Seen from above: pressed plants on paper sheets, laid out in neat rows, with a hand lens on one sheet.

More than 100 interviews show where people get stuck.

We did not theorize the AI gap; we interviewed our way into it, one conversation at a time. This piece describes that interview program: who we talked to, how we score what they say, and the pain points they described.

Pressed plants, one to a sheet, sorted by kind.

Tanush Jagdish · June 2026

We did not theorize the AI gap. We interviewed our way into it, one conversation at a time. This piece describes the interview program behind our research: who we talked to, how we score interviews, and what participants said hurts. Two companion analyses draw on the same program: The Widening AI Gap reports capability scores for 54 interview transcripts, and The Confidence-Capability Gap compares those scores with how confident each person sounded.

#1How the program started

Olito first built an AI product for regulatory compliance at banks, called Oliver. To sell it, we did what founders are told to do and interviewed the buyers: more than 50 banking and consulting professionals, about their work.

Two things then happened at once. Banks were stuck at the most basic uses of Copilot, and regulatory rollbacks shrank the compliance market we were building for. Meanwhile, in conversation after conversation, participants raised the same question unprompted: how do I use AI better? So we followed that demand, moving from compliance software to teaching professionals how to use these tools, with the same domain expertise and relationships. Interviews that began as customer discovery became the basis of our curriculum and our most durable research asset. Our program has since passed 100 interviews and continues.

#2Who we talked to

Our program spans more than 100 interviews at 47 or more companies. Two analyses from February 2026 draw on it. Our first analysis coded 38 interviews, from 37 people, for pain points, tools, and willingness to pay. On February 24 we re-scored all 54 transcripts then available, from 50 people, with the published rubric; every capability and confidence score on this site comes from that re-scoring. Each figure below names the set it comes from.

Consulting dominates the 54 re-scored transcripts: 21 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. Most participants are based in the United States; others work in the United Kingdom, Spain, the Middle East, and Asia.

From the first 38 interviews we coded 212 distinct pain points, 131 tool mentions, and 105 statements of what participants want from AI, each tied to a specific person and workflow.

#3Pain clusters the same way everywhere

Across firm types and seniority, pain clusters the same way. Tool limitations account for 43% of the 212 pain points and manual processes for 23%, followed by gaps in awareness (16%), output quality (14%), and time drain (5%).

Tool limits and manual work dominate212 pain points · first 38 interviews
Tool Limitations43%

It doesn’t work well enough for serious knowledge work

Manual Processes23%

Tasks that obviously could be automated but aren’t

AI Awareness Gap16%

They sense there’s more possible but can’t see what

Output Quality14%

Answers that need heavy checking or rework

Time Drain5%

Time lost doing it the old way

Most complaints come from people who use AI enough to hit its limits: evidence of progress, not failure.

Exhibit 1: Where the pain concentrates. Shares of the 212 pain points coded from the first 38 interviews (February 2026).

Tool limitations lead because AI does not yet work well enough for serious knowledge work, not because it fails outright. Sessions degrade, documents get misread, and context windows fill up; as one participant put it, the model starts giving back the same regurgitated answer over and over. These are not beginner complaints. They come from people who use AI enough to hit its ceiling. Manual processes, at 23%, are tasks that obviously could be automated and aren’t. Awareness gaps, at 16%, are the most telling: people sense that more is possible but cannot see what.

Specific complaints are practical. Spreadsheet and data work came up most often, followed by slide production and managing context across sessions. Of the 105 things participants said they wanted from AI, 68% concerned efficiency. People are not asking to be transformed. They want the tedious manual part to stop: tracking everything by hand, or exporting from one tool to a CSV file just to upload it into the next.

“There are so many things that I have to download and export as a CSV and then upload into the next one. That really frustrates me.”

a consultant, on daily workflow

#4Tool sentiment splits sharply

ChatGPT was the most-mentioned tool, and 64% of its 36 mentions were positive. Microsoft Copilot drew the highest share of negative mentions, 4 of 12, and one participant said they had stopped using it altogether. Internal firm tools drew mixed reviews, with the occasional exception that improved from useless to useful. Neither Claude nor Claude Code drew a single negative mention, though far fewer participants used them.

