Kepler's CEO says trusting AI work means having someone to blame when it is wrong

Vinoo Ganesh, CEO of the AI firm Kepler, argued that people trust AI work only when someone can be blamed for its errors. He made the case on the AI Engineer show. Kepler sells a verification system, so he speaks as a vendor. His talk covered how firms might check AI output before they rely on it.

What was said

Ganesh used to lead business engineering at Citadel. He began with a puzzle. Many buyers are spending heavily on AI, he said, yet investing has not been transformed. His explanation: to produce work you believe in, "you have to trust it, meaning you have to blame somebody who gets it wrong."

He named three causes of wrong numbers. The system may fetch the wrong document. It may pull the wrong figure from the right document, such as a 2025 value when the question was about 2023. Or the model may try arithmetic. He said models cannot do maths, because they only predict text.

Kepler's fix starts with ordinary code. It extracts every fact ahead of time and gives each one an ID. The model then writes a plan that points to those IDs. It never types a figure itself. Anything it cannot resolve is corrected or refused.

On the wider market, Ganesh was blunt. He said "the vast majority of finance AI tools right now are being used to format PowerPoint presentations."

Why it matters

Our reading: the blame test turns a technical gap into a question of who pays. If a tool cannot show where a figure came from, the analyst who signs off owns the mistake. That gives cautious firms a reason to keep AI on low-stakes jobs.

Ganesh gave buyers questions for any vendor. Can you click a number and see its source? Is every citation checked? Is the arithmetic checked? What happens to a figure that came from nowhere? These questions test whether a vendor can show its work.

The other side

Ganesh sells this approach, so his view of the market comes from an interested party. He also said the fix is not perfect. In Kepler's production data, refusals dropped sharply, but not to zero. Models are probabilistic and can still invent numbers during inference. Kepler adds a further check afterward.

He called retrieving the right document complicated. As he described it, the ID system blocks numbers that cannot be resolved. We infer it does not decide which document was the right one. That is our inference, not his claim. The talk also does not say who is blamed when a verified system still errs.