Ten percent of mid-market firms have scaled all of their AI initiatives past the pilot stage (IT Brief UK, 2026). The other ninety percent are buying the same tools, from the same vendors, at roughly the same price.
That is the part worth sitting with. If the tool produced the return, the returns would not disperse like that.
A new NBER working paper gives the dispersion a name. Babina, He and Jiang trace early AI productivity gains not to the technology but to organization capital โ the durable, firm-specific knowledge a company accumulates through learning-by-doing (NBER, 2026). The tool is the occasion for the gain. It is not the asset.
The Same Investment, Two Different Decades
"Canaries in the Gold Mine: Early Productivity Gains from Artificial Intelligence Creating Organization Capital" was issued in August 2026 by Tania Babina, Alex Xi He and Renhao Jiang (NBER, 2026).
Their instrument is the interesting part. Rather than counting licences, vendor spend or self-reported adoption โ all of which measure distribution โ they build a firm-level measure of AI investment from AI-skilled employment, spanning machine learning through generative and agentic AI. Who a firm hires is harder to fake than what a firm says it is doing.
Run against that measure, the headline result is a contrast rather than a coefficient:
AI investments are associated with productivity growth in recent years, but not over the previous decade.
Same measure. Same class of technology. Two different answers.
The authors then do the work most adoption research skips: they ask what changed. Their answer is the accumulation of organization capital โ and to test it they construct a second novel measure, built from workers' job descriptions, to identify which roles actually build that capital. The productivity gains, they find, are driven by AI-skilled jobs that build organization capital. Not AI headcount in general. Not tooling in general.
One caveat I will state before anyone builds a slide on it: the working paper body is gated, and I have verified the design and the direction of the findings, not effect sizes. Nobody should be attaching a percentage to this study yet โ including me.
Why the previous decade paid nothing is not settled by the paper, and I will flag my own reading as inference rather than finding: the machine-learning wave largely bought prediction that sat beside the workflow โ a score, a forecast, a dashboard someone consulted. The current wave sits inside the workflow, which is the only position from which a tool can generate learning about how the work is actually done. Prediction adjacent to a process does not teach you the process.
What "Organization Capital" Actually Means
The term is not new and it is not soft. Eisfeldt and Papanikolaou formalised it in the Journal of Finance: organization capital is a production factor embodied in a firm's key talent, carrying firm-specific efficiency, with a claim on the firm's cash flows. Firms holding more of it earned average returns roughly 4.6% higher than firms holding less (Eisfeldt and Papanikolaou, 2013).
In operational language: it is the accumulated knowledge of how your work gets done. Which exceptions matter. Which customer signals predict churn. Which step in your close process is the one that actually breaks.
A licence gives you a general-purpose capability. Organization capital is what happens when that capability is repeatedly pointed at your specific work and the result is retained somewhere other than one person's head.
The first is purchasable. The second is not.
Why Your Peer Bought the Same Stack and Built No Organization Capital
Hold the NBER finding against the mid-market evidence and the pattern stops being abstract.
Klarus surveyed 500 senior decision-makers at UK and Ireland firms between ยฃ200M and ยฃ2B in revenue, 300โ3,000 employees. Seventy-three percent had partially or fully deployed AI. Ten percent had scaled all initiatives past pilot. Ninety percent were stalled or still early (IT Brief UK, 2026).
The same survey contains an inversion worth reading twice: 91% said they were confident in their internal AI expertise โ while 48% named a lack of AI expertise as the main reason their projects failed to advance. The same population is simultaneously sure it has the capability and sure the capability is missing.
Underneath both: 83% reported poor data quality, and 69% said poor data had prevented or delayed AI work outright.
CLA's Heartbeat Index found the same shape on the other side of the Atlantic. Across 722 small and mid-market organisations, 69.9% believed their workforce had the skills needed for future success. Only 49.5% reported any positive efficiency improvement from recent technology or AI investment, and just 20.2% reported highly positive impact (CLA, 2026).
Two independent mid-market samples. Both report capability present and results absent.
That combination rules out the diagnosis most firms are funding. If skills were the binding constraint, the population under study would report a skills shortfall. It reports the opposite. What is missing is not fluency โ it is the mechanism that turns fluency into something the firm owns.
