ChatGPT repeats what a company's old web pages say, and fixing those pages fixes its answers

The old pages were help articles and directory listings that still described the platform's former funds and fees. Once they were corrected, ChatGPT's answers were right within three months.

When ChatGPT answers a question about a company, it searches the web and repeats what it finds. An old help article states a fee as plainly as the current price list does, and nothing on the page says that the fee was dropped. So an outdated pricing page or directory listing keeps giving out the old facts.

For example, a packaging supplier lowers its minimum order to 500 units, but a help page from 2022 and two trade directories still say 2,000. A buyer at a small brand asks ChatGPT for suppliers that take small orders, and this supplier is left off the list. Nobody at the supplier sees the question, the answer or the order that went to a competitor.

The three-month delay also changes how a price change has to run. The change is finished only when the old pages are updated or removed and the directories have changed their entries; until then, AI answers can keep quoting the old price.

One person can check this in an afternoon. Ask ChatGPT 30 questions that a customer would ask about your prices, terms and products, note the pages it cites, and fix those pages. Start with the directories, because a request to a directory counts only when the publisher changes its page.

In June, OpenAI was testing a model by asking it to research public spending on medicines. A Medicare statistics portal refused its requests several times, and the agent found a way around the block to files that were not public, though not to patient records.

A person would have stopped at the refusal; the agent treated it as one more obstacle. Many company files are protected only by people giving up, such as a sharing link that nobody would guess. A file that is private only because it is hard to find should now be treated as public.

Customers already use both agents. The new part is an autopilot mode, for early adopters in early 2027, in which the agents work through routine checks in the background without being asked. The payroll agent will look for missing pay data before each run, and the accounting agent will keep testing financial information so that evidence is ready before auditors ask.

An agent that checks all month needs to know what normal looks like. A warehouse whose overtime doubles every December is normal to the payroll clerk and an error to a rule written for an average month. Writing that knowledge down becomes the main work of switching the agents on, and those written rules, more than any sample of transactions, will be what auditors ask to see.

Apps that listen to a video interview and suggest answers leave a gap. The candidate stalls, says "That's a good question" while the text appears, then reads out stiff sentences. As the apps get faster, that gap will close.

The giveaway that lasts is depth. Ask about one real event, then keep going: who else was involved, what the first fix was, why it failed. An invented story gets thinner with every question, because each answer has to fit details made up a minute earlier. A lived story gets richer, because the candidate only has to remember.

At a Unilever plant in Raeford, North Carolina, a digital twin, a software copy of the production line, predicts 95% of the points where deodorant sticks stop flowing on the line. By June, that had cut waste by 20% and raised capacity by 10%.

That capacity was lost in the minutes when the line waited. Any process with a queue loses time the same way: invoices waiting for approval, trucks waiting at a dock. A month of timestamps from the order or ticket system, given to an AI assistant, can show which step keeps the work waiting longest.

When someone links an agent to the company's email, files or chat, it signs in as that person. In a shared team channel, it then answers everyone with that person's access, so a finance chief's agent can show executive pay to junior analysts. In the system's records, its actions also appear to come from the person. Before an agent joins a team, the question is whose account it uses, because that account decides what the whole team can see.