OpenAI shipped workspace agents on April 22, 2026. They run inside ChatGPT for Business, Enterprise, and Edu plans. They stay on when you close your laptop. They connect to Slack, Google Drive, Salesforce, SharePoint, and custom data sources. And they replace the old custom GPTs model that never quite worked for teams.

This is the first time a major AI company has shipped persistent, shared, always-on agents directly inside its main product for business users. Not a developer API. Not a separate platform. Inside ChatGPT, where your people already are.

TL;DR

  • OpenAI launched workspace agents inside ChatGPT (April 22, 2026) for Business, Enterprise, and Edu subscribers
  • These are always-on agents that run in the cloud, connect to your existing tools, and handle recurring workflows without repeated prompting
  • Free through May 6, then credit-based pricing kicks in
  • 88% of organizations now use AI for at least one business function (Stanford HAI, 2026), but most are still using it as a search bar, not as a worker
  • The real question is not whether to adopt. It is which five workflows to hand off first

What Actually Shipped

Workspace agents are the successor to custom GPTs, rebuilt from scratch on OpenAI’s Codex system. The difference matters. Custom GPTs were single-turn assistants you had to prompt every time. Workspace agents are persistent. You describe a workflow once. The system maps the steps, connects the tools, and then runs on a schedule or trigger you set.

They operate in the cloud. When your marketing lead goes home at 6pm, the agent that monitors customer feedback channels in Slack keeps running. When your ops manager is in meetings all morning, the agent that pulls weekly metrics from three different dashboards and formats a summary keeps running.

According to OpenAI, more than 4 million developers were using Codex weekly by mid-April 2026, up from 3 million just two weeks prior. That adoption velocity tells you something about where the gravity is pulling.

The agents can be shared across an organization. One person builds it, the whole team uses it. Admins control permissions, restrict data access, and require approval before sensitive actions execute. Enterprise customers using Enterprise Key Management are excluded from the preview for now.

The Operational Shift That Matters

Here is where most coverage of this launch misses the point. The technology is interesting. The operational implications are what matter.

Every team has a set of recurring tasks that are predictable, multi-step, and time-consuming but not complex. Preparing a weekly report from multiple data sources. Routing customer feedback to the right internal channel. Researching a sales lead before a meeting and updating the CRM. Checking software requests against internal policy and filing IT tickets.

These are not hard problems. They are tedious ones. And tedious work has a compounding cost. According to a 2025 McKinsey survey, knowledge workers spend roughly 28% of their time on information gathering and 19% on internal communication. That is nearly half the workweek consumed by activities that a properly configured workspace agent can handle.

The winner here is not the company with the best technology team. It is the company that correctly identifies which workflows to hand off.

That is an operations call, not a technology call. It requires someone who understands the actual work being done, where the bottlenecks are, and which handoffs create friction. The CIO’s office might approve the license. But the person who makes this work is the operations leader who can map a workflow clearly enough for an agent to execute it.

What a Real Workflow Looks Like

OpenAI published several example configurations. The most telling one is what they call the Software Reviewer agent. Here is how it works:

  1. An employee submits a software request through the company’s internal channel
  2. The agent picks up the request automatically
  3. It checks the request against the company’s software policy (stored as a document the agent can reference)
  4. It evaluates cost, security posture, and whether an approved alternative already exists
  5. If the request passes, it files an IT ticket. If not, it responds to the employee with the reason

That entire workflow used to require a human to read, research, compare, decide, and respond. Now it requires a human to set the policy document and review edge cases.

Gartner’s 2025 forecast projected that by 2028, 33% of enterprise software applications would include built-in decision-making capabilities. OpenAI just moved that timeline forward by shipping decision-capable agents inside a tool that 400 million people already use.

What to Watch Out For

Three risks are real.

First, credit-based pricing after May 6 means costs will scale with usage. A workspace agent that runs 24/7 monitoring Slack channels will consume more credits than one that runs a weekly report. The free preview period is a trial, not a business model. Plan your budget before you build dependencies.

Second, data exposure. These agents connect to your real systems. Google Drive. Salesforce. SharePoint. An agent with broad permissions and a poorly scoped workflow can surface information in the wrong context. The security controls exist, but someone has to configure them correctly. That someone should not be the intern.

Third, the “automate everything” trap. The fastest way to waste this capability is to build thirty agents in the first week. Start with two or three workflows that are clearly defined, clearly repetitive, and clearly measurable. Prove the value. Then expand.

The Gap Is Compounding

Here is the part that should keep you up at night. 88% of organizations report using AI for at least one business function (Stanford HAI, 2026). But the distance between “using AI” and “using AI as persistent operational infrastructure” is enormous. Most companies are still in the first category. They gave people access to ChatGPT and called it adoption.

Workspace agents are that next category. AI that does not wait to be asked. AI that runs processes on behalf of the team, maintains state across sessions, and operates inside the tools where work already happens.

OpenAI’s enterprise revenue now exceeds 40% of total revenue and is on track to hit parity with consumer by end of 2026. Global corporate AI investment reached $581.7 billion in 2025, a 130% year-over-year increase. The capital is flowing. The products are shipping. The question stopped being whether your competitors will adopt this. Now it is how many quarters of operational advantage they stack before you do.

FAQ

Can I use workspace agents on a free ChatGPT plan? No. Workspace agents are available only on ChatGPT Business, Enterprise, Edu, and Teachers plans. They launched as a research preview on April 22, 2026, and are free to use through May 6, after which credit-based pricing applies.

What happens to my existing custom GPTs? They still work. OpenAI has announced a migration path to convert existing GPTs into workspace agents, but no firm timeline. If you have GPTs that work well today, keep using them while you experiment with workspace agents for new workflows.

How do I decide which workflows to automate first? Start with tasks that are recurring, multi-step, and clearly defined. Weekly reports, lead research, feedback routing, and policy-checking are strong candidates. Avoid workflows that require real judgment calls or context that changes frequently. The best first agents handle work that is predictable and measurable.

Research and structure: Mai. Direction and voice: John Lipe. SN client experience: first-hand.