A tall rock on the shore at low tide. Mussels and barnacles crowd its lower half, and only a few cling near the top.

Most regular AI users score in the lower half.

I scored 54 interview transcripts from consultants and investors on what they can actually do with AI. Nearly all use AI at least weekly. A few have built systems that multiply their output, while most have plateaued at chat. That split runs through the regular users themselves, and it widens at every tier of skill.

On a tidal rock, mussels and barnacles crowd the lower half and thin out toward the top.

Tanush Jagdish · February 2026

In February 2026 I scored 54 interview transcripts, from 50 professionals at strategy firms, the Big 4, and venture capital and private equity funds, on demonstrated AI capability. Nearly all of them use AI at least weekly, and half use it daily. A small group has built systems they say multiply their output several times over; most have plateaued at basic chat. This article covers how capability is distributed and the three tiers behind it. A companion article, Confidence in AI runs far ahead of capability, compares the same capability scores with how confident each person sounded, and More than 100 interviews show where people get stuck describes the wider interview program that both analyses draw on.

This gap does not separate people who use AI from people who don’t. It runs through the group that already uses AI regularly, and it is widening.

#1Most regular users score in the lower half

Most participants work in consulting: 21 transcripts come from the three largest strategy firms (known as MBB), 11 from boutique firms, and 6 from independent consultants. Another 4 come from the Big 4, 4 from venture capital and private equity funds, and 8 from research, in-house, public-sector, and banking roles. I scored each transcript from 0 to 100 on demonstrated capability, using a published rubric, and separately on three tiers, from everyday chat to agents that run tasks on their own.

Most transcripts score in the lower half54 transcripts · Feb 2026
4
3
16
8
8
9
3
3
0
10
20
30
40
50
60
70
80
90
Median 38Upper half 15 of 54Top score 78

Only 3 of 54 transcripts score above 70: this gap sits inside the user base, not at its edges.

Exhibit 1: AI capability score in 10-point bands across 54 transcripts. Median 38; 15 of 54 score above 50; the highest score is 78.

Most transcripts land in the lower half: 39 of 54 score 50 or below. Most of these people use AI for one thing, everyday chat: asking questions, drafting emails, summarizing documents. Then they stop. Their organizations would call many of them “AI-forward.”

Economic historian Paul David documented the same pattern with electricity. Factories had electric motors for decades before productivity rose, because the gains required redesigning work around the new technology. AI has arrived the same way: the tools are in place, and the redesign of work around them has barely begun.

Scores run from 5 to 78, and the middle half sits between 22 and 52. Only 3 transcripts score above 70. That distance between the median of 38 and the top scores separates very different ways of working, as the archetypes below show. This gap sits inside the user base, not at its edges.

#2Three tiers, and the split widens at each

A composite score shows that most people land in the lower half. It does not show why. Scoring the three tiers separately shows where people stop.

Everyday Chat is where confidence lives. Scores sit highest here, with a median of 55, and for most people “using AI” begins and ends with chat.

Advanced Prompting is where people separate: the median falls to 45, and scores spread from 0 to 82. Some write detailed instructions, ask for structured output, and keep context across sessions; others have not tried any of it. Prompting well takes deliberate practice, so people who put in that time pull away, and people who don’t stay at chat.

These tiers form a ladder, not a menu. Each depends on the one below it: you cannot automate work you have not learned to prompt well, and you cannot run agents without the scaffolding that automation provides. Stalling at any rung cuts off everything above it.

Capability falls at each tiermedian by tier · 54 transcripts
55
45
12
Everyday Chat
Advanced Prompting
Automation & Agentic35 of 54 score below 20

Each tier depends on the one below it, so missing one stalls every tier above.

Exhibit 2: Median tier score across 54 transcripts: 55 on Everyday Chat, 45 on Advanced Prompting, and 12 on Automation & Agentic, where 35 of 54 score below 20.

Beyond prompting, the frontier is nearly empty. Automation, connected tools, command-line agents, and autonomous workflows are where productivity should compound, yet 35 of the 54 transcripts, about two-thirds, score below 20 on this tier. One outlier, the System Builder described below, built an automated system that connects deal sourcing, CRM, and due diligence.

