By Scott Sumner | AI at Work | Published September 30, 2026

A client brings a purple folder to a professional working at a bench, beside a reception counter. A useful tool helps the customer do something, not just find the business.

ChatGPT plugins give businesses another place to help customers do useful work. OpenAI’s latest changes make that opportunity worth a look, but a new place to put your logo is not a reason to build software.

Answer: Build a ChatGPT plugin when it helps customers complete a specific task using capabilities or data your business provides. Do not build one merely to repeat your website. Check demand, current platform rules, maintenance needs and a measurable customer outcome before investing.

The first decision is a product decision, not a coding decision. What can someone finish with your business inside ChatGPT that they cannot finish with an ordinary answer?

Jump ahead

What OpenAI announced on September 29

In its DevDay 2026 recap, OpenAI announced plugin extensions that can give a plugin a home in ChatGPT’s sidebar, support interactive panels beside a conversation and provide viewers for supported file types.

The company also announced Plugin Creator, a revised submission flow and improved ranking and recommendations for discovering plugins in the directory and in conversations. The recap lists these plugin updates as available to all plans. Users choose the plugins they use and approve the access each receives.

That is the announcement. It is not evidence that a particular business will receive customers from the directory, nor a promise that every submitted plugin will be approved.

The accompanying developer documentation explains the practical shape of the experience: an app can open beside a conversation, expose product-specific settings, handle supported files or ask for structured input through forms.

For a business owner, the important change is not the list of developer features. It is the possibility of letting someone use part of your service without leaving the work they are already doing in ChatGPT.

Discovery and doing are different jobs

A prospect asking which agency handles a website migration is trying to find and evaluate a supplier. An existing client checking which migration tasks are complete is trying to use a service.

Those are different needs. They should not automatically produce the same investment.

Your website explains the offer, shows evidence and gives buyers a route to contact you. A plugin can expose an actual capability: retrieve the client’s current project status, work with a file in your product or help assemble a plan using information your system maintains.

Building a plugin is not a substitute for AI Findability. Nor does a directory listing establish that your business will be recommended in general AI answers. Treat discovery of the business and use of the product as separate things to evaluate.

I would not commission a plugin whose entire job is to introduce the company and repeat the service page. An accurate page may already do that job with much less upkeep.

A useful plugin earns its place by helping the customer finish something. The logo is incidental.

Choose a customer task worth building

Start with an action customers already take, not a feature someone demonstrated on a conference stage.

Here is a hypothetical example for an agency with an existing client portal. A client wants to know what is preventing a campaign launch. Today, they open the portal, inspect approvals and copy the relevant items into a message.

A proposed plugin could retrieve that client’s current project information and return the outstanding decisions with links to the underlying records. Its value would come from live, authorized project data, not from being another chatbot that can explain campaign planning.

This is a product concept, not a claim that we have built or tested it. It also might be unnecessary. If clients already use the portal comfortably and rarely ask that question, the plugin could become another interface to maintain without solving much.

A different business might have a specialist file viewer or a configuration tool that fits naturally beside a conversation. The new extensions give developers ways to present those capabilities. They do not supply the customer demand.

Before building, write down the task in the customer’s words. “Show me which approvals are holding up launch” is a useful start. “Increase our presence in the AI space” is not a product requirement.

Check the rules before the business case

OpenAI’s current plugin guidelines require a clear purpose, reliable behavior and functionality or workflows beyond the products’ built-in capabilities. They reject static frames with no meaningful interaction and say trial or demo plugins will not be accepted.

That matters if your plan is to wrap an existing marketing page and call it an app. A new location for the same sales copy is a weak customer experience and does not meet the stated standard for meaningful interaction.

There is also a commercial constraint worth reading before anyone writes a proposal. At the time checked, the guidelines permit commerce for physical goods and prohibit selling digital products or services, including subscriptions, directly or indirectly through the plugin.

They separately allow users to sign into an existing paid account and use features already included in their subscription. They do not allow the plugin to display subscription plans, initiate new subscriptions or promote upgrades. The guidelines describe limited informational explanations for unavailable entitlements, not an unrestricted checkout route.

For a service business, “help an existing customer use what they already have” and “sell a new service inside ChatGPT” therefore need different treatment. Do not assume the second is permitted because the first is possible. These are platform requirements, not legal advice, and they can change.

Public submissions also need a privacy policy. Data collection must be limited to what the function requires. Have the implementer review the current requirements against the actual experience before committing to a launch date.

Test usefulness before chasing distribution

The directory is not a demand forecast. OpenAI says strong real-world utility and user satisfaction may make a plugin eligible for enhanced distribution. “May” is doing useful work in that sentence.

There are two separate tests: does the right tool get selected, and does it help the customer complete the task?

OpenAI’s discovery guidance recommends testing direct prompts that name your product, indirect prompts that describe the desired outcome and negative prompts that should not activate your tool.

In the agency example, “Check the launch approvals in our client portal” could be a direct test. “What is holding up our campaign?” could be an indirect one if the relevant account is connected. “Explain what a campaign approval process is” may be better answered without calling the plugin at all.

Being invoked more often is not automatically better. Being invoked for the wrong task creates confusion.

My suggested pilot would also record whether the result is correct, whether the customer can find the original record and whether they still need to ask a person to reconstruct the answer. Compare that experience with the existing portal. Do not count a successful connection as a successful customer outcome.

Include the maintenance work in the decision: authentication changes, API changes, error handling, mobile behavior and support. Someone has to keep the experience useful after the launch screenshot.

Write the brief before buying the build

A sensible brief can fit on one page: the customer, the task, the unique capability or data you supply, the current way of doing it and the evidence that the new experience is better.

If the unique contribution is only public company information, improve the website first. If customers repeatedly need a real capability while working in ChatGPT, a focused plugin experiment may be worth the effort.

Keep your website understandable to agents either way. A plugin is an additional product interface, not an excuse to neglect the public explanation of the business.

Want help separating a useful AI integration from an expensive brochure? Bring one customer task and the way it works today. That is enough to start a practical conversation.

Sources

Sources checked September 30, 2026. Product capabilities and platform restrictions above are attributed to OpenAI. The investment criteria and agency example are analysis, not measured results or an endorsement of an untested implementation.