
Confidence in AI runs far ahead of capability.
We scored 54 interview transcripts twice: once for demonstrated AI capability, and once for how confident each person sounded about AI. Confidence averaged 66 out of 100 and capability 37, and confidence ran higher in 53 of the 54. This article examines that gap: what it measures, where it comes from, and what closes it.
At low sun, a small tree casts a shadow many times its own height.
Tanush Jagdish · February 2026
In February 2026 we scored 54 interview transcripts, from 50 professionals, on two measures. A capability score rates what each person demonstrably does with AI. A confidence score rates how confident they sound: it starts at 50 and moves up or down with signals in the interview, such as enthusiasm, an early-adopter identity, anxiety, or skepticism. Confidence averaged 66 out of 100 (median 67); capability averaged 37 (median 38). This article examines that gap. A companion piece, The Widening AI Gap, covers the capability distribution, and Market Research: 100+ Interviews describes the wider interview program behind both.
Most people in this group sound ahead on AI. They are ahead of most workers, and far behind what the tools can already do.
#1Confidence runs well ahead of capability
These professionals talk about AI with energy: 45 of the 54 transcripts show unprompted enthusiasm, 44 describe experimenting freely, and 31 show an early-adopter identity. Measured capability tells a different story.
Average score; median 67.
Average score; median 38.
Same transcripts, both scored 0 to 100. Confidence starts at 50 and moves with what people say; capability starts at 0 and rises with demonstrated skill.
Confidence ran above capability in 53 of the 54 transcripts, and the median gap per person was 29 points. Both measures do move together (correlation 0.65): people who can do more also sound more confident. Almost everyone, though, sounds more confident than their demonstrated skill supports.
Most of these people are not beginners. Nearly all use AI at least weekly (27 transcripts describe daily use, 23 weekly), which already puts them ahead of most workers (see “Ahead of most, behind the frontier” below). Adoption is not proficiency, though, and the difference between using AI and using it well is where this gap lives.
An executive AI trainer who coaches Fortune 500 leaders gave the clearest illustration of the gap:
“I was working with the SVP of digital tech for a Fortune 250. He rated himself an 8 out of 10. An hour in, he stopped me and said, “I think I’m actually a 2 out of 10. I didn’t even realize what I didn’t know.”
an executive AI trainer who coaches Fortune 500 leaders
That exchange shows the mechanism. Overconfidence here is not arrogance; it is a horizon problem. People calibrate against what they have seen work, and few in this group have used the agentic tier: 39 of the 54 knew such tools existed before the interview, but only 12 had used one. So they judge themselves against chat, where they are comfortable, and feel ahead. Measured against the whole ladder, the typical score is closer to a 4 out of 10.
#2Reading the size of the gap
Same people, same 0 to 100 scale: the space between the two distributions is the gap.
Both scores run from 0 to 100, but they are built differently, so read the size of the gap with care. Capability starts at 0 and rises with demonstrated skill. Confidence starts at 50 and moves with what people say, so part of the distance comes from that design.
Its pattern does not depend on that design. Of the 24 transcripts with a confidence score of 70 or more, half score below 20 on automation and agentic work. Among the 31 transcripts scoring below 40 on capability, 22 expressed enthusiasm about AI and 21 said they experiment freely.
These results also hold up to reasonable changes in the sample. Dropping the one workshop transcript, or counting each person once, leaves the median gap at 29 points.
#3Confidence tracks the tools people use
Scoring the three tiers separately shows what confidence follows. Confidence rises with chat and prompting skill more than with automation skill: its correlation is 0.53 with Everyday Chat and 0.59 with Advanced Prompting, against 0.36 with Automation & Agentic. People judge their readiness by the tools they use every day, and for most of them those tools stop at chat and prompting.
“Most people are still on Copilot. 90 to 95% of the time they don’t want to become AI experts. They want this to operate the same as when they had to learn email.”
an executive AI trainer who coaches Fortune 500 leaders
Most people want AI to work like email: learned once, then invisible. Climbing past chat takes deliberate practice that few choose to put in, so their self-image stays anchored to the tiers they know.
#4What recalibrates confidence
Pain points offer a first clue. Tool limitations, the largest category among the 212 pain points coded from the first 38 interviews, come from people who have used AI enough to hit its limits: sessions that degrade, context that fills up, answers that repeat themselves. Nobody is frustrated by a context window they have never filled. Market Research: 100+ Interviews analyzes those pain points in full.
Seeing the next tier at work is what moves people. When participants saw a multi-step task get done, rather than a question get answered, their mental model changed quickly, and their language was strikingly consistent. They said it felt like cheating, that they had been introduced to a whole new world, and that it was the same as giving instructions to a business analyst, except that this one is a robot. Moving from “AI answers questions” to “AI does work” is a conceptual shift, not a technical one.
That same exposure took the SVP above from an 8 to a 2 within an hour. A gap like this is not a verdict on anyone’s intelligence. It measures how much room is left, and seeing what is on the other side is where closing it starts.
#5Ahead of most, behind the frontier
Both statements hold at once.
Ahead of most. In Pew Research Center’s September 2025 survey, 21% of U.S. workers said at least some of their work is done with AI. In Gallup’s survey for the fourth quarter of 2025, 49% of U.S. employees said they never use AI in their role. Among the 15,000 employees in EY’s 2025 Work Reimagined survey, 88% used AI at work, mostly for search and summarization, and only 5% used it in ways that transform their work. Against that baseline, this group’s confidence is earned.
Behind the frontier. In McKinsey’s 2025 global survey, 88% of organizations used AI in at least one business function, yet only 23% were scaling AI agents anywhere in the business, with another 39% experimenting. In BCG’s 2025 AI at Work survey, 13% reported agents integrated into broader workflows. Among the people in this study, the median automation and agentic score is 12 out of 100. What these professionals do with AI sits far from what is now technically possible.
#6Related
- The Widening AI Gap: the capability distribution for the same 54 transcripts, the three tiers, and three archetypes of what the gap costs.
- Market Research: 100+ Interviews: the wider interview program, how interviews are scored, and the pain points participants described.
- Part 1 · Chat: its mental-model section addresses the same calibration failure that leads people to treat AI as an oracle.
#7Sources
- 1.Olito Labs primary research: 54 interview transcripts from 50 professionals, scored February 24, 2026 for demonstrated capability and for confidence coded from interview signals. Primary.
- 2.Executive AI trainer who coaches Fortune 500 leaders: expert interview, February 2026. Primary.
- 3.Pew Research Center, “About 1 in 5 U.S. workers now use AI in their job, up since last year,” October 2025 (survey fielded September 2025). Context.
- 4.Gallup, workplace AI use survey, fourth quarter 2025: 49% of U.S. employees never use AI in their role. Context.
- 5.EY, Work Reimagined Survey 2025: 15,000 employees in 29 countries; 88% use AI at work, mostly for search and summarization; 5% use it to transform their work. Context.
- 6.McKinsey, The State of AI in 2025, November 2025: 88% of organizations use AI in at least one business function; 23% are scaling AI agents and 39% are experimenting with them. Context.
- 7.BCG, AI at Work 2025, June 2025: 13% report AI agents integrated into broader workflows. Context.
More from our research.
Working on the same question?
We look for people to interview, data to compare against, and teams willing to test what we find. Leave a note and we will write back.