OpenAI Brings Shared Workspace Agents to ChatGPT: What Businesses Should Assess Before Adoption

OpenAI’s new feature lets teams create Codex-based agents for complex workflows, with integrations, controls and approvals. An operational read for CIOs, IT leaders and digital teams.

The update in brief

OpenAI has introduced a new generation of shared agents for teams in ChatGPT. According to the announcement, they are based on Codex, run in the cloud and can manage complex, long-running tasks while staying within permissions and controls defined by the organization.

The feature is presented as an evolution of GPTs: it does not simply respond to a prompt, but can gather context, use connected tools, write or execute code, remember useful information and continue activities across multiple steps. Agents can be used in ChatGPT and Slack; OpenAI says additional surfaces will come later.

Availability is in research preview for ChatGPT Business, Enterprise, Edu and Teachers. For Enterprise and Edu, administrators can enable the feature through role-based controls. OpenAI also states that usage will be free until May 6, 2026, with credit-based pricing starting from that date.

Where they can be genuinely useful

The examples cited by OpenAI are concrete and help define the scope: reviewing software requests, routing product feedback, producing weekly metrics reports, qualifying leads and analyzing third-party risk.

From an operational standpoint, these use cases share common traits: repeatable rules, distributed information sources, steps across different tools and moments where human review is required. They are not simple chatbots, but processes that can be turned into shared procedures.

For a company, the first filter should be pragmatic: choose workflows that are frequent, reasonably standardized and already documented. If the process is currently ambiguous or changes depending on who performs it, an agent may amplify disorder rather than reduce it. Before automating, it is better to clarify roles, exceptions, allowed data and quality criteria.

Governance is the central issue

OpenAI highlights several controls: selecting which tools and data are accessible, defining allowed actions and requiring approval for sensitive steps such as modifying a spreadsheet, sending an email or adding a calendar event. Enterprise and Edu administrators can manage access to tools and actions by user group, as well as permissions to use, create and share agents.

Visibility elements are also included: usage analytics, number of runs and number of people using an agent. According to OpenAI, the Compliance API provides visibility into configurations, updates and executions; administrators can also suspend agents if needed. The company also cites safeguards against misleading external content, including prompt injection attacks.

For CIOs and IT leaders, the takeaway is clear: adoption should be treated as an extension of business processes, not as an isolated experiment. Organizations need at least a basic matrix linking each agent to its owner, data sources, connected tools, permitted actions, required approvals and available logs. These are the same kinds of digital governance activities addressed in Dedaloop’s Strategic IT advisory and governance work.

Checklist for a controlled pilot

Before deploying shared agents on real workflows, it is advisable to set up a tightly scoped pilot. An essential checklist may include:

Metrics should go beyond time saved. It is useful to measure the percentage of outputs accepted without changes, number of human interventions, detected errors, requests blocked because of missing permissions, tickets generated and team adoption level. Only then can an organization understand whether automation is improving the process or simply moving control work downstream.

Security, cloud and integrations

The fact that agents run in the cloud and can interact with business tools makes access design central. Every integration increases both usefulness and the risk surface. For this reason, companies should apply the principle of least privilege: fewer connectors, fewer automatic actions and more checks on sensitive steps.

OpenAI mentions the ability to run scheduled activities and deploy agents in Slack, so requests can be handled where work already happens. This is an operational advantage, but it requires clear rules: which channels are allowed, which data can be read, when the agent should respond and when it should open a ticket or ask for confirmation.

For organizations that are already consolidating cloud environments, observability and security, the topic naturally connects with cybersecurity and cloud infrastructure practices. Monitoring, incident management, identity control and configuration review become part of the same operating architecture.

What to watch in the coming months

OpenAI anticipates new automatic triggers, better dashboards for evaluating performance, more actions across business tools and support in the Codex app. These developments are worth monitoring closely, especially for organizations that have already evaluated recurring automations based on Codex.

The direction is clear: generative AI is moving from individual assistant to component of enterprise workflows. The best choice today is to start with limited but measurable use cases, with explicit governance and security integrated from the beginning. There is no need to automate everything: the priority is to understand which processes truly deserve an agent and which should first be simplified.

Approfondimenti Dedaloop

Fonti