McKinsey published their validating ladder study in mid-2026, showing how degrees of AI adoption show up in valuation multiples. It closely maps to the Talbot West onion model of AI adoption. The onion is the better framework for understanding where value lies.
A ladder has one axis. An onion has two: how deep your AI runs, and how well you built it at that depth. The second axis is the one nobody prices and the one that decides whether the first ever pays.
The rungs and the layers, side by side
Picture your company as an onion. The outer layer is people using AI in their own work. Under that sits AI wired into systems and processes more than one person depends on. At the core is AI designed into your capabilities, your products, and your business model. The further in it runs, the more weight it carries.
What McKinsey found
They looked at 471 private-equity-backed companies across 31 industries and 30 countries, all acquired in deals from 2023 onward.
The jumps between levels are uneven. Going from scattered individual use to a reworked operating model barely moved the multiple, 13x to 14x. The premium showed up when AI started changing what the company sells. Companies using AI broadly carried multiples about 130% higher than companies using it in one function.
Where the onion goes further
- It looks forward.McKinsey reports what buyers have already paid. We built the onion to guide a build before the value exists.
- It reads two axes instead of one.Depth is the first. How well you build at that depth is the second, and it decides whether the first one pays.
- It covers the companies McKinsey did not look at.That sample is large, leveraged, and headed for an exit, and it sits well ahead of the rest of the economy. Most of the companies deciding what to build next look nothing like it.
The pattern outside private equity
MIT's Center for Information Systems Research built a four-stage maturity model of its own, with no reference to McKinsey's ladder or our onion. It grades companies on growth and profit against their industry average rather than on what an acquirer will pay. In a 2025 survey of 152 companies, the two shallow stages came in below industry average and the two deep stages well above.
Money spent, ground lost.
Three models, built separately, measuring different things, pointing the same way.
What a multiple cannot see
The clearest evidence sits in manufacturing. Using Census Bureau data on US manufacturers, four researchers found J-curve returns to industrial AI. Performance drops first and the gains come later. A one-standard-deviation increase in AI use came with 1.33 percent lower total factor productivity in the short run.
Who took the hit is the interesting part. The losses concentrated in older plants, and among those, walking away from structured production-management practices explained about a third of the damage. Companies with growth-oriented strategies got off lighter. Early adopters grew faster later on. Depth did not decide those outcomes. What the company did around the technology decided them.
So the level you sit on matters less than how you sit there.
Every layer can be built well or badly
Each layer has a version that pays and a default that quietly does not. The default is always the easier path, which is why companies end up there without deciding to.
Tool sprawl: a dozen AI subscriptions, scattered usage, nothing measured, company material sitting in tools nobody vetted.
Knowledge management. Point AI at your own documents and data, get answers grounded in your material instead of the open internet, and people stop hunting for things they already own. McKinsey says the market barely rewards this layer. Your own P&L will.
The pilot that dodges production. It demos well because it avoids the data problems, the permissions, the edge cases, and the human handoffs waiting on the other side.
This layer has what a model needs around it: clean data plumbing, evaluation that catches regressions, humans in the loop where judgment belongs, and a straight account of what the system can and cannot be trusted with.
MIT's NANDA researchers name brittle workflows and misalignment with how the work actually runs as the causes of the drop-off, not weak models.
Magical thinking. The roadmap gets treated like a pilot, the proprietary data is assumed to exist in usable shape, and nobody budgets for how much the operating model has to change.
These layers rest on a real data advantage and an operating model rebuilt to feed and govern the capability. At the far end, the new AI-native business runs as its own company instead of a feature bolted onto the old one.
A study of roughly 2,000 companies from 2010 to 2018 tied AI investment to growth in sales, employment, and valuation, driven by product innovation rather than cost cutting. The 20x and 31x rungs are the market noticing, years later.
Value at every layer
The outer layer pays for itself when you build it on purpose, in recovered hours and sharper decisions, long before an acquirer re-rates you. You do not have to reach the core to get paid.
A product-layer build on weak foundations destroys more value than a workflow-layer build ever could. It costs more, and it fails in front of more people. Sequence by dependency and let ambition wait its turn. A company that masters the outer layers arrives at the inner ones with the data, the discipline, and the habits those layers demand.
How we work the layers
Before anything gets built, we prioritize. We use a method we call APEX, which weighs candidate opportunities on stakeholder urgency, revenue impact, technical feasibility, implementation complexity, and strategic alignment. The point is to build the thing most likely to pay off rather than the thing easiest to demo.
- 08-A / Outer
We configure tooling against your own knowledge and measure what comes back.
- 08-B / Middle
We build the harnesses that let a pilot survive production.
- 08-C / Core
We get honest about the data and operating-model dependencies before the money goes out.
Integration makes a capability possible. It does not make it free. Naming the tradeoffs early is what keeps the effort pointed at the payoff.
The question worth asking
Everyone wants to know which level to aim for. The better question: at the level you already occupy, are you building well or coasting on the default? Look at where your AI lives today and ask how it got there. The answer shows up in your own financials long before it shows up in your multiple.
