Shield is using AI agents to carry out supported tasks in older IT portals, according to an OpenAI customer account published on November 13, 2026. For IT providers and the small businesses they support, this could remove some waits for a technician to navigate a website.
The useful distinction is between reaching a system, having permission to change something and completing the customer’s request. A support team needs an answer to each before handing over a recurring job.
What the browser agent does
The OpenAI account describes Forge, Shield’s IT-services platform, assigning supported browser tasks to GPT-6 Astra. The agent reads the screen, clicks and types, allowing it to work in portals without an API—a software connection that would otherwise let programs exchange instructions directly.
Technicians decide which actions can run without them and which need approval. Shield says these browser workflows are live, while phone calls, on-site work and remote-desktop tasks still involve people. Expanded remote-desktop workflows are being tested, not presented as deployed browser capabilities.
Shield lists password resets, employee onboarding and offboarding, licence renewals and firewall repairs among its broader AI-supported helpdesk work. That is not a confirmed list of browser-only tasks.
The Forge product page says the platform brings together accounts, devices, tickets, contracts, licences and operating procedures. It also advertises separate customer permissions and data, role-based access, approval requirements for sensitive actions and action logs. These are product descriptions, not an independent audit of those controls.
That customer-specific setup matters. Shield describes engineers spending weeks alongside providers and adapting workflows with technicians. Its helpdesk hands work to experts when technician-set rules require it. This is a configured service operation, not simply a model given a login.
A completed action is not a completed request
In its October 30 evaluation report, Shield examined 52 tickets its helpdesk agent had previously handled incorrectly. Technicians and engineers defined successful outcomes, and human reviewers checked an AI judge’s assessments.
One published example involved a customer requesting a phone call. An agent instead emailed to arrange remote maintenance; that did not satisfy the defined outcome. This was a benchmark case, not an observed incident in the live browser deployment.
The selected difficult tickets cannot establish a representative success rate for everyday support. Shield also discloses that OpenAI is an equity owner and technology partner of Thrive Holdings, whose benchmark includes the evaluation.
Three questions before delegating a workflow
For a recurring support task, use these three questions. This is a suggested assessment framework, not a description of Forge’s internal checks.
Can it perform this exact portal task? Ask the provider to name the supported customer system, action and sign-in requirements. Agree where it must stop if access fails or the interface changes. “Account administration” is too broad to define a delegation.
May it make this change for this customer? Record the customer’s prerequisites, permitted actions and technician approval points. Ask how the action log identifies the customer and how that customer’s permissions remain separate throughout the task.
What proves the request is finished? Agree the completion evidence and name the person responsible for exceptions. A successful portal step should not close a request that still needs a call, an on-site visit or another customer-specific step.
Request a walkthrough of one actual supported workflow, including an interrupted attempt and its handoff. The reviewed public pages do not supply a complete portal-and-action catalogue, authentication arrangements or contractual escalation commitments.
What to ask an IT provider before buying
Shield announced Forge on September 24, 2026, with access through a Shield partnership and deployment tailored to each provider. The November customer story is not a new self-service launch. The reviewed product and launch pages publish no subscription price; ask for provider-specific terms rather than assuming you can buy individual seats.
The OpenAI customer account reports faster execution and higher satisfaction, but does not establish a controlled comparison that predicts another business’s results. Ask for outcomes by supported task type, including reopened requests and technician involvement.
Shield’s evaluation costs and times exclude human checking, corrections and follow-through. They are not all-in service prices.
For a service comparison, count platform and usage charges alongside setup, review, correction and escalation costs for the same accepted outcome. Track customer waiting time separately from staff effort. A cheaper agent run is useful only if the completed service justifies the total cost.
For occasional document batches, see the complementary guide to AI agents for one-off office paperwork. Recurring IT support needs an ongoing agreement about customer procedures, permitted changes and who takes over unfinished work. Start the discussion with one named workflow and its responsible technician, not a promise to automate the whole helpdesk.