Example / Order support

Order support in chat: resolve the question before the refund.

Support gets expensive when a person has to find the order, piece together its state, look up the policy, and then switch to another screen to do anything about it.

Reference workflow. It is deliberately not a promise that an AI chat should make payments, refunds, or changes without your product's approval.

Request“Where is my order, and can I return it?”
Readapp.read_order
Policyapp.read_policy
Approvea person confirms the action
Resolvedreplacement requested · #4417
Works in
ChatGPT
Claude
Gemini
Coding agents

Direct answer

Use the chat to pull together the order and the policy, explain the next step the customer qualifies for, and stop for confirmation before requesting a refund, a replacement, a cancellation, or an address change

Read first01

Start with an order the customer can verify.

The first action finds the right order and returns a short, readable summary: the status, where the shipment stands, the dates, and the fields needed to decide what happens next. The rest of the customer's history stays out of it by default.

If the request is vague, the chat should ask for an order number, a date, or another reliable reference instead of guessing.

Apply policy02

Your policy lives in your product, not in the AI's memory.

A policy rule needs a source and an owner. Your product can hand back the current rule, the condition that matched, and the actions this person is allowed. That gives the assistant a grounded explanation, and a policy change never means editing a prompt.

A short, precise answer from your product is usually safer than sending an entire policy manual to a model on every order request.

Recommend, then act03

Separate the suggestion from the action.

The chat can suggest a replacement, a handoff to support, or a refund request. The action that changes something should demand the exact order, the action, the reason, and a confirmation. When it succeeds, it returns your product's record, not a reassuring sentence.

OpenAI's Apps SDK documentation keeps server tools, sign-in, and optional interface pieces separate. That split helps here: a result can make sense to the person without asking them to trust an effect they never saw.

OpenAI Apps SDK · build an MCP server

Escalate well04

Make the handoff an outcome, not a failure.

Some requests need a person: damaged goods, signs of fraud, exceptions, or a case the policy does not cover. The workflow should open a useful support handoff with the order details and the person's request already attached.

That still counts as a finished task. It avoids handing over a blank ticket, and it leaves the business decision with the team that owns it.

AI / Second opinion05

Not sure where to start? Ask an AI.

Paste this question into ChatGPT, Claude, or Gemini. It walks you through finding the one task your customers would most want to finish by chatting, instead of clicking through your app.

“I run a software product. Help me identify one high-value workflow my customers could finish inside ChatGPT instead of switching between tabs. Ask me about the product, the user, the action, the data it needs, permissions, and the safest small first release.”

The prompt is copied as a backup. Some AI chats may ask you to paste it after sign-in.

FAQ06

Questions, answered.

Is this just an MCP server?

No. The server is one piece. We do the whole job: we pick the customer task, set who can do what, handle the quirks of each AI chat, make the actions run fast, add monitoring, test it, and ship it.

Do we have to rebuild our product?

Almost never. Our layer sits in front of what your product already does. We start with one small workflow and grow from what works.

Will it work in ChatGPT, Claude, and Gemini?

The same build can serve all three. Each one has its own sign-in, look, approval step, and publishing rules, so we set up and test each one separately.

How do you keep it safe?

Your product decides what is allowed, and we keep that in charge. Every action is spelled out, every input is checked, and the customer sees an approval step before anything happens. Our layer sits beside your product, not inside it.

Where does it run? Does it touch our code?

It runs as a separate service in front of your product. It never goes into your code, and you still own your data.

What does it cost, and how fast?

A first workflow starts at €3,200 and goes live in about 7 days.

Contact us07

Now build yours

Send us your product, who uses it, and the action you have in mind. We will tell you the smallest version worth building, and whether it is a €3,200 first workflow or something that needs a closer look first.

Protected by reCAPTCHA. Google's Privacy and Terms apply.