Why Pasting Your Dashboard Into ChatGPT Gives You Garbage Advice (and the Context Doc That Fixes It)

"I screenshotted my whole dashboard, gave it to the AI, and it told me to kill the channel that's actually carrying the account." A founder said that to me on a call a few weeks back, half laughing, half rattled, because for a second he'd nearly done it.

I've watched this exact thing happen enough times now that I want to pull one example apart in detail, because the lesson sitting inside it is the most important thing to understand about using AI on your numbers in 2026. The tool isn't broken. It's missing something. And once you see what, you can't unsee it.

The teardown: watch it confidently get it wrong

Here's the setup, close to one I ran myself. Take a screenshot of a business dashboard. Month-to-date pacing against target, dozens of metrics, each one colour-coded. Green means ahead of plan, red means behind. Contribution margin at the top, then the business-level numbers, then each ad platform below.

Paste it into the model and ask a reasonable question. "Build me a plan for the right actions to hit our contribution margin goal this month. How should I prioritise?"

It starts confidently. It correctly spots that contribution margin is pacing behind. Good. Then it says: the problem is almost entirely Google. Spend up 51%, revenue down 45%. And off it goes, building a whole plan around fixing Google.

There's just one issue. Google was green. On the actual dashboard, Google was well ahead of target on both spend efficiency and revenue. It wasn't the problem. It was one of the things going right. The model had invented a crisis out of thin air and was about to have me pour my attention into the one place that didn't need it.

So I pushed back. "Google's ahead on spend and efficiency, it's green, it's not a problem." And it folded instantly. "You're right, I misread the colour coding. Green is good, red is bad. Let me re-read." Polite, agreeable, and completely backwards thirty seconds earlier.

That's the experience so many founders are having and quietly distrusting. The numbers come out wrong. It hallucinates a story. And because it says it with total confidence, you have to already know the answer to catch it. Which rather defeats the point of asking.

Why it happens (and it's not the model being dumb)

Here's the part worth slowing down on. The model didn't fail because it's stupid. It failed because nobody told it how to read what it was looking at.

It saw a wall of numbers and colours with no idea what any of it was for. It doesn't know that green means good on this particular dashboard. It doesn't know contribution margin sits at the top because it's the goal and everything underneath exists to explain it. It doesn't know which metric to look at first, or what "good" even means for your business. So it does what these tools do when they're under-briefed: it pattern-matches, picks a number that looks dramatic, and writes a confident story around it.

I sat in on a session once where a big ad platform was showing off its own AI media-buying assistant. Someone asked the obvious question: what methodology is it trained on to make recommendations? The honest answer back was that there isn't one. It's built to do what you ask, not to hold a point of view about what's right. And that's the whole thing in a sentence. Most of these tools are execution engines. They'll enact your request beautifully. They will not supply the judgement you didn't give them.

So when you paste a screenshot in and ask "what should I do", you're asking a brilliant, fast, eager analyst who has never seen your business, doesn't know your margins, and has no framework for what matters, to make a senior call. Of course it hallucinates. You'd be amazed if it didn't.

The fix is a layer, not a better prompt

The reflex when this happens is to go hunting for a cleverer prompt. That's not where the fix is. The fix is to hand the model the missing layer between your raw data and a good decision.

Think of it as three layers stacked up. At the bottom, the data: all your numbers in one place, sitting against your targets so the tool can see what's ahead and what's behind. In the middle, the methodology: how to actually read that data, what to look at first, what "good" means. At the top, you, making the final call.

Almost everyone trying to use AI on their numbers has the bottom layer and is missing the middle one entirely. They've got the data. They've got no shared way of interpreting it. So they're asking the tool to skip straight from raw numbers to a senior recommendation with nothing in between, and then they're surprised when the recommendation is rubbish.

Watch what changes when you add the middle layer. In that same teardown, instead of arguing with it metric by metric, I gave it the framework. Here's the hierarchy. Contribution margin is the scoreboard, the only number that's actually the goal. Everything below it exists to diagnose and protect that number, in this order. Read it top down.

The output flipped completely. Now it opened with the scoreboard, contribution margin, and read everything beneath it as a diagnosis of that one number. It noticed new customer acquisition was running hot, which reframed the whole gap. It correctly read the channels, including the fact that Google was performing. Same tool, same screenshot, same model. The only thing that changed was that I'd told it how to think. It went from useless to genuinely sharp in one step.

The context doc that does the heavy lifting

So the practical move is to write that middle layer down once, properly, and feed it in every time. I'd build a short context doc and keep it to hand. A few things belong in it.

• Your definitions. What counts as revenue here, gross or net? How do you define contribution margin, and what's included in it? These sound pedantic right up until the model assumes the wrong one and every number it gives you is built on sand.
• Your metric hierarchy. What's the one number that's the actual goal, and what's the order of everything else underneath it? This is the single highest-value thing in the whole doc. It's what stopped the Google hallucination dead.
• Your targets. What are you pacing against? A number with no target attached is just trivia. "Spend is up 51%" means nothing until the tool knows whether 51% up is on plan or a fire.
• Your guardrails. Your CPA or efficiency targets, and the brand lines you'd never cross. The discounts you won't run, the claims you won't make, the things that are simply off the table no matter how good they'd look in a forecast.

That's most of it. Definitions, hierarchy, targets, guardrails. None of it is fancy. It's just the stuff that lives in your head and your team's heads and has never been written down in one place where a tool could read it.

And here's the quiet payoff. The moment you've written it down for the AI, you've also written down the thing your own team has probably never had explicit: a shared, on-paper view of what good looks like and what order to think in. The doc that stops the model hallucinating is the same doc that stops two of your people reading the same dashboard and reaching opposite conclusions.

The part underneath all of this

There's a deeper thing here, and it's the bit I keep coming back to.

These tools are about to give you more information than you've ever had. More analysis, more reports, more "here's what I noticed", faster than you can read it. And more information is not the same as a better decision. I've watched people drown in it, surfacing twelve problems a day and assuming all twelve need solving, when half of them were never problems at all.

A framework is what turns all that information back into a decision. Without one, the AI will happily run in any direction you point it, with total confidence, and some of those directions are straight off a cliff, like killing the channel that's carrying you. The data layer is getting commoditised. Everyone will have it soon. The thing that's actually scarce, and getting scarcer, is a clear point of view about what your numbers mean and which way is up.

So before you paste another screenshot into a chat box and ask it what to do, the question I'd sit with is this one: if you handed your dashboard to a sharp stranger with no knowledge of your business, would they read it the way you do? If the answer is no, that gap isn't the AI's to close. It's yours. And it's worth asking whether you've ever actually written down how your business is supposed to be read, or whether it's only ever lived in your head.