90.5% Build Confidence: What We Measured and How

90.5% of surveyed BNEDai Agent Lab participants reported moderate or extreme confidence to build their next AI agent independently. [1] That number comes from 21 Google Forms respondents out of at least 122 participants who built agents across the first seven weeks of the program. [1][2] This page explains the instrument, the sample, what the number captures, and what it does not.

What was the survey instrument?

The post-session survey was administered through Google Forms. [1] It asked participants to rate their confidence in building their next AI agent independently, on a scale from no confidence to extreme confidence. It also asked them to rate instruction clarity on a 1-to-5 scale and whether they would commit to teaching one other person what they learned. [1][3]

The survey was voluntary. No incentive was offered for completion. Participants received the form after their live session.

What does 90.5% represent?

19 of the 21 respondents selected moderate or extreme confidence. That is 90.5%. [1]

The question measures self-reported confidence, which is a participant's belief that they can perform the task again on their own. It does not measure whether they did perform it again, how long they retained the skill, or whether the agents they went on to build worked correctly. Confidence and competence are different things. This survey measured the first.

What else did the survey find?

Instruction clarity: 4.95 out of 5. Across the same 21 respondents, the average instruction clarity rating was 4.95 on a 5-point scale. [3]

Teach-one commitment: 100%. Every one of the 21 respondents committed to teaching one other person what they learned. [4] This was a yes-or-no question, not a scale. It measures stated intent at the moment of survey completion. Whether those 21 people followed through is a different question that this survey does not answer.

How large is the sample relative to the population?

21 survey respondents out of at least 122 participants who built agents is a response rate of roughly 17%. [1][2] There are several reasons this matters.

First, the 21 who chose to respond may differ systematically from the 101 who did not. Participants who felt more confident after their session may have been more willing to fill out the form. Participants who struggled may have closed the tab. Voluntary surveys carry this bias by design, and no statistical correction was applied.

Second, the 90.5% figure describes these 21 people. It is not extrapolated to all 122 builders, and BNEDai does not claim it represents the full cohort. The Impact Report and the press release both report it as a survey finding from 21 respondents, not as a program-wide statistic.

Third, the 132 source-record registrations are a different population still. Some registrants never attended. The 21 respondents are a subset of the 122 builders, who are a subset of the 132 registrants. [2][5] These three numbers measure different things, and conflating them would misrepresent the data.

How does this compare to the Luma feedback data?

A separate feedback channel ran through Luma's built-in rating system. 14 participants left ratings there: 13 five-star and 1 four-star (Sean Kosofsky, August 3 session). [6] The Luma ratings and the Google Forms survey are independent instruments with different respondent pools, different question formats, and different sample sizes. Some individuals may appear in both. The two cannot be combined into a single figure.

The Luma data is useful for a different question. Where the Google Forms survey measures confidence to build independently, the Luma ratings capture overall session satisfaction. Both are high. Neither tells you what happened after the session ended.

The named Luma reviews provide qualitative texture the survey's numeric scales cannot. Susan Quinn rated 5 out of 5 on August 10 and wrote: "Super great and not as scary as I thought it was going to be." [6] That word, "scary," captures something the confidence scale misses. The 90.5% figure tells you most respondents felt confident afterward. Quinn's comment suggests that at least some of them did not feel confident beforehand, and the session changed that. The survey instrument did not ask a "before" question, so no pre-post comparison is possible from this data. Quinn's review is a single data point, not a trend, but it names the emotional distance the session covered for at least one participant.

What about the session-level data?

Two sessions have detailed engagement records that add context around the survey findings.

The June 29 session, the first one, had 24 invited and 15 built live, a 63% show rate. [7] The July 20 session had 28 registered and 21 attended live, a 75% show rate, with a 53-minute average stay in a 73-minute session and 5.0 feedback across three rating categories. [8]

These session-level figures measure attendance and engagement, not confidence. They are included here because a 53-minute average stay in a 73-minute session suggests sustained participation, which provides some behavioral context for the self-reported confidence scores. But average session duration is not a proxy for learning outcomes.

What does this number not tell you?

The 90.5% confidence figure does not tell you:

- Whether participants retained the ability to build agents weeks or months later - Whether the agents they built after the session performed correctly - Whether confidence translated into action (building a second agent, integrating agents into their work, or changing their team's processes) - How the 101 non-respondents would have answered - Whether participants who encountered barriers (managed-device restrictions, billing questions, trial expiration concerns) rated their confidence differently from those who did not

These are questions for follow-up instruments that do not yet exist. One behavioral signal does exist outside the survey: 9 participants registered for two or more sessions without being prompted to return. [9] Repeat attendance is a different kind of evidence than a survey response. It measures revealed preference, what someone chose to do with their time, rather than stated confidence. It does not tell you why they came back (to build a new agent, to troubleshoot a previous one, or for the community). But it tells you that at least 9 people found enough value in the first session to choose the second.

If BNEDai collects 30-day or 90-day follow-up data, it will be published as a separate evidence page with its own methodology section.

Why publish the methodology at all?

Because a number without its method is a marketing claim. This page is the method behind the 90.5% figure that appears in the Agent Lab Impact Report, on the Impact Report page, and in the press release. Anyone citing that number should be able to find, on this page, exactly how it was collected, from whom, and what it does and does not prove.

For a definition of AI readiness and how BNEDai measures it more broadly, the AI Readiness Index is a separate instrument with its own methodology. The confidence survey and the readiness index measure different constructs: the survey captures post-session confidence for people who already built something, while the index assesses readiness for people who have not yet started.

Methodology

BNEDai administered a voluntary post-session Google Forms survey to Agent Lab participants between June 29 and August 14, 2026. The survey included three items: build confidence (a multi-point scale from no confidence to extreme confidence), instruction clarity (a 1-to-5 numeric scale), and a yes-or-no question on commitment to teach another person. 21 participants completed the form.

The 90.5% figure is calculated as the proportion of respondents who selected moderate or extreme confidence. The 4.95/5 instruction clarity figure is the arithmetic mean of the 21 numeric clarity ratings. The 100% teach-one commitment figure is the count of yes responses divided by total respondents.

No weighting, imputation, or statistical adjustment was applied. The survey was voluntary with no incentive. No third-party audit was conducted. Luma's separate feedback system collected 14 ratings independently, not as part of this survey instrument.

Source citation table

| # | Claim | Source | Date | Link | |---|---|---|---|---| | 1 | 90.5% reported moderate or extreme confidence (19 of 21 respondents) | Google Forms survey, 21 respondents | Aug 2026 | Impact Report | | 2 | At least 122 participants built working agents | Instructor's session records, attested | Aug 15, 2026 | Impact Report | | 3 | 4.95/5 instruction clarity rating; 100% teach-one commitment | Google Forms survey, 21 respondents | Aug 2026 | Impact Report | | 4 | 100% committed to teaching one other person | Google Forms survey, 21 respondents | Aug 2026 | Impact Report | | 5 | 132 source-record registrations | Manual calendar + bnedai.com landing page + Luma | June 29 to Aug 14, 2026 | Impact Report | | 6 | Luma feedback: 13/14 five-star, 1 four-star | Luma feedback export, 14 respondents | June-Aug 2026 | Impact Report | | 7 | June 29: 24 invited, 15 built live, 63% show rate | Session records | June 29, 2026 | Impact Report | | 8 | July 20: 28 registered, 21 attended, 75% show rate, 53-min avg stay in 73-min session, 5.0 feedback | Session records and analytics | July 20, 2026 | Impact Report | | 9 | 9 repeat registrants (2+ sessions, unprompted) | Registration records | June-Aug 2026 | Impact Report |