Tool sentiment splits sharplyshare of mentions · first 38 interviews
ChatGPT64% positive · 6% negative · 36 mentions

Most mentioned, and mostly liked

Claude50% positive · 0% negative · 10 mentions

No negative mentions

Internal firm tools33% positive · 26% negative · 27 mentions

Mixed reviews

Microsoft Copilot25% positive · 33% negative · 12 mentions

Highest share of negative mentions

A participant on Copilot: “Once the trust is gone, you don’t go back to it anymore.”

Exhibit 2: Positive and negative shares of mentions for the four most-discussed tools, with mention counts, from the 131 tool mentions coded in the first 38 interviews.

“Why am I going to have an army of BAs? Where? Like this. This deck took me nothing.”

a strategy consultant, on Claude Code

#5Many would pay for help

Many participants named a monthly amount they would pay for this kind of help. Among the 37 people in the first analysis, the most common answer was $100 to $300 a month (30%), and 39% named $500 or more. Most preferred a monthly subscription to a one-time fee. One participant put it plainly: they would pay $500 a month to have this set up on their computer.

What participants would pay each monthshare of 37 participants · Feb 2026
24%
30%
0%
14%
14%
11%
$0 to 100
$100 to 300
$300 to 500
$500 to 800
$800 to 1,500
$1,500+

“I would pay you 500 per month for having this installed in my computer.” A participant

Exhibit 3: Monthly amount participants said they would pay, as a share of the 37 people in the first analysis. Another 8% named no amount.

Others expected their employer to pay, which says as much about what firms are not yet providing:

“I would want [firm] to invest in it for me.”

a consultant at a large strategy firm

#6What participants think is at stake

One line from the interviews reframed the company for me:

“I don’t think the thesis is whether AI is going to take your job. I think it is whether I can do AI better than the guy next to me.”

a consultant

Seen that way, the gap matters to each individual, not only to the market.

#7How we score interviews

Each interview is recorded and transcribed. A language model then applies a fixed codebook to the transcript and extracts structured fields: the capability score, a confidence score coded from what the participant says, pain points with intensity and frequency, desired outcomes, tool mentions with sentiment, stories, and willingness to pay. Pairing the confidence score with the demonstrated capability score is what produces the confidence gap.

A companion article, The Widening AI Gap, explains how five weighted dimensions make up the capability score. 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. Scores do shift between model versions, so we compare scores only within one scoring run. Before anything is published, an export step removes duplicates, scrubs names, firms, and internal tool names, and then checks the output for leaks.

Keep two limits in mind. Participants are a self-selected group of consultants, investors, and other professionals: people who agreed to a research interview about AI are not a random draw from the workforce, so read the findings as directional, not representative. And every quote is anonymized: names become [colleague], companies become [firm], and internal tools become “Internal AI Tool.”

#8Check our work

Our scoring rubric, extraction schema, and test-retest results are open source. A methodology page, linked from the Research page as “Method and score calculator,” documents them, so anyone can run the same codebook on their own interviews and compare results.

#9What the interviews changed

Our interviews became the curriculum. Participants’ most common concrete pains, spreadsheet and data work and slide production, are exactly the work Part 2 · Agent shows an agent producing directly. Their top complaint, sessions that degrade as context fills up, is the ceiling Part 1 · Chat teaches readers to recognize. Context management, the need behind many of the rest, is what Part 4 · Personal brain addresses.

#10Related

  • The Widening AI Gap: capability scores for 54 interview transcripts, the three tiers, and three archetypes of what the gap costs.
  • The Confidence-Capability Gap: confidence coded from the same 54 transcripts, set against demonstrated capability.

#11Sources

  • 1.Olito Labs interview program, 2024 to 2026: more than 100 interviews at 47 or more companies. First analysis, February 2026: 38 interviews from 37 people, 212 coded pain points. Re-scoring, February 24, 2026: 54 transcripts from 50 people. 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.

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.