A number without its source

In January 2019 I kept a note called “Concise Pitch.” It opens with the eight questions George Heilmeier used to judge research proposals when he ran DARPA in the 1970s:

• What are you trying to do? Articulate your objectives using absolutely no jargon.
• How is it done today, and what are the limits of current practice?
• What is new in your approach and why do you think it will be successful?
• Who cares? If you are successful, what difference will it make?
• What are the risks?
• How much will it cost?
• How long will it take?
• What are the mid-term and final “exams” to check for success?

Under the questions, the note says: “If you know a number, say it. Restate vague sentences with specific facts.”

I answered the first three. “Apres automates AI training.” Then: “Today AI training is a manual process involving data labeling and maintaining the surrounding infrastructure internally. This is expensive, technical and custom.” The third answer was four words long, and the other five questions, including the exams, stayed blank.

I took the advice about numbers to heart, and used plenty of them. What I didn’t always keep was where each one came from. My own files show two of them drifting.

22% and 300%

In November 2018 my recruiting script told candidates: “Enterprise data has the problem of being 22% inaccurate and growing 300% YOY.”

The 22% is on the statistics page I kept that winter, with its context: the average share of inaccurate data had risen to 22% from 17% in a year. The closest thing to 300% on the same page is a survey line. Of the people asked, 27% expected their data volume to grow by somewhere between 75% and nearly 300%. That is an expectation held by about a quarter of respondents. By the time it reached my script, it was a growth rate for all enterprise data.

85%

The same statistics page quotes a Gartner prediction: through 2022, 85% of AI projects would deliver erroneous outcomes because of bias in data, in algorithms or in the teams managing them.

Two years later, in January 2021, the company overview we used at Apres said: “Today, when AI fails, we all lose and models fail 85% of the time.” The draft of our seed announcement that month said that even successful models “will have an 85% instance of error.”

A prediction about projects had become a present-tense fact about models. Nobody set out to mislead anyone. Each step was a small rounding, made by someone in a hurry who wanted the sentence to land. Some of those someones were me.

It isn’t only me

In June 2025 a trade headline said enterprise AI agents fail 65% of multi-turn tasks. The number traces back to CRMArena-Pro, a benchmark from Salesforce AI Research. In the paper, 35.1% is the multi-turn success rate of the best of nine models in its B2B setting. The median model succeeded on 23.7% of those jobs.

So the 65% was the best case. The typical model did worse. Excellent published a close read of the paper on November 3, 2026, with every number tied to the table it came from, because the headline had kept the figure and dropped the context.

The rule I use now

Every number has a source, and the source travels with it. In the plan for my own posts, written in November 2026, every draft carries an Evidence field naming the file each number came from. Before I repeat a number, I ask four questions:

1. Who measured it, and when?
2. Is it a measurement, a forecast or a survey of what people expect?
3. Whose number is it: the best case, the median or one customer?
4. What did the sentence I’m about to write leave out?

The same rule is most of what Excellent does for agent work. An agent’s “done” is a claim with no source attached. The receipt is the source.

Heilmeier’s last question is the one I left blank in 2019: what are the exams that check for success? I would answer it first now.