Will OpenAI Dots Finally Join All the Dots?

For years, we have been learning how to ask AI better questions.

Then we started giving AI tasks.

Write this. Analyse that. Research this company. Summarise these documents. Build this presentation. Check these numbers. Draft this campaign.

OpenAI's new Dots makes me think we are approaching a much more important transition.

We may be moving from giving AI tasks to giving AI responsibility.

That distinction is much bigger than it sounds.

A chatbot waits.

An assistant responds.

An agent acts.

But an always-on agentic AI system can potentially do something fundamentally different: retain an objective over time, observe what changes, decide what needs attention, carry out permitted work and return to the human when judgment is required.

That begins to change the relationship between us and software.

The problem was never just completing the task

AI has become extraordinarily capable at individual tasks.

It can research markets, analyse documents, write code, prepare reports, search the web, build presentations and increasingly interact with other software through screen agents.

But most meaningful work is not one task.

It is a chain of tasks.

Take something as ordinary as pursuing a business opportunity.

Someone sends an email.

You research the company.

You schedule a meeting.

You prepare notes.

You attend the meeting.

Someone promises to send a document.

Three days later, you realise it never arrived.

You follow up.

The document comes.

You review it.

You update your team.

Someone has another question.

You schedule another discussion.

A proposal has to be prepared.

Then revised.

Then followed up again.

The actual work is scattered across email, calendars, documents, messaging platforms, browsers, spreadsheets and dozens of tiny acts of human memory.

AI can already help with almost every individual step.

What it has struggled to do is connect the steps into one continuous objective.

That is why Dots interests me.

The bigger opportunity is not necessarily an AI that performs one task dramatically better.

It is an AI that remembers why all those tasks are happening in the first place.

The unit of AI may be changing

For most of the generative AI era, the fundamental unit has been the prompt.

You prompt.

The machine responds.

Then tools and AI agents moved us toward the task.

You define something that needs to be accomplished, and AI performs multiple steps to get there.

Persistent agents may take us one level higher.

The new unit could become the objective, a shift closely tied to how objective-driven AI is transforming intelligence.

Instead of saying:

Research these ten companies.

You might say:

Help me identify the most interesting companies in this sector and keep me informed when something materially changes.

Instead of:

Draft a follow-up email.

It becomes:

Make sure this opportunity does not quietly die because somebody forgot to follow up.

Instead of:

Prepare my meeting notes.

It becomes:

Keep this project moving and bring me in when a decision requires my judgment.

That is not simply faster software.

It is a different operating model.

Software has always depended on human attention

Think about most software we use today.

Your inbox contains information.

Your CRM contains information.

Your calendar contains information.

Your documents contain information.

Your project-management tool contains information.

But somebody still has to repeatedly look at all of them and ask:

What changed?

What matters?

What have I forgotten?

What should happen next?

Humans have effectively been the integration layer between our software systems.

We carry the context from one application to another.

We remember that an email is connected to a meeting that is connected to a proposal that is connected to a deadline.

And we regularly fail.

Not because people are incapable, but because modern work has created an absurd volume of small things that need remembering.

This may be where persistent agents become truly interesting.

The agent does not merely perform work.

It potentially becomes part of the continuity of the work.

The biggest productivity gain may come from very small things

There is a tendency to judge AI by spectacular demonstrations.

Can it build an app?

Can it write a book?

Can it design a building?

Can it replace an analyst?

Those questions make good headlines.

But I suspect a huge part of the economic value of agentic AI will come from far less dramatic things.

Remembering that someone did not reply.

Noticing that a number changed.

Checking whether a deadline slipped.

Reading the new version of a document.

Updating a status.

Moving information from one system to another.

Preparing context before a meeting.

Following up after the meeting.

Flagging that two pieces of information contradict one another.

Individually, none of these activities is revolutionary.

Collectively, they consume an extraordinary amount of human attention.

The productivity breakthrough may therefore not be AI becoming 10 times better at completing an isolated task.

It may be AI connecting thousands of small tasks around a larger objective.

That is a much more profound proposition.

It could also create something organisations badly need: memory

Companies forget constantly.

People leave.

Teams change.

Strategies evolve.

Documents get buried.

Important conversations disappear inside inboxes.

Six months later, someone asks:

Why did we decide this?

And nobody quite remembers.

A persistent agent that has been appropriately connected to a project could eventually create something resembling institutional memory built on knowledge graphs.

Not merely a database.

Context.

Why a particular decision was taken.

Which alternatives were considered.

What customers said.

What failed.

Who was supposed to do what.

Which assumptions subsequently changed.

Imagine the value of an AI agent that had been working alongside a project for three years and could explain not only what happened, but how one decision led to another.

That begins to feel considerably more important than another productivity application.

