Executive Summary
AI adoption is saturated. 89% of organisations now use it regularly in at least one function. 37% can attribute any earnings impact to it. 6% can attribute more than five points — and both of those numbers are flat year on year, against rising spend.
The gap is not model quality. The companies capturing earnings impact are around three times more likely to have rebuilt the workflow around the AI rather than bolting AI onto the workflow they had. The constraint was never the model. It is whether anyone will make the decision to change the process the model sits inside.
That is the studio’s thesis in one line, and it is the same thesis that governs our operating-partner work: AI does not rescue an undisciplined operation — it brings the breakdown forward. On a governed estate the gains compound. On an ungoverned one, the same tooling industrialises the decay. Discipline is the product; the tooling is downstream of it.
Three findings organise this report. The curve is a data curve, not an AI curve: at every scale the binding constraint is whether the systems of record agree with each other. The interface is the illusion: buyers purchase the visible layer, which is cheap, and skip the data and integration layers, which are slow, expensive and decisive. And the mid-market is falling behind on the part that matters: above $1B revenue, 40% of organisations are scaling agents inside at least one function, up from 27%. Below $1B the figure is flat at 22%, and the gap is widening.
McKinsey State of AI 2026 · 1,719 respondents, 97 countries · attribution by executives, not audited accounts.
The Interface Illusion
Ask an executive to describe their AI programme and they will name the interface — the copilot, the chatbot, the agent. It is the only part of the estate anyone can see, and it is the cheapest part. The programme lives or dies underneath it.
Top — cheap, fast, visible
Bottom — slow, expensive, decisive
Every enterprise AI estate is the same seven layers. What changes with scale is how many of them are deliberate rather than accidental. Failed programmes buy layer 7 and skip layers 1 and 2, then conclude the model was not good enough.
A Worked Example from the Public Record
Heineken’s estate, as described in the group’s own reporting and a Microsoft customer story published in December 2025, is instructive precisely because the sequence ran bottom-up. A group data platform connecting more than 70 operating companies came first. A build layer took five years and 7,500 internal builders, producing over 10,000 applications and 42,000 automation flows on top of SAP as the system of record. Governance was retrofitted under pressure — tiered environments, enforced sharing limits, deployment pipelines — and became the thing that made scale possible. The agents arrived last, and are the only part anyone describes.
The number worth stealing is not an AI number. A product team of five administers more than 8,000 environments. That is a governance result — and it is the reason the AI layer can exist at all.
The commercially useful version of the insight: every buyer describes their programme by naming the interface. The first diagnostic question is always what sits underneath it.
The rest of the paper opens here, with your work email.
6 more chapters. One email opens every paper on this site.
Why the Returns Don’t Arrive
Four-fifths of users report feeling more productive. Two-thirds of organisations cannot find the productivity in the accounts. That is not a measurement failure. It is what happens when time is saved inside a process that was never rebuilt.
The saved hour returns to the process that produced it. It is absorbed by the next task in the same queue, and never converts into fewer people, faster cash or better price. The mechanism is the same one that produces the delivery discount in vertical software: value absorbed invisibly into an unchanged structure. The accounts measure what was spent; the loss sits in what was absorbed.
The 6% who do capture earnings impact are not distinguished by better models or bigger budgets. They are roughly three times more likely to have fundamentally redesigned the workflow, and three times more likely to be scaling agents inside a single function rather than across a demo. The differentiator is operating discipline, and it is external, published evidence — not this firm’s opinion.
| Measure | 2026 | Read |
|---|---|---|
| Regularly use AI in at least one function | 89% | Saturated. Adoption is no longer a differentiator |
| Scaling AI across the enterprise | 44% | Up from 38%. Spreading sideways |
| Report improved individual productivity | 80% | People feel faster |
| Attribute any EBIT impact | 37% | Flat year on year. Finance does not feel it |
| Attribute more than 5% of EBIT | 6% | Flat, against rising spend |
| $1B+ firms scaling agents | 40% | Up from 27%. Accelerating |
| Sub-$1B firms scaling agents | 22% | Flat. The gap is widening |
| Say AI operating costs constrain deployment | 20% | Token spend is now a budget line |
Fielded 4 May to 8 June 2026; 1,719 respondents across 97 countries, weighted by contribution to global GDP. Attribution by executives, not an audit of accounts — which is the evidentiary standard of the entire category, and part of why the earnings number is stuck. A margin movement that survives a quality-of-earnings read is the bar this firm builds to, and the bar the delivered figures in Chapter 6 are classed against.
