OpenAI's dots: what leaders need to define before delegating to AI

Delegating a task to AI requires deciding how far it can go. Who checks the data and approves sending? If the proposal includes a wrong commercial term, the company needs to know who answers for it.

Presented by OpenAI on September 29, 2026, dots can work across conversations, with a cloud computer and authorized apps. An example the company released shows a dot that identified a forgotten charge, prepared the invoice and sent it after the user's approval. Read the official announcement

In this case, the user's approval separated preparing the invoice from sending it. In the company, this point needs to be defined before the work starts.

When choosing a task, look for the cause of the delay. Information may be missing, there may be too much rework or a decision with no owner. Each situation calls for a different adjustment.

Consider preparing sales proposals as a test hypothesis. If salespeople use different price lists and the service terms are out of date, the company first needs to organize these references.

With inconsistent sources, the agent can produce a proposal that looks complete but requires many corrections. You need to know which document is valid and who can confirm a doubt before delegating its preparation.

Define the documents the agent can consult, the material it should prepare and the moment it needs to ask for approval. The team should know who will receive this request and what that person will check.

In the proposal test, for example, the agent could consult a selection of documents and prepare the draft. Discounts, deadlines and sending to the client would be subject to validation by the sales owner.

Having access to a price list doesn't authorize the agent to negotiate a discount. This distinction should appear in the request and in the permissions, respecting the tool's rules.

OpenAI says dots follow action rules and safety checks, with situations that require the user's confirmation or direct participation. The company acknowledges that errors are still possible. These limits need to be part of planning the test. Understand the safeguards OpenAI describes

Pick a small task, with known sources and limited consequences if it needs to be redone. Name someone to follow the test and decide whether the deliverable can be used.

Before starting, record how long the activity takes today, how many reviews it usually requires and which errors repeat. That will be the reference for evaluating what changed.

During the test, also count the time spent on guidance and checking. Note the corrections and the situations that required intervention. Getting the draft faster helps little if the team needs to spend more time getting it ready.

Whoever reviews needs to know the approval criteria and be able to stop the process when information is missing. In the sales proposal, that includes checking whether the values come from the current price list and whether the deadlines were confirmed by the responsible area.

Record the problems so you can fix their origin. A commercial term interpreted the wrong way may require a more precise instruction, an updated document or a change in the delegated task.

At the end of the test, decide based on the work observed: expand use, adjust and test again or keep the activity with the team. Consider both the time spent and the errors and risks that remained.