Ryan Brady, the chief information officer at Select Health, a nonprofit health plan in Utah, told InformationWeek at the Forbes CIO Summit about a meeting where he kept questioning a slide deck until the presenter admitted that Claude had written it and blamed Claude for the weak parts. His point was that people should stay in the flow of AI-assisted work, not just wait at the end to approve it.
The story is worth more than the anecdote. It shows that AI changes how much of your own work you have actually read, and it changes nothing about who answers for it.
The person receiving the work never sees the tool
A boss reading a board pack cannot ask questions of Claude. A client reading a report cannot either. They deal with you, so you are the one who gets the question, and "the tool did it" is an answer to a question nobody asked.
In October 2025 Deloitte Australia agreed to refund part of the A$440,000 the Australian government paid for a report that had AI-generated errors in it, including an invented quote from a federal court judgment. The government had paid for the firm's judgment, not for a tool, and the firm gave money back.
Checking belongs inside the work, not at the end
Brady prefers "human in the flow" to "human in the loop". A loop puts a person at the end to approve what the AI produced. A flow keeps the person present while it is produced, and that is where mistakes are cheapest to catch, because you still know what the AI was asked and what it was given.
For example, a finance manager asks AI to turn a monthly spreadsheet into a 30-slide board deck. The draft takes ten minutes and the same deck used to take half a day. Reading it properly means tracing every number back to its cell and asking the AI where each claim came from, and that is about 45 minutes. The manager still gets most of a morning back every month, and the deck is theirs in a way it was not when they only skimmed it.
The context is the part only you have
Mark Wanish, who leads architecture at Salesforce's regulated-industries unit and spoke on the same stage, gave a plain example of why. A health plan's data may list one doctor at three addresses, and the model treats them as one person only when something tells it so. The model works from context, and people supply that context.
The same is true in your own files. The AI does not know that two customer names are one customer, that a supplier's price list is out of date, or that a figure in last year's report was restated. The person who knows these things is the one who catches the error, and that knowledge is what a sent document is vouching for.
So before sending something AI helped write, pick the three claims the reader would act on and trace each one to its source yourself. If you could not answer a question about a slide without opening the deck, the work is not ready to send.