Pilots Don't Fail. They Fail to Deposit.
Valliance surveyed 1,000 senior leaders at large European businesses and found that 40% of AI initiatives remain pilots by design โ never killed, never scaled. In organisations with mature, established AI programmes, that figure rises to 48% (Consultancy.uk, 2026).
More maturity, more perpetual pilots. That is not what a learning curve looks like.
The cost is in the same dataset. Firms stuck in pilot report strong ROI 20% of the time. Firms that scale report it 76% of the time. The stuck cohort also waits longer for value โ 6.6 months against 5.9 (Consultancy.uk, 2026).
Read that through the NBER lens and the gap is legible. A pilot that ends without changing a documented process, a role definition or a decision rule has produced a result and deposited nothing. The team learned. The firm did not. Run twelve of those and you have twelve experiences and zero organization capital โ which is precisely the configuration that shows AI investment with no productivity growth attached.
The Deposit Test
Make it arithmetic. A 200-FTE services firm runs six AI pilots this year at roughly two months of part-time effort each โ call it a full FTE-year of senior attention, before licences. Standard reporting asks what each pilot saved. Ask instead how many produced a durable artefact: a rewritten SOP, a maintained prompt-and-evaluation set, a decision rule now used by people who never touched the pilot.
If the answer is one, you converted about two months of that year into organization capital and spent the rest on tuition. The pilots were not failures โ five of them were simply uncapitalised. That distinction does not appear anywhere in a standard ROI deck, which is exactly why it keeps getting funded again.
Valliance's own framing is that experimentation is not the problem; what happens next is. The NBER paper explains why that sentence is an accounting statement rather than a slogan. Learning-by-doing only compounds if the learning is captured in something durable.
The Honest Counter
Four limits, stated plainly.
Association is not causation. The NBER result is an association between AI investment and productivity growth, mediated by a constructed measure of organization capital. Firms that build organization capital well are plausibly better-run in ways that also make them better AI adopters. The paper's design is careful; it is not a randomised trial.
Both novel measures are proxies. AI investment inferred from AI-skilled hiring will understate firms that buy capability rather than hire it. Organization capital inferred from job descriptions will understate firms whose knowledge lives in systems rather than in role text. Proxies of this kind get direction right more reliably than magnitude.
The mid-market corroboration is regional and self-reported. Klarus is UK and Ireland; CLA is a US client panel. Both are self-report. They agree with each other, which is meaningful, and neither is a national probability sample, which matters.
Organization capital has an exit door. This is the sharpest objection, and Eisfeldt and Papanikolaou raised it thirteen years ago: organization capital is embodied in key talent, and that talent holds a claim on its returns. Knowledge that lives in three senior people is firm-specific right up to the week one of them resigns. Codification is not bureaucratic overhead here โ it is the difference between an asset and a retention risk.
What to Decide This Quarter
Four moves. None of them require a new vendor.
- Score every live pilot on what it left behind, not what it sped up. One question per pilot: name the process document, decision rule, or role definition that changed because of it. If the honest answer is "the team got faster," you funded an experience. Valliance's 40% is made entirely of experiences.
- Rewrite one AI job description around codification. The NBER result runs through roles that build firm-specific knowledge. Most AI-adjacent job descriptions in the mid-market specify tools. Specify the deliverable instead: documented workflows, maintained evaluation sets, decision rules others can run.
- Fix the data foundation before the next seat. 83% poor data quality is not a side note, it is the reason learning-by-doing cannot accumulate โ an agent that cannot reach your history cannot learn your firm. That is a Q4 line item competing directly with licence expansion, and it should win.
- Audit where your organization capital is embodied. List the five processes AI has measurably improved this year, then name where each improvement is stored. Every one that resolves to a person rather than an artefact is a productivity gain with a resignation letter attached.
Your competitor can buy your stack this afternoon. They cannot buy the eighteen months your team spent learning which exceptions matter โ unless you left it undocumented, in which case they can simply hire it.
Organization capital is the only part of your AI investment that is actually yours. Decide this quarter whether you are building it, or just renting the occasion to.