Capability at this tier can also move quickly. One participant, the Skill Builder below, went from no command-line experience to advanced use in a few weeks of structured practice. Because the tools keep improving, people who practice keep compounding while people who wait stand still. Anyone who misses one tier stalls on every tier above it.

#3What the gap costs

This separation shows up in deals reviewed, hours spent, and projects that fail. Three patterns came out of the interviews, and each turns on one question: did the person restructure their work around the tools?

The System Builder: general partner, venture capital fund

5×deal throughput, his estimate
ChatPromptAuto + Agent

Works across all three tiers. His system is only half connected, and his team can’t yet replicate it.

The Skill Builder: incoming associate, MBB firm

80%time saved
ChatPromptAuto + Agent

Started with zero command-line experience. Climbed the ladder deliberately in a few weeks of structured practice.

The Wary Expert: healthcare strategy consultant

0her firm’s AI-scoped projects on time
ChatPromptAuto + Agent

Wrong answers and blocked tools taught her caution, so she stayed with chat. When her firm priced in automation-tier productivity, the manual work fell to her.

Exhibit 3: Three archetypes, each with the tiers the person works in, as described in the interview.

A general partner at a venture capital fund is the System Builder. He built an automated pipeline that connects deal sourcing, CRM, and due diligence, and his first-day diligence now matches what used to take a full manual process.

“We can probably see five times as many deals.”

a general partner at a venture capital fund

His estimate implies roughly five times the deal throughput. His system is only half connected, though, and his team cannot yet replicate it.

An incoming associate at an MBB firm is the Skill Builder. He started with no command-line experience. Over a few weeks of structured practice, he built reusable AI instructions, memory that persists between conversations, and workflows that span several sessions. By his account, he produced in one day a deliverable that normally takes an experienced associate a full week, and senior colleagues with years of domain expertise could not match his throughput. They had the knowledge but not the tools.

“I am 99% sure no one at the firm will be using it that way.”

an incoming associate at an MBB firm

One day instead of five is an 80% time saving on a real client deliverable.

A healthcare strategy consultant is the Wary Expert, and her caution was earned. Her first chatbot gave her wrong information. Her firm later blocked that tool over data concerns, and some clients did not want AI used on their work. She stayed with chat, and she wrote one shared prompt so her team’s research stayed consistent. Then her firm sold projects scoped to AI-accelerated timelines, and the gap landed on her team: automation-tier deadlines with chat-tier tools. Every AI answer needed checking, so she absorbed the manual work herself and burned out.

“All the projects sold relying on AI in the proposal… they imploded.”

a healthcare strategy consultant

By her account, none of those projects was delivered on the AI-scoped timeline.

Each archetype follows the tier structure. Working across all three tiers, the System Builder saw his output multiply because each tier built on the last. Over a few weeks, the Skill Builder climbed the same ladder on purpose. The Wary Expert was given no ladder to climb: her tools had failed her, and when her firm priced in automation-tier productivity, the manual work fell to her.

Such differences add up to role compression. When one person reviews five times the deals and another saves 80% of the time on a real deliverable, their distance from everyone else grows by multiplication, not addition.

Between these extremes sits the majority. One participant named it:

“I am the silent middle. If it is there, I will use it. If it is not there, you know.”

a consultant

Comfortable with chat, aware that more exists, and waiting for someone else to build the path: that is the most common position in this sample, and the most precarious. What keeps people there is a lack of investment, not a lack of access or intelligence.

Nobody has to stay there. With no structured program from his firm (no participant described one), the Skill Builder closed the distance from chat to command-line tools in a few weeks on his own. Closing the gap is possible, but it takes an investment that almost no one is making.

Other participants described similar builds. One consultant built a pipeline that reads 50 earnings transcripts and produces a comparative analysis, turning a two-day team task into a 15-minute run. Another set up an agent that scans regulatory publications daily and surfaces only what matters to a client. A third automated the path from raw interview notes to cited first drafts, cutting drafting time by 80%. Each had used only basic chat six months before the interview.

