The Five Rungs of AI Adoption: Why the Bottleneck Is Never the Tool

The five-rung structure below is adapted from a model of Claude adoption by Boris Cherny, creator and head of Claude Code at Anthropic. The rungs are his; the re-basing onto mid-market business operations — where most companies do not ship software at all — is ours.

Most owners we meet are not behind on AI. They are stuck in front of it.

They have a shortlist. They have had the demos. Someone forwarded them a comparison chart, and someone else forwarded them a different one that reached the opposite conclusion. There is a free trial running somewhere that nobody has opened in three weeks. The question "which tool should we use" has been open for eight months, and every month it gets harder to answer because three more tools have appeared.

That is not indecision. That is what happens when you apply a purchasing process to something that is not a purchasing problem.

The Gap Nobody Talks About

Here is the number that reframed this for us.

86% of middle-market firms report AI is partially or fully integrated into their operations. That is from RSM's Middle Market AI Survey, fielded in March 2026 across just over a thousand companies.

But only 13% have actually deployed it broadly across the business, and 84% are sitting on at least one pilot that never made it to production. Those come from Zapier's survey of more than eight hundred mid-market and enterprise leaders.

Read those together and the picture is not "companies haven't adopted AI." Almost everyone has. The picture is that adoption is shallow nearly everywhere. The money is spent. The licenses are active. The tools are in the building. And the work has not changed.

So if you feel stuck despite having bought things, you are not behind the curve. You are exactly on it.

Five Rungs

It helps to stop thinking about AI as a product you own and start thinking about it as a rung you are standing on. There are five, and you can usually place yourself in about ninety seconds.

Rung 0 — Gated. You're blocked before you start. Nothing has been approved, or what was approved is so narrow that nothing real ships. Work happens on someone's laptop and stays there. What holds you here? Approval built for a different era — and no one able to change it.

Rung 1 — Assisted. You and one assistant. One person, one task at a time, everything checked before it goes anywhere. It is genuinely faster than not having it. What holds you here? Your attention. You don't trust it enough to look away — so the work moves only as fast as you can watch it.

Rung 2 — Parallel. You orchestrate several. Several streams of work run at once, and the work checks itself before it reaches you. Your job shifts from doing to directing. What holds you here? Reviewing output. You now read six streams instead of writing one.

Rung 3 — Supervised Autonomy. You manage the managers. Work runs in the background without being started by hand. The question stops being "did you check this" and becomes "what context was missing, and how do we supply it next time." What holds you here? Trust in the loop — and how fast your team can actually decide.

Rung 4 — AI-Native. You steer by intent. Most work is kicked off without you and you monitor by exception. What holds you here? Knowing what to automate, and the right guardrails for each kind.

If rung 0 sounds like you, that is not a failure of nerve. It is usually the accumulated residue of approval processes that were reasonable when they were written. That is worth saying plainly, because a lot of writing on this subject treats rung 0 as embarrassing, and the embarrassment is a large part of why people stay there.

Now Read It Again, Sideways

Go back and look only at what holds a company on each rung.

Approval processes. Your attention. Review capacity. Trust in the loop. Knowing what to automate.

Attention. Trust. Review. Ownership. Decision speed.

Not one of them is a limit of the technology.

That is the whole argument, and it survives scrutiny better than most claims in this space. Across three separate surveys — Zapier's, RSM's, and the World Economic Forum's Future of Jobs report — the barriers leaders name to getting AI into production are infrastructure, integration, compliance, unclear ownership, unprovable return, and skills. Model capability does not appear. Not as a minor entry. It is simply not on the list.

Which means the thing most companies are doing right now is aimed at the wrong target. Comparing tools forever would eventually produce an answer if the tool were the binding constraint. It never does, because it isn't.

The company on rung 3 did not buy better software than the company on rung 0. In a great many cases they are running the same software. They built trust in a verification loop, and they gave someone ownership.

What Standing Still Costs

The gap between rungs is not a percentage. It compounds.

PwC's AI Performance Study, published in April 2026, found that 74% of the economic value generated by AI is being captured by 20% of companies, and that those leaders are generating substantially more value than their competitors at higher margins.

One honest caveat, because this is exactly the kind of number that gets abused: PwC's respondents are primarily large, publicly listed companies. That figure tells you the divide between depth and shallowness is real and widening. It does not tell you what a forty-person firm will earn. Anyone who quotes it at you as your expected return is selling something.

Where We Are, and Why We'll Show You

We are on rung 3, working toward 4. We think you should ask any partner you are considering the same question, and ask them to show you rather than tell you.

So: over twenty-five working days this summer, our own telemetry recorded 97 working sessions, 28,283 individual actions, and 950 separate pieces of work delegated to agents running on their own. Peak day was 192.

The honest framing matters more than the number. That is one workstation — one person's orchestration depth, not a company-wide fleet. Stated accurately it is a stronger claim, not a weaker one: one operator, 950 delegated work streams in a month. That is what rung 3 looks like from the inside, and it is why we are comfortable describing the rungs above it as work rather than theory.

Most firms selling AI to the mid-market are themselves on rung 1. It is a fair question to ask.

The One Next Move

If you take one thing from this: stop shopping and name your rung.

Not because the naming is magic, but because it converts an unbounded question into a bounded one. "Which AI tool should we buy" has no answer — the menu is infinite and grows weekly. "We are on rung 1, and rung 2 requires one person to own a verification loop we trust" has an owner, a scope, and a next step. One of those is a decision you can make this quarter. The other is the reason you have been stuck for eight months.

The next rung is a specific, nameable change. It is almost never a purchase.

Notes for Review — Not for Publication

Both review items are closed.

Attribution — RESOLVED 2026-08-03. Attributed to Boris Cherny, creator and head of Claude Code at Anthropic, in an italic line directly beneath the H1. Verified his title against his LinkedIn profile and press coverage rather than assuming it. The line credits him with the rungs and claims only the mid-market re-basing for Pando, which is the accurate split.

Heading case — RESOLVED 2026-08-03. Founder ruling: headings in Pando-branded written material are Title Case. Applied here. The rule is now written into Marketing/Branding/brand_guidelines.md §5 under "Capitalization" and the brand registry corrected to cite it — the registry had previously asserted sentence case with no source behind it. Scope is Pando-brand only; the other registered product brands carry their own canon. This post is now consistent with the three existing blogs in this folder.

Wiring when the whitepaper lands: this post deliberately does not exhaust the subject — the rung-to-rung transitions, the three stalls in full, and the vendor-rung purchasing checklist are all held back for the gated piece. Add the gated CTA after the "one next move" section when it exists.