can check. For example: missing source context is making the draft hard to trust. Change that input, inspect the next result, and see whether the obstacle remains. Do not turn a provisional explanation into a permanent employee label.

4. Inspect high consumption before calling it success or waste

A workflow using substantial capacity deserves attention, but its size does not supply a verdict. It may be doing demanding work, generating useful alternatives, retrying a poor approach, or responding to a larger workload. You need the work sample and its owner to distinguish those possibilities.

Look first at what changed. Did the team take on a more complex deliverable? Did the input become longer? Did the chosen model or settings change? Did a skill gain broader use? These questions can explain a trend without presuming that employees are careless or exceptionally productive.

The current Insights guide says plugin and skill credit allocations can overlap. Adding those views together can double count consumption. Keep each measure's meaning in the report instead of building an impressive total from incompatible parts.

Imagine a hypothetical team creating several design directions for one client. Producing alternatives may be a deliberate part of the job. The relevant question is whether the alternatives help the client choose and whether the team can finish within its available resources. Cutting every second attempt would confuse exploration with failure.

A client considers several design directions within a bounded workspace. View image detail

Choose Actual size to read the graphic closely.

A different team might keep regenerating a document because its source table is incomplete. There, higher consumption may be a reason to repair the input. Asking for fewer messages without examining that table would move the problem rather than solve it.

Choose a meaningful comparison if the team wants to test a cheaper or faster setup. Keep the task, input and result criteria similar. Include the effort required to review and correct the output. A shorter first response is not enough if someone must reconstruct the missing details afterward. This is a proposed comparison, not a recommendation that any particular model will win.

Avoid equating capacity use with an invoice. The financial effect belongs to the organization's agreement and billing settings. Our discussion of AI business value beyond vendor formulas addresses the separate accounting question. Here, the support decision is whether the workflow's setup fits the work it is trying to do.

5. Choose training that changes a specific result

Useful enablement starts with a task someone wants to complete. Once you understand its obstacle, teach the smallest change that helps them produce and check that result. Training does not need to become a tour of every feature.

For a hypothetical account-planning workflow, the obstacle might be inconsistent source material. A useful session could show how to assemble the relevant account records and mark missing information before drafting. If the obstacle is review, the session could instead teach how to compare claims with those records and keep unresolved questions visible. Those are separate lessons.

Ask participants to bring a permitted example or use a clearly staged one. Explain what good work looks like before showing a prompt. That gives people a way to judge the output, rather than encouraging them to copy wording and hope the result is appropriate for their own case.

A shared skill needs the same care as a shared procedure. Name the task it supports, the inputs it expects and what the user must check afterward. Assign someone to check the shared skill when its inputs, examples or expected output change.

A maintained task procedure connects expected inputs with a separate result review. View image detail

Choose Actual size to read the graphic closely.

Do not treat using more tools as a curriculum objective. A simple workflow that produces a reliable deliverable may be preferable to an elaborate chain that is harder for the team to maintain. Teach an additional capability when it addresses the observed obstacle. Avoid turning the training program into a competition to invoke the most features.

Check a fresh piece of work after the session. Did the person understand the input? Could they identify a missing fact? Did the manager accept the result with less reconstruction? A follow-up conversation about one example is more useful than treating attendance as proof that the workflow improved.

Record what still needs help. Perhaps the session improved the draft, but approval remains delayed because nobody owns the final decision. That is useful progress with a remaining organizational problem. Keep both parts in the readout so the next intervention addresses the right step.

6. Keep engineering activity connected to engineering outcomes

Codex analytics can help engineering leaders choose where to investigate, but a count of contributions or reviews does not replace their normal checks of code. Use the measure as context for a bounded change in a repository the team understands.

OpenAI's announcement describes Outcomes measures for Codex contributions to merged commits and lines of code. The current Insights guide separately describes Code review measures, including reviewed pull requests and findings. A reviewed pull request is not necessarily merged or deployed.

That distinction changes the question you ask. If review activity rises, ask whether useful findings were acted on. If contributed code rises, ask what changes reached the intended destination and whether they behaved as expected. Do not combine these into a generic claim that engineering productivity rose.

For a hypothetical recurring bug fix, choose issues with a comparable problem and enough repository context to judge the fix. Ask the engineer to show how the failure was reproduced, what changed and what evidence supports the result. Include investigation, review and correction in the effort discussion. Time waiting for release is a separate question from time spent doing the work.

A recurring fix is examined through the reproduced problem, the change and its result evidence, separate from release waiting. View image detail

Choose Actual size to read the graphic closely.

For tests, examine what they catch. A larger test file or a greater number of assertions may still miss the relevant failure. For a refactor, inspect the affected behavior and call sites. These are useful ways to judge the deliverable; they are not invented dashboard fields.

Give failures room in the account. An abandoned attempt, a reverted change or extra review can reveal a context problem worth fixing. A report that records only successful merged work can miss the effort spent getting there. The team can supply that context from its normal engineering records.

