Register for the upcoming Ai Governance cohort starting October 8.
I know the appeal of handing something over and trusting that it will keep moving.
Running a small business, raising children, and keeping several projects underway, I have plenty of work that lives between conversations. The unanswered question. The document waiting for one more detail. The follow-up I meant to send.
So when OpenAI describes Dots as an always-on agent that can carry work forward between conversations, I understand the attraction.
According to OpenAI’s documentation, a dot can work across connected tools, retain context, and bring back results or decisions that need your attention. Its cloud computer can keep working when yours is off.
That is a useful possibility. It also asks us to become much more precise about what we mean when we say, “Take care of this.”
What did we actually hand over?
Imagine a foundation asking an AI agent to keep grantee follow-ups moving.
The assignment sounds ordinary. Staff spend hours tracking reports, checking correspondence, and figuring out what is still needed. Some help could be welcome.
But follow the work.
The agent reads a report and checks it against the grant requirements. It finds an unanswered question. It reviews earlier correspondence to see whether someone already addressed it. It drafts a message.
So far, we might be comfortable.
Then it decides which organizations need attention first. Perhaps it uses late reports, missing information, or the tone of correspondence to make that judgment. It labels one organization a concern and sends a more urgent message.
We have moved through several different kinds of work: finding information, interpreting it, setting priorities, assessing an organization, and communicating on the foundation’s behalf.
Each step deserves its own decision about authority.
“Keep follow-ups moving” did not tell us which of those decisions we intended to delegate.
A connected tool is a form of access
Connecting an agent to an inbox or grantmaking system gives it a way into the work. We still have to establish what information it may use and for what purpose.
May it read every grantee’s correspondence? Combine reports with site visit notes? Use information from one relationship to inform another?
In Who Gets to Shape What Comes Next?, I argued that meaningful participation includes the authority to set boundaries, challenge decisions, and decide that a particular use should not proceed.
Delegating work to an AI agent makes those questions immediate. What have we authorized it to do?
One way to examine the assignment is to follow the information through the work:
Collect → Combine → Infer → Decide → Act
At each transition, something can change: the sensitivity of the information, the judgment being made, or the authority needed to proceed. An agent may move through several of these steps within a single assignment. We need to establish where it can continue and where it must return to a person.
Delegation makes the transitions in that chain especially important. An agent may need to make intermediate choices to complete an assignment. Some are routine. Others can affect how an organization is understood or treated.
A late report might justify a reminder. It does not, by itself, establish that a grantee is poorly managed.
A difficult exchange might warrant a conversation. It does not establish that a leader is unreliable.
We need to decide which interpretations are appropriate, what evidence supports them, and what consequences they may carry.
Make the assignment specific
This is where a Practice Guideline becomes useful.
For this workflow, the foundation might authorize the agent to check reporting dates, identify missing items against agreed requirements, and prepare draft reminders.
A staff member would review the draft and the evidence behind it before anything goes out. Assessments of organizational performance would require a separate process. Conflicting records or unclear requirements would send the work back to the person responsible.
That could fit into a short instruction:
Check reporting deadlines and agreed requirements. Prepare draft email reminders with links to the relevant records. Bring unclear or conflicting information to the assigned program officer. Do not assign grantee risk ratings, change requirements, or send messages without review and approval.
The exact boundaries will differ by organization. The point is to establish what the agent may do, what remains human, who is accountable, and when the work must stop.
Oversight needs somewhere to happen
We often say a person will remain “in the loop.”
Where, exactly?
If a person sees a message only after it has been sent, they can respond to the consequence. They cannot review the decision before it occurs.
For consequential actions, meaningful control may require an actual pause. The person reviewing needs to see what is proposed, what evidence informed it, and what remains uncertain. They also need the authority and time to change the outcome.
Machines can move faster than people can evaluate. That makes the design of the handoff part of governance.
We also need to know how to stop the work. OpenAI’s documentation notes that pausing a dot’s main task does not necessarily stop delegated or scheduled tasks. The practical lesson extends beyond this product: understand what continues running and how to end the whole assignment.
Does this improve the work?
Efficiency is the floor.
An agent may prepare reminders quickly. We still need to account for the time spent checking its interpretations, correcting errors, maintaining access, and supervising its actions.
We should also ask what the grantee experiences. Does the process make expectations clearer? Does it recognize an extension already agreed by email? Does it leave room for someone to explain what is happening?
More automated follow-up could make a burdensome reporting process even more burdensome. Before accelerating it, examine whether the requirement serves a useful purpose.
This is Align → Steward → Enable in practice.
Align around the outcome and whether the workflow needs to change. Steward the information, judgment, relationships, and authority involved. Enable a small test with clear boundaries and someone responsible for learning from it.
Start with preparing work for review. Examine the results before expanding authority. Let evidence determine the next step.
I want help with the work that keeps accumulating between conversations. I also want to know what I have handed over.
Before asking AI to keep working without us, we need to establish what it may access, what it may decide, what it may do, and when it must return.
Delegation is a governance decision. We should make it deliberately.
Sources and further reading
OpenAI, “Meet dots.” Product capabilities, connections, permissions, and controls. Availability and features are subject to rollout.
This grantee follow-up example illustrates governance choices. It is not a description of a tested Dots implementation.