But responsibility changes the risk too

There is an obvious other side to this, and it sharpens the old debate over whether AI is a blessing or a beast.

The more responsibility we give AI, the more carefully we need to think about authority.

There is a vast difference between allowing AI to draft an email and allowing it to send one.

Between analysing a transaction and approving one.

Between suggesting a change and implementing it.

Between reading a document and deleting it.

Persistent agents therefore make permissions, approvals, auditability and human oversight considerably more important.

The right model cannot simply be:

Give AI access to everything and hope it behaves intelligently.

It has to be:

Give AI enough authority to accomplish its objective, but make the boundaries explicit.

In some ways, this begins to resemble how organisations already manage humans.

People have roles.

They have access.

They have approval limits.

They have escalation paths.

They are trusted to make some decisions and required to seek approval for others.

We may eventually need to design equivalent structures for AI agents.

Prompt engineering may become less important than delegation

One of the defining ideas of the early generative AI era was prompt engineering.

People learned frameworks for telling machines exactly what to do, from casual tips to a strategic guide to prompting ChatGPT.

Give it context.

Specify the output.

Define the role.

List the steps.

Provide examples.

That remains useful.

But increasingly capable agents may shift the premium toward a different skill.

Delegation.

Delegation is harder than prompting.

A good manager does not tell someone every keystroke required to accomplish a goal.

A good manager defines the outcome, the constraints and the decision rights.

What are we trying to achieve?

How will we know when we have succeeded?

What may you decide independently?

What requires approval?

What are the boundaries?

When should you escalate?

Those may eventually become some of the most important questions we ask when deploying AI.

The better AI gets, the more important clear human thinking becomes.

The perfect prompt may matter less than the right objective

There is an interesting paradox here.

We have spent several years trying to become better at telling AI exactly what to do.

But the more capable AI becomes, the less useful it may be to prescribe every step.

Instead, we may need to become much better at defining what actually matters.

Imagine telling an autonomous system:

Increase sales.

That is not a good objective.

At what cost?

Among which customers?

Over what period?

Can it offer discounts?

Can it contact customers automatically?

What brand standards must it follow?

What risks are unacceptable?

Should profitability matter more than volume?

Human organisations already discover this problem constantly.

A poorly defined KPI creates poor behaviour even when the people pursuing it are intelligent.

AI will not magically solve that problem.

It could amplify it.

The ability to define the right objective may become more valuable than the ability to write the perfect prompt.

We may eventually have AI organisational charts

Follow this idea far enough and something else becomes possible.

Companies may eventually operate not with one AI assistant, but with many persistent agents.

A research agent.

A sales-intelligence agent.

A customer-support agent.

A recruitment agent.

A finance-monitoring agent.

A competitive-intelligence agent.

A project agent.

Perhaps some of these agents will coordinate with other agents, which is exactly where agent orchestration engineering comes in.

Perhaps each employee will manage several of them.

Perhaps organisations will create permissions, responsibilities and reporting structures specifically for their AI workforce.

That sounds futuristic.

But so did having a conversation with a machine a few years ago.

The interesting part is that the organisational question may become almost as important as the technological one.

It will not simply be:

Which AI model should we use?

It may become:

Which responsibilities should belong to humans, which should belong to agents, and how should the two work together?

The human role moves upward

I do not think the most interesting interpretation of this transition is that humans disappear, and I have argued why humans are irreplaceable in an AI-driven world before.

I think our role moves.

If AI increasingly handles monitoring, information retrieval, routine coordination and repetitive execution, humans can spend relatively more time on activities where judgment matters.

Choosing the objective.

Understanding ambiguity.

Building relationships.

Negotiating.

Making trade-offs.

Exercising taste.

Taking responsibility.

Deciding what should not be done.

Those are very different activities from moving information between systems and remembering to send the third follow-up email.

And perhaps that is ultimately the promise of agents.

Not merely doing our work faster.

But allowing us to spend less of our lives managing the machinery surrounding the work.

From prompting to delegation

The first era of generative AI taught us how to talk to machines.

Then we gave machines tools.

Now we are beginning to give them something closer to ongoing responsibility.

That is a meaningful transition.

Because when an AI can retain an objective, observe what changes, use permitted applications, perform recurring work and return when human judgment is required, it is no longer simply waiting inside a chat window.

It becomes part of the operating system around the objective.

We spent the first few years of generative AI learning how to prompt machines.

The next few years may be about learning how to delegate to them.

And perhaps the most valuable skill in that world will not be asking AI the perfect question.

It will be knowing exactly what outcome we want, what authority we are prepared to delegate and where human judgment must remain.

OpenAI calls them Dots.

The much bigger question is whether this is the point at which AI finally begins to join all the dots.