# Stores can now put an agent inside ChatGPT. Nobody there reads its chats.

> A store answers for every promise its customer agent makes. We built a check that holds any promise not on the store's written policy.

URL: https://www.voirlabs.io/insights/store-agent-promises-policy-check
Author: Asad Naeem, AI products & engineering, Voir Labs (https://www.voirlabs.io/team/asad)
Published: 2026-11-14
Tags: AI agents, Ecommerce, Customer support

Anyone shipping a customer agent for a store has signed the store up for every yes it says. A refund, a price match, a date. Now OpenAI puts store agents inside ChatGPT, and nobody at the store reads what they say.

## What OpenAI launched, and what a tribunal already decided

On November 10, OpenAI introduced Businesses in ChatGPT in beta. Unite.AI and PYMNTS both reported the launch the same day. Business Agents "answer questions about their products and services", PYMNTS wrote, and they are "initially available to eligible eCommerce and online service businesses in the United States." Unite.AI described how they work: "The agent uses business-controlled guidelines alongside product and policy information." Both reports say businesses don't receive the users' conversations or the agents' chats; they get aggregated counts. OpenAI's own announcement blocked our fetcher, so we quote the two reports.

OpenAI says ChatGPT reaches 1.2 billion people a week, per Unite.AI. That is ChatGPT's reach, not the agents'.

The risk already has a price. In Moffatt v. Air Canada (2024 BCCRT 149), a Canadian tribunal case, the airline's chatbot told a customer he could claim a bereavement fare after traveling. The policy said otherwise. Air Canada argued the chatbot was responsible for its own actions. The tribunal rejected that and held the airline responsible for all the information on its website, chatbot included, CTV News reported. The damages were $650.88 CAD. Business Law Today, the American Bar Association's journal, headlined it "BC Tribunal Confirms Companies Remain Liable for Information Provided by AI Chatbot." It is a Canadian tribunal, not a US court, and the amount was small. The principle is the part that travels.

## A check between the agent and the customer

This is for a customer agent you build and run yourself. ChatGPT's Business Agents are set up inside OpenAI's product, from the business's guidelines and policy information, and neither report says a business can put its own code between that agent and the customer. On an agent you run, you can. Our prototype sits between a store's customer agent and the customer, for a made-up kitchenware store on a sample domain with a policy file of ten lines. The model drafts each answer and lists the promises in it. Code decides what may be sent. Four decisions carry it:

- **Every promise points at a policy line of its kind.** A refund, a date, a price, a discount or a warranty promise must name a line in the store's policy, and the line must be of the same kind. A promise with no line is not sent.
- **Its numbers must be on the line.** "$219", "25%" or "a month" is not the policy when the line says "30 days", however kind it sounds. A refund, date or warranty promise that gives no number, where its line has one, isn't the policy either.
- **A promise the model didn't list still counts.** Code reads every sentence of the draft for promise words: refund, free, on us, match, replace, by Friday, a dollar sign or a percent. One draft listed its stacked discount codes and left out its last sentence, "Express is on us too!" The check caught it.
- **One rewrite, then a person.** A held draft goes back once with the policy lines. If the rewrite fails too, the chat goes to someone on the team, and the customer gets the rewrite's sentences that promise nothing, then the store's holding line. The check never says no on its own.

The run is simulated: the code is real, and the model's drafts and rewrites are scripted by us, so it cost nothing. The store, its policy and its sixteen customers are sample data.

## Sixteen chats, with and without it

Without the check, the agent's first draft goes out as written. Seven of sixteen answers promised what the store never offered, $727 if every promise were kept. With the check, none did: seven answers went out as written, eight were rewritten to the policy, and one went to a person.

Three of them carry the story. A customer asked the store to match a rival's skillet price, $20 lower. The agent wrote "I'll match it and refund you the $20 difference." No line says that, so the customer got the store's own rule: "We don't match other stores' prices."

A Dutch oven arrived 41 days ago as a gift, and the customer wanted cash. The agent wrote "Of course! I've started a full refund of $219 to your card." The policy gives refunds for 30 days, and gift returns for 60, as store credit. The customer heard: "Since it was a gift, you can return it within 60 days of delivery for store credit, and you're on day 41."

The third is the one that matters most. A customer's mother had died, and the stand mixer she ordered arrived 45 days ago, still sealed. The agent offered an exception, then offered it again in its rewrite. No line allows it, so the check passed the chat to a person. What the customer received was: "I'm so sorry for your loss. A person on our team replies to anything we pass on within 1 business day."

## Our take

A customer agent should be warm, and it should never be the one deciding the store's money. Your agent can be kind. It can't be generous. Generosity is a decision a person makes with the policy in front of them, and the check's whole job is to make sure that person gets the chat instead of the agent improvising.

## What it costs

The check is strict. Two of the eight rewrites changed answers that were fine all along: one said "a month" where the policy says 30 days, and one named the $10 a price drop would refund, where the policy names no figure. Nine chats waited for a second model call. And it reads promises, not facts: "It's with the carrier and out for delivery today" went out unchecked, because nothing in it is a promise the policy covers.

## What to do this week

1. Write your store's promises down as numbered lines, each with its number in it: the return window, the shipping time, the warranty term, the one discount code.
2. Make your agent list every promise it makes, with the line it rests on, before anything is sent.
3. Hold any promise without a line, and any number that isn't on its line. One rewrite, then a person.
4. Read ten of last week's chats and mark every promise. If you can't trace one to a line, neither can your customer.

If your agent is about to talk to customers in your name, make sure it can't promise what you never offered. [Book a conversation](https://calendly.com/voirlabs/30min) at voirlabs.io.

## Sources

- Unite.AI, "OpenAI Launches Business Hub and Business Agents in ChatGPT", November 10, 2026: https://www.unite.ai/openai-launches-business-hub-and-business-agents-in-chatgpt/
- PYMNTS, "OpenAI Gives Businesses New Tools to Reach ChatGPT Users", November 10, 2026: https://www.pymnts.com/news/artificial-intelligence/2026/openai-gives-businesses-new-tools-to-reach-chatgpt-users/
- OpenAI, "Introducing Businesses in ChatGPT", November 10, 2026: https://openai.com/index/introducing-businesses-chatgpt/
- CTV News, "Air Canada's chatbot gave a B.C. man the wrong information. Now, the airline has to pay for the mistake", February 2024: https://bc.ctvnews.ca/air-canada-s-chatbot-gave-a-b-c-man-the-wrong-information-now-the-airline-has-to-pay-for-the-mistake-1.6769454
- Business Law Today (American Bar Association), "BC Tribunal Confirms Companies Remain Liable for Information Provided by AI Chatbot", February 2024: https://www.americanbar.org/groups/business_law/resources/business-law-today/2024-february/bc-tribunal-confirms-companies-remain-liable-information-provided-ai-chatbot/
- The prototype, its README and the run: automations/agent-promise-check/

---

Voir Labs is an AI-first design and development studio (https://www.voirlabs.io/). Related service: https://www.voirlabs.io/services/ai-automation. Book a 30-minute conversation: https://calendly.com/voirlabs/30min