OpenAI gave companies an AI control panel

Welcome back.

OpenAI released something yesterday that will probably receive less attention than a new model.

It may matter more to companies trying to make AI useful.

The new Admin plugin for ChatGPT Work and Codex lets an administrator ask questions about workspace activity, manage access, change usage limits, review spending requests, and complete routine actions from a conversation.

In plain language, it gives companies a control panel for AI.

That is a sign of where enterprise adoption is heading. AI is no longer only a tool an employee opens when they need help. It is becoming a managed layer of work, with users, permissions, budgets, policies, and recurring processes.

Companies need that layer. But they should be careful about what they conclude from it.

An administrator can see that a team is using AI every day. They can see which models people can access, how many credits they consume, and whether spending is rising.

None of that proves the work improved.

My view is that the next failure in AI adoption will not come from having too little usage data. It will come from mistaking usage data for business evidence.

The control panel can tell you whether AI is being used responsibly. Your operating system still needs to tell you whether AI is worth using at all.

That distinction matters now because the easiest numbers to collect tend to become the numbers leaders manage. If the dashboard foregrounds seats, credits, and active users, those measures can quietly become the definition of success. A useful control panel should begin a conversation about the work. It should never be allowed to end it.

AI has become an operations job

The Admin plugin announced by OpenAI can review activity and credit usage across ChatGPT Work and Codex. It can add or remove members, update groups, diagnose access problems, control which models people can use, and approve or deny requests to increase spending limits.

It can also automate recurring tasks. A request for more usage can be sent to Slack or Microsoft Teams for approval. Access can be granted automatically when a request meets predefined criteria, while exceptions are sent to a person for review.

This sounds like administrative plumbing. That is exactly why it matters.

Every important technology eventually develops an operating layer around it. Someone needs to decide who receives access, what they can do, what the company will pay for, and what happens when something goes wrong.

AI arrived inside many companies before that layer existed. Employees created accounts, teams bought separate tools, and useful experiments spread through informal recommendations. That helped people move quickly, but it also created fragmented access, unclear spending, inconsistent controls, and very little visibility into what was happening.

The new plugin is one response to that problem. It makes the administrative work easier and gives companies a clearer way to apply existing permissions and approval rules.

But there is an important distinction.

AI administration controls the system. AI operations improves the work.

Administration asks whether the right people have access and whether the company is staying within budget. Operations asks which workflows are changing, whether quality is holding, and whether the organization is creating more value.

Most companies need both. If they build only the administrative layer, they may end up governing a large amount of activity that was never connected to a meaningful outcome.

This changes the job for whoever leads AI internally. They cannot remain only a buyer, trainer, or security gatekeeper. They need enough authority to connect access decisions to workflow decisions and enough curiosity to ask why one team creates value with AI while another creates only more usage.

A dashboard can make the wrong thing look successful

Imagine a thirty person sales team using AI to research accounts, prepare calls, draft follow ups, and update the customer system.

An administrator opens the dashboard and sees encouraging signs:

• Twenty seven people used AI this week

• Total usage increased by 40%

• More employees requested access to the strongest model

• Nobody exceeded the approved spending limit

From an administration perspective, the rollout looks healthy. Access is controlled, usage is growing, and spending is visible.

Now ask the operating questions:

• Are representatives preparing for calls faster?

• Are more qualified prospects replying?

• Is the customer system becoming more accurate or filling with generated noise?

• Are managers spending more time correcting weak research?

• Is the cost per qualified meeting improving?

The answers could tell a very different story.

The team might be using AI constantly because leadership asked them to. Representatives might save ten minutes drafting an email, then spend fifteen minutes correcting account details. Managers might see more activity while conversion stays flat. The company could be operating AI safely and still operating it badly.

OpenAI reports that one of its own Slack based IT agents resolved about 45% of ticket volume at the time measured. It also says operational dashboards helped its IT team eliminate a backlog while support volume roughly doubled. Those are company reported results, so they should not be treated as a universal benchmark. But the shape of the evidence is useful.

The meaningful claims are not that employees opened the tool or generated more tokens. They are that tickets were resolved, a backlog disappeared, and the team handled more demand.

That is the level at which your AI program should be judged.

There is a practical benefit to being strict about this. Outcome measures make weak workflows easier to stop without turning the decision into a debate about enthusiasm. They also make strong workflows easier to fund because leaders can see the result, the cost, and the risk in the same picture. Evidence creates a better conversation than advocacy.

Every AI dashboard needs a second dashboard

The first dashboard should measure control.

It answers questions such as:

• Who has access?

• Which systems and models can they use?

• What is the company spending?

• Which requests require approval?

• Where are policies being ignored?

This protects the company. It makes access visible, applies rules consistently, and gives someone responsibility for the system.

The second dashboard should measure work.

It answers different questions:

• Which workflow changed?

• How much time did it save from beginning to end?

• Did quality improve or decline?

• What new errors appeared?

• Did the change reduce cost, increase speed, improve quality, or support growth?

This protects the investment. It prevents leaders from confusing widespread activity with actual progress.

The two dashboards should connect. If one team is consuming far more AI than another, the business evidence should help explain why. Perhaps that team found a valuable workflow worth expanding. Perhaps it is paying frontier model prices for routine work. Perhaps it created a process that generates more review than value.

Usage data identifies where to look. Outcome data tells you what you are looking at.

This is also why the owner of AI adoption cannot be responsible only for licenses and permissions. The role needs a line of sight into real work. That person must be able to sit with a team, understand the workflow, define the expected outcome, and decide whether the evidence justifies further investment.

Without that connection, AI administration becomes a cleaner way to manage the same old problem: lots of tools, lots of activity, and no reliable answer to what improved.

The second dashboard does not need to become a giant reporting project. Start with one baseline and one quality measure for each important workflow. The goal is not perfect attribution. It is enough visibility to distinguish a useful change in work from a popular new habit.

Use the new controls to ask better questions

OpenAI’s new plugin is useful. Companies should want simpler ways to understand access, apply permissions, control spending, and automate routine administration.

But better controls should create better decisions, not simply more controlled usage.

This week, take one AI usage report and add four columns beside it:

• Workflow: What complete piece of work is this usage supporting?

• Outcome: What should become faster, cheaper, better, or more valuable?

• Review: Who checks the work and what failure are they looking for?

• Decision: What evidence would make us expand, change, or stop this workflow?

If those columns are empty, the company does not yet have an adoption problem. It has a definition problem.

Do not solve that by encouraging more people to experiment. Choose one consequential workflow, establish a baseline, and connect the administrative data to the result you expect.

For example, if the finance team uses AI to prepare a weekly cash report, they track more than how often they open the tool. Measure how long the complete report takes, how many corrections are required, whether important exceptions are caught, and how quickly leaders receive an accurate view.

Then the control layer becomes useful. You can see who needs access, what the workflow costs, which permissions it requires, and whether the result supports further investment.

AI is becoming easier to administer. That is progress.

It also creates a useful moment for leaders to reset expectations. You can tell teams that access will expand when a workflow has a clear owner, a defined result, and a review rule. That makes governance feel less like an arbitrary restriction and more like a shared standard for doing valuable work.

The companies that benefit most will be the ones that make it just as easy to see whether the work is improving.

Haroon

If you already have an AI usage dashboard, reply and tell me what business outcome sits beside it.