Webinar · 120 minutes

15% versus 3%

Three serious studies measured the return on artificial intelligence over the same pair of years and arrived at 95%, 85%, and 80%. None of the three lied. This session shows why the numbers diverge by seventy-five points, and hands over the three questions that settle the next statistic to land on your desk.

What the session delivers
120minone session
3yardstick questions
1live exercise
1public case, with its source

Every number cited comes with who measured it, on whom it was measured, and who paid for it — including the ones that hold up this session's own argument.

The problem

Three sources. Three answers. The same question.

This is not a debate between good research and bad research. All three are defensible, all three were published by serious institutions, and even so the distance between the first and the last is seventy-five percentage points. Understanding why is more useful than picking a side.

95%
MIT Media Lab · Project NANDA · 2025

of pilots with no measurable impact on P&L

The test is the ledger: the money has to show up. It is the hardest yardstick of the three, and MIT does not sell artificial intelligence.

85%
Lopez Research · Enterprise AI Benchmark · 2025

of companies struggling to find ROI

Not "there was no return" — "we could not measure it". This is the most honest yardstick of the three: it admits its own difficulty instead of forcing a number.

80%
Anthropic · State of AI Agents · 2026

of organizations with measurable ROI

More than five hundred technical leaders at companies that have already deployed agents. Those who tried, failed, and gave up are not in the sample.

The instrument

Three questions settle it, and they hold for good

You do not need to know the studies to take them apart. The same three questions work on any statistic that lands on your desk after the session, including the ones you have already used.

First

What was measured

Impact on the ledger and return self-reported by those who deployed are not the same test. They are different questions, not different answers.

Second

On whom it was measured

Surveying satisfaction only among those who are still customers always gives a good result. Survivorship bias is not bad faith, it is how the sample was cut.

Third

Who paid for the research

Interest is not fraud, but it is a fact about the fact. It does not invalidate the study; it invalidates using the study alone.

And then the yardstick turns on the session itself

Applied to the number I use

The figure that holds up the title — 15% versus 3% — goes through the same three questions in front of the room, with the funder's interest stated out loud.

The limits, stated in the session

What this session does not do

A hundred and twenty minutes do not replace months of work. Saying so at the opening is part of the method, not modesty.

Does not deliver

A ready-made cost dashboard

The session shows why seeing the cost is hard — a thousand times more consumption on an agentic task than on a chat, and thirty times of variation between two runs of the same task. Building that visibility is a project, not a slide.

Does not promise

That 15% is a good number

Five times more chance of a return is still little chance. Anyone who leaves thinking they installed a dashboard and solved the problem understood the opposite of what was said.

Format

A hundred and twenty minutes, live, with an exercise in the middle

It works for a group from a single company or for an open audience. The four-minute exercise, in which each person applies the three questions to a statistic they themselves have repeated, is what changes the conversation afterwards.

Length

120 minutes

It fits into a board-level calendar window, without requiring half a day to be blocked.

For whom

Those who decide, not those who implement

No technical background required. What it requires is having to defend or challenge a number about AI in front of someone.

Where it leads

To the masterclass, if that makes sense

The webinar hands over a yardstick for reading numbers. The Executive Judgment Masterclass takes on the next decision: how far AI can go in each process.

What stays

A question to take with you

Which AI metric did you institute this year, and which of your own values is it contradicting right now?

Has your board already repeated one of these numbers?

A hundred and twenty minutes, and the next statistic that reaches the table is read with a yardstick. Tell me whether it is for you or for your team and I will reply with the possible dates.

I want to attend this webinar
Before or after this session