#4What this study does not capture

Data governance limits adoption, and for good reason. Most firms in the sample restrict what proprietary data may go to external AI tools. Client data, deal terms, and internal analyses belong behind access controls, and no productivity gain justifies compromising them. So the real question is how to build governance that enables adoption instead of blocking it: firm-approved tools where they exist, providers that meet enterprise standards (SOC 2 Type II, data-processing agreements, zero data retention), and clear rules for which data goes where. Most firms have not started on that design problem. AI governance is a data infrastructure problem describes what that layer looks like.

AI agents are confident when they are wrong. Every participant who moved past basic chat ran into this. Higher up the ladder, people depend more on AI output and find it harder to verify; several said checking AI work took almost as long as doing it by hand. Better tools change the shape of the verification burden without removing it. Part 1, Push chat to its ceiling covers this in its sections on the mental model and on verification.

Participants chose to take part. These 54 transcripts come from 50 self-selected consultants, investors, and other professionals who agreed to talk about AI; four people were interviewed twice, one transcript records a workshop rather than an interview, and each transcript is scored separately. Its scores describe this group, not the wider population, so read the findings as directional. Dropping the workshop or counting each person once moves the headline numbers by a few points at most: the median score stays at 38, while the highest score falls from 78 to 72.

Scores also depend on the scoring model. On the model used for this re-scoring, the same ten interviews scored about 12 points lower than on the earlier model used for the reliability check described below. Compare scores within one scoring run, not across runs.

One snapshot cannot measure change over time. My claim that the gap is widening rests on the tier structure and on individual trajectories like the Skill Builder’s, not on repeated measurement of the same people.

Tools also change. Features that needed a command line a year before these interviews now have graphical interfaces, so the tier boundaries will shift. These results describe the gap as it stood when the transcripts were scored in February 2026. Its structure, with each tier building on the last, is likely to last; the specific distances may not. Agents, for example, no longer need a command line: Anthropic made Claude Cowork generally available in its desktop app in April 2026, and OpenAI launched ChatGPT Work in July 2026.

#5How the score works

Each transcript’s capability score (called the AI-Native Score in the published methodology) measures demonstrated capability from 0 to 100. Self-assessment plays no part: every component rests on evidence from the interview, such as tools named, workflows described, and integrations built.

The AI-Native Score0 to 100 · demonstrated, not self-rated
Highest Capability Tier
30%
Usage Sophistication
25%
Tool Diversity
20%
Self-Directed Learning
15%
Integration Depth
10%
Tier codebook
0

No demonstrated capability

25

Everyday chat only

50

Advanced prompting, persistent context

100

Automation & agentic systems

Open-source protocol. Test-retest on an earlier model version: 10 interviews scored in five independent runs of the same model, ICC(3,1) = 0.92.

Exhibit 4: Five weighted dimensions (highest capability tier 30%, usage sophistication 25%, tool diversity 20%, self-directed learning 15%, integration depth 10%) and the codebook that scores the tier dimension at 0, 25, 50, or 100.

Composite scoring separates people that a single tier label would lump together. Among the 25 transcripts whose highest tier was Advanced Prompting, scores ran from 22 to 72; sophistication, tool diversity, and integration depth made the difference. A one-question “which tier are you?” measure would have missed that.

Scoring runs on an open-source pipeline: a language model applies a fixed, published codebook to each transcript. In a test-retest check on an earlier version of the scoring model, the same model scored 10 interviews in five independent runs each, and the runs agreed closely: intraclass correlation ICC(3,1) = 0.92. A methodology page, linked from the Research page as “Method and score calculator,” publishes the codebook, rubric, and reproducibility results.

#6Where this goes

For each person, the choice is the same: invest in climbing the tiers, or watch the gap widen. No participant described a structured program for making that climb, so Olito built one. Our Learn curriculum follows the same principle that each step builds on the last, moving from Chat to Agent, Skills, and a personal Brain, and one-on-one coaching adds eight weeks of guided practice. Coaching is by referral or invitation; olitolabs.com/learn has the interest form.

#7Sources

  • 1.Olito Labs primary research: 54 interview transcripts from 50 professionals, scored February 24, 2026 with the published rubric. Primary.
  • 2.Olito Labs, “Scoring the AI gap,” methodology and open-source scoring protocol, February 2026. Test-retest check on an earlier model version: 10 interviews scored in five independent runs of the same model, ICC(3,1) = 0.92. Primary.
  • 3.Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” American Economic Review, 1990. 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.