What ChatGPT Workspace Agents Can Do

OpenAI introduced workspace agents in ChatGPT on 22 April 2026.

The useful part is not that ChatGPT can do another impressive party trick. The useful part is that agents are becoming easier for teams to share, schedule, connect to work tools, and improve over time.

That matters.

It also does not mean a business has hired an assistant.

The difference is important, because "assistant" is a lovely word for hiding messy decisions. A human assistant notices context, asks awkward questions, learns politics, spots when a request is badly framed, and knows that "send this today" sometimes means "check whether this should be sent at all".

An AI agent does not have that kind of judgement by default. It has instructions, permissions, connected tools, memory, and a task.

That can be enough for useful work.

It is not enough for vague delegation.

Workspace agents make shared AI work easier

OpenAI says workspace agents can be built inside ChatGPT for repeatable team workflows. They can use connected apps, run on a schedule, work in ChatGPT or Slack, and be shared through a team directory. The examples include software request review, product feedback routing, weekly reporting, lead outreach, and vendor risk checks.

OpenAI's help docs add a practical detail that matters: agents can be previewed before publishing, scheduled, connected to tools, given files and skills, and controlled through permissions. Write actions, the kind that send, edit, post, or delete things, ask by default unless someone changes the approval settings.

That is the bit small businesses should pay attention to.

Not the launch language. Not the screenshots. The approval settings.

Agents are moving from "one person asks ChatGPT to do a thing" towards "the team has a reusable worker for a defined job". That is a real shift. It brings AI closer to the place work happens.

But it also makes the design of the job more important.

A useful agent needs a narrow job

The first mistake is treating the agent as a spare pair of hands.

"Help with sales" is not a job.

"Every weekday morning, look at new inbound enquiries, compare them against our good-fit criteria, draft a short summary, and queue a suggested first reply for review" is much closer.

The second version has edges. It has inputs, judgement criteria, output, frequency, and a review point.

That is where AI agents become useful.

They are good at work that is repeated often enough to justify the setup, structured enough to describe, and annoying enough that people keep putting it off.

For example, a small consultancy might receive ten new enquiries a week. Some are a good fit. Some are not. Some need a fast reply because they mention a live deadline. Some need a careful reply because the request is unclear. A workspace agent could read the enquiry, check it against a fit checklist, draft a summary, suggest a next step, and prepare a reply.

That is useful.

It should probably not send the reply automatically on day one.

The gain is not "we replaced a person". The gain is "the person starts from a prepared summary instead of a blank inbox and a mild sense of dread".

Very glamorous. Also exactly where the value tends to live.

Check risk, repetition and review

If you are considering an agent workflow, use three questions before buying or building anything.

Is the task repeated?

If it happens once a quarter, it may not deserve an agent yet. Use a checklist, a template, or a normal ChatGPT conversation.

Is the task bounded?

If nobody can describe what good looks like, the agent will not magically discover it. The business needs a clearer process before it needs automation.

Can the risky step be reviewed?

If the agent can draft, classify, summarise, research, sort, or prepare, review is easy. If it can send, delete, refund, approve, publish, or commit money, review needs to be explicit and hard to bypass.

That is the practical test: repetition, boundaries, review.

If all three are present, an agent may be worth testing.

If one is missing, fix that first.

Repeated jobs suit workspace agents

The best near-term use cases are not dramatic.

They are the jobs that already have a shape:

• weekly reports that pull from the same sources
• inbound enquiries that need the same first pass
• support themes that need turning into a product list
• meeting notes that need actions and follow-ups
• website leads that need routing by service, budget, and urgency
• supplier checks that need a consistent evidence pack
• project updates that need chasing before the meeting

These are not glamorous tasks. That is why they are good candidates.

In our own studio, the agents that stuck are the ones attached to a real recurring job: project triage, live software error sorting, report drafting, code review support, and shared memory capture. They work because they clock on for a specific task, not because we gave them a job title and hoped they would become useful.

That is a sentence only we can write because we have had the slightly odd experience of building AI tools that remind us what we forgot to tell the AI tools.

Agents fail without clear review rules

Workspace agents will still break in ordinary ways.

They will misunderstand messy inputs. They will surface old information if the connected source is stale. They will over-prioritise the wrong signal if the criteria are weak. They will produce a clean summary of a confused situation, which is better than nothing and still not the same as judgement.

They also introduce permission risk.

OpenAI's own guidance is clear that write actions need care, especially where an agent can send, edit, post, or delete. Their older ChatGPT agent safety notes also spend a lot of time on prompt injection, where malicious content tries to manipulate an agent while it works on the web.

That is not a theoretical problem for large companies only.

Small businesses have customer data, supplier emails, calendar access, payment systems, internal documents, and enough operational fragility to make one bad automated action painful.

The sensible answer is not "never use agents".

It is "do not give the first version the keys to the building".

Start with read-only access where possible. Let the agent prepare, classify, draft, and recommend. Put a person in front of anything that leaves the business, changes a record, moves money, or affects a customer.

Then watch what actually happens.

Workspace agents do not replace staff

Do not rewrite your whole operation because OpenAI added a better way to create shared agents.

Also, do not ignore it.

This is the direction of travel: AI moving from individual chat to reusable team workflows. For small businesses, that is more interesting than another clever demo because it starts to match how work actually happens. Work is shared. Work has rules. Work has review. Work happens in inboxes, calendars, files, project tools, websites, and the places people already have open.

The opportunity is not to hire a pretend assistant.

The opportunity is to give one narrow, repeated job a better first pass.

That is less exciting to say. It is much more likely to survive Tuesday morning.

If you are looking at workspace agents, start with the task your team already describes with a sigh. Then ask: can we define the input, the output, the review point, and the point where the agent must stop?

If yes, you may have a good first agent.

If no, you have found the real work.

That is useful too.

We build Agent Workflows around exactly that distinction: not "AI as an assistant", but AI attached to a real job, with limits, review, and a clear reason to exist.