The support decision should remain concrete. Better reproduction steps, clearer repository instructions or a narrower initial task may help more than a blanket push for additional use. Engineering owners should choose the relevant checks for their code, and the next review should examine whether the chosen change made the work easier to complete correctly.

7. Automate reporting after its meaning is stable

Recurring reports are worth automating when they answer a useful question consistently. Automating a misleading comparison only gives the mistake a reliable schedule.

OpenAI's current Analytics API documentation describes aggregated workspace reporting, distinct from raw audit logs. It directs implementers to the current Admin API reference for access, fields and time semantics. Confirm availability for the workspace before planning an integration.

Start by deciding who needs the report and what decision it supports. An admin investigating one team's support needs does not necessarily need the same level of detail as a leadership summary. Define the audience before joining product activity to internal records or distributing the result more widely.

For example, a weekly support report could show the period and recorded activity beside completed tickets from the team's own system. The two measures should remain separate. Write down what each one measures. Keep its source, unit, period and scope alongside the resulting chart. If a field or classification changes, the person maintaining the report should know which comparison needs another look. This is especially important when a trend crosses the start of a new reporting capability.

A report keeps activity and completed-ticket records separate, with a reference beside each source. View image detail

Choose Actual size to read the graphic closely.

Include a place for context that cannot come from the product metric. A task owner can explain a change in demand, a missing input or a planned trial. Do not pretend an automated narrative can establish those facts by reading a graph. The owner should check consequential interpretations before they are sent as findings.

Different administration views also have different boundaries. OpenAI's Workspace analytics Help article describes aggregated reporting rather than raw message or file content. Confirm which view supplies a measure before combining it with Work Insights. A limitation in one view should stay visible in the summary.

Test the recurring report against a small known period before relying on it. That means checking the actual reference and the organization's records, not assuming an example endpoint or a dashboard label supplies a stable integration contract. No API implementation has been tested for this article, so these are design choices to make with the person responsible for the reporting system.

8. Leave the review with a decision and an owner

A good adoption review ends with a change someone can make and a result someone can inspect. It should not end with a demand that every line on the chart go up.

Choose among several plausible actions: share a useful workflow, repair its inputs, improve review, investigate access, test a different setup, or leave it alone. Reduce scope when the team cannot yet explain what good completion looks like. Expanding a confusing workflow makes the confusion more expensive.

Put the decision in a short working note. For the hypothetical proposal team, it might say: the operations owner will improve the client-input template before the next proposal because missing scope details caused revision work. Name the actual change and the person who will make it. Specify the next piece of work to inspect and when the owner expects enough activity to judge it. A suitable review interval follows the workflow's cadence, not a universal thirty-day rule.

If access or seat allocation is under discussion, involve the responsible owner and check the actual agreement and supported administration path. A low-use chart does not establish that automatic removal is supported, that history will be preserved after a change, or that a specialist no longer needs access. Those are separate facts and decisions.

An owner reviews access beside an agreement and an administration path rather than removing a seat from an activity signal. View image detail

Choose Actual size to read the graphic closely.

Keep the record proportionate. A small team does not need a new governance program to learn whether a recurring task has become easier. A large organization may need a shared format so related decisions can be compared. Either way, the note should make the next action clearer rather than generate another reporting obligation.

When you review the next result, compare it with the original obstacle. Did the missing input improve? Was review easier? Did the output meet the team's criteria? If the answer remains uncertain, say what evidence is missing and decide whether another attempt is worth the effort. Uncertainty can be a sensible outcome of a small investigation.

The next useful step is simple: choose one task that matters, pair its activity pattern with its owner, and inspect the work together. That turns analytics into a way to help people deliver something useful, without pretending the chart has already done their thinking for them.

Checked for this article

Sources

  1. How to connect AI usage to business valueOpenAI
  2. Usage Insights for ChatGPT Work in CodexOpenAI
  3. Analytics APIOpenAI
  4. Generative AI at WorkQuarterly Journal of Economics
  5. Workspace analytics for ChatGPT Enterprise and EduOpenAI

Keep reading

All articles
  • Automation and Agents

    Codex Cloud Environment Access: A Practical Sharing Checklist

    Sharing a prepared environment is not the same as sharing a teammate's task. It can still make prepared files, credentials or connected resources available, so check the access path before rollout.

    14 min read

  • Automation and Agents

    Codex Cloud Environments: Is Reusable Setup Worth It?

    A reusable coding environment can remove repeated preparation, but it does not prove that work finishes faster. Here is a practical way to test the whole workflow.

    14 min read

  • Automation and Agents

    Can You Trust a ChatGPT Work Data Agent Dashboard?

    A practical review framework for checking Data agent evidence, calculations, connected-account access, published dashboard contents, audience, and downstream actions.

    17 min read