The AI Data Diet

Ask an AI for last month's CPA while the real dashboard sits open next to it, and it will answer instantly, confidently, and wrong. Patrick Neyland, founder of Neyland Solutions, opens this AI Exchange with that exact moment, because the danger is not the mistake. It is that nothing in the answer tells you a mistake was made. If you have to validate everything AI gives you, for a lot of roles it is faster to do the work the old-fashioned way.

The costs are real. Wrong numbers become wrong decisions. Lawyers have been disbarred over hallucinated citations that never existed. And simple tasks burn real money: the same work that runs on a $200-a-month plan became a $2,200 bill in twelve days on the API, one lane over.

The middle of the session is the mental model. Same employee every time; only where the paper lives and how much you tell them changes. Hand your assistant three tidy pages and the answer comes back fast and right. Send them to a locked room of filing cabinets three floors up and the answer takes a hundred times longer and comes back stale. Context is working memory: the fuller the window gets, the worse the model performs, and a long fruitless search fills the window with everything it is not.

The fix is the same one you would use with a brand-new employee: get on the same page first. Do you understand what I am asking for? Do you have everything you need to do this? Patrick's client story makes the point. A Google Ads account where page views had been accidentally marked as conversions had the AI reporting "we are killing it." AI is only as right as your setup.

From there he lays out the Hierarchy of Data Efficiency, which he is quick to call a spectrum: chat copy/paste, an MCP connection with no instructions, connected with instructions, aggregated, and aggregated with instructions. Not everyone needs the top. Sometimes a screenshot into the chat is exactly right. But a small move up, usually just adding instructions, dramatically improves accuracy, speed, and cost.

The live build lands it. A deliberately messy Airtable ad table: spend in two currencies, a mid-dataset audience switch, conversion columns where blank means not tracked rather than zero. The naive analysis walks into every trap. Then Patrick writes one plain-English skill file, ten things only the team knows, and asks the same questions again. The refusal is the point. It knew what it could not claim.

The deck and Patrick's key-takeaways handout are published below, along with the two prompts that power the build.