On 13 November 2026, OpenAI published a customer story about Shield. Shield uses OpenAI to help IT technicians hand off work in legacy portals. Its Forge agents navigate those systems, so technicians can focus on the issues that need their judgement.

I could not load the full article, so I am working from OpenAI’s summary only. I won’t guess at numbers or results that I haven’t seen.

The excitement is about agents, but the point is the portal

Most of the talk around stories like this is about the agent. How smart is it, which model runs it, what else could it do. I think that is the wrong place to look.

The useful part is where the agent works. It works inside legacy portals. Those are the old systems that businesses already depend on, the ones with no modern integration and a login screen that someone has to type into every day. A technician opens the portal, finds the record, checks a few fields, makes a change, and moves to the next one. Nobody enjoys that work, and nobody needs judgement to do most of it.

Is there a real need? Yes, and it is an old one

My first question about any AI story is whether there is a need for it. Here the need is easy to see. Legacy portals exist in almost every business. Replacing them is slow and expensive, so the work around them stays manual.

What changed is that an agent can now operate those systems the way a person does. It navigates the screens and carries the task through. That is the real change.

What did not change is that someone still has to decide the hard cases. The summary says technicians keep the issues that need their judgement. That split is the sensible one. The routine goes to the agent. The exceptions go to the person.

Automate the process, not the chat

I keep coming back to one position. Automate the workflow, not the chat. The value is a process that runs end to end, not a smarter answer.

A chatbot can tell a technician what to do in the portal. That saves a little time. An agent that does it inside the portal saves the whole task. One answers and the other acts. For an operator, only the second one takes work off the team.

This is also why I don’t care much about the label “agentic”. What matters is where the thing sits and what it does in your systems. Shield’s agents sit inside the portals that the technicians use. That is a concrete place with a concrete job.

What this looks like in a normal business

Take a clinic that books appointments in one system, bills in another, and keeps patient records in a third older one. Front desk staff retype the same details across all three. Or take a service business where staff copy job details from an email into a supplier portal, then copy the confirmation back.

None of that work needs judgement. But a few cases in every batch do. A booking clash, a billing dispute, a supplier rejection. Those are the ones you want your people on.

Rules matter here too. Anything that commits money, promises or dates should be limited by the system, not left to the agent’s judgement. When the agent hits a case outside those limits, it should pass the case to a person with the details attached. I read Shield’s split between routine work and judgement work the same way, though I can’t say how Shield sets its own limits.

Where this meets what I build

This is the same thinking behind AeroChat. It is chat infrastructure that lets you run your customer service and sales workflows on autopilot. You describe the workflow you want, and the system turns it into instructions and runs it. It connects to your own data systems and backend processes, so it can perform actions, not just answer questions. Shield’s portals are a different setting from a customer chat, but the principle is the same. The process runs inside the systems you already have.

What a business should do

Start with a list, not a tool. Ask each person on your team which tasks they repeat in an old portal or system. Write them down, and mark the ones that need no judgement.

Pick one. Describe it in your own words, step by step, as you would to a new hire. Say what the agent may change and what it must hand to a person. Then run it on a small batch and check every result before you widen it.

If one task works, add the next one. Keep your people on the exceptions.

Source: OpenAI News: Shield helps technicians hand off legacy portal work with OpenAI