LLMs Are Bad at Sales Forecasting. Everyone Else Is Too.

Should you trust an "AI Sales Forecast"?

My friend Peter Eigenschink and I recently posted about AI and sales forecasting.

This prompted the Funnelcast team to run a simple test. We gave several LLMs a sample sales dataset and asked them to forecast.

They defaulted to the metrics most sales teams are trained to look at:

• Velocity

• Coverage ratios

• Run rates

• Pipeline reviews

• Slippage

• Stage-to-stage conversions

...

Widely used. Easy to communicate.

But there are much better and more predictive was to model a sales engine.

So we tried something else. We stripped sales context from the dataset. Renamed the columns. Removed anything that looked like CRM language.

It did better.

Gemini (that's all we tried at this point) stopped parroting sales folklore and adopted a generic data scientist persona—groping it's way towards a recommended forecast model.

But even then, it took heavy expert prompting to reach something reasonable—logistic regression. (Stats 201 stuff.)

It didn’t present us with the best math for the problem.

It had to be led there.

We asked Gemini why it took such different approaches. It explained:

When data is labeled “Sales,” models revert to industry heuristics — even if they’re mathematically shallow.

LLMs don’t "think."

They don’t independently derive the mathematical structure of the problem.

They "predict" words based on what people usually say about sales forecasting.

And that’s the real issue.

LLMs didn’t invent bad sales forecasting. They learned it—from us.

And most of the conversation around sales forecasting is built on familiar—but weak—heuristics: velocity, coverage, run rates, activity, stage conversions...

LLMs didn’t choose those.

They reflected them.

The weak methods are not the fault of sales leaders. They aren’t statisticians—and they shouldn’t have to be. They’re operating with the frameworks they’ve been taught. By others who similarly were not statisticians.

But forecasting software should be held to a higher standard.

Which makes me wonder what’s under the hood of the “AI forecasting” tools:

1. Pure LLMs?

2. Statistically rigorous, purpose-built models?

3. Conventional sales arithmetic branded as AI?

If you don’t have a solid statistical framework—and raw LLMs aren’t sufficient—what’s left? Familiar heuristics in a modern UI?

AI is powerful. We use it too. But sales forecasting isn’t a language prediction problem.

It’s a statistics problem.

LLMs are bad at sales forecasting.

But that because everyone else is too.

When someone says “AI forecasting,” the real question isn’t whether they use AI.

It’s whether they use sound math.