The Diligence View
For the reader who diagnoses rather than owns: an AI estate is now a diligence item, and most estates fail the same way. Hours-saved dashboards with no baseline. Adoption metrics self-reported by the people who built the thing being measured. Token spend with no forecast. Shadow tools carrying customer data. And no answer to the classification question the regulation now requires. The finding usually arrives in a sale process — the one moment the seller cannot act on it. The same analysis, run two years earlier, is a different instrument.
Ask the management team which P&L line the AI programme moved. The pause is the finding.
The Curve by Scale
What actually gets deployed, how it fails, and where the money is — by stage. Companies below roughly $1M buy courses and templates, not engagements; the curve starts where a data layer becomes possible.
| Stage | What is deployed | How it fails | Where the money is |
|---|---|---|---|
| PRIORITIZE $1–5M · 10–25 people | Consumer AI seats, a notetaker, AI features inside tools already owned. No data layer, no owner | Tool sprawl with no system of record; a data-protection exposure nobody has priced | Proposal and content velocity; deflecting the easiest third of support. Shows up as the founder’s evenings, not as margin |
| PRODUCTIZE $5–15M · 25–60 | The first custom build — and the first encounter with the real constraint | The pilot works and never scales, because the data it needs sits in four systems that do not reconcile. Misdiagnosed as a model problem; the vendor gets changed | Cost-to-serve in support and sales-cycle compression. The first tier where a number can be defended |
| OPTIMIZE $15–50M · 60–150 | The platform decision — the one that sets the next five years of cost | Three teams on three stacks: no shared auth, observability or evaluation. The first serious token-spend surprise | The first tier where AI can move a margin line rather than produce an anecdote — and the first with enough money to waste |
| CATEGORIZE $50–150M · 150–500 | Agents move from adjacent to the workflow into the workflow | Governance arrives as a brake and the shadow stack returns — or the platform is excellent and unadopted, because the operating model never changed | Headcount avoidance in high-volume functions; cycle time on revenue-critical paths. Engagements underwritable against a P&L line |
| SPECIALIZE $150M–1B · 500–5,000 | Citizen development at scale, held by a small central team | Measurement theatre: dashboards of self-reported hours that never reconcile to the P&L | The governance model itself — harder to achieve and worth more than any single agent in the estate |
| CAPITALIZE $1B+ · 5,000+ | Everything above, plus proprietary data as the differentiator | The distance between centre and edge: an excellent group platform while an operating company still runs the month on a spreadsheet | Only two lines reliably reach EBIT: cost-to-serve in functions carrying real headcount, and pricing. Everything else is absorbed productivity |
The commercial seam sits at the Optimize and Categorize stages: revenue, a real data problem and a budget — but no platform team and no governance model. Above them, the large caps build their own. Below them, nobody can fund it. That seam is who this studio builds for.
What Changes Underneath: The Next Three Years
Generation cost has collapsed. Verification cost has not. Everything in this chapter follows from that asymmetry — and it is a forecast, labelled as one.
The craft of writing code is declining; the craft of building software is compounding. As generation approaches free, most defects stop being coding defects and become intake defects — the system built the wrong thing correctly. Which moves the scarce work upstream: written intent, priced boundaries, acceptance criteria. The intake is the product.
Review moves the same direction. Line-level human review is already uneconomic in greenfield agentic work; in regulated vertical software it survives longer, then consolidates upward. The gate does not disappear — it moves to architecture, invariants and acceptance, where a human owns intent and machines verify execution. Hand-written unit tests decline; machine-run verification against executable specifications grows. The cost of correctness relocates; it does not vanish.
“Good code” survives as an economic property, not a style. Entangled code burns model context and widens blast radius: agents pay for structure in tokens what humans paid in hours. And delivery cost shifts from payroll toward compute — one organisation in five already reports token spend as a deployment constraint.
| Now | Through 2027 | 2027–2029 |
|---|---|---|
| Line review dying in greenfield agentic work; verification the cost centre; intake defects dominate | Agent fleets standard under thin senior direction; the economic definition of seniority converges on holding intent under pressure | Agents verify agents; artifact-level human review uneconomic; the gate consolidates at architecture and acceptance |
| Ungoverned 2025–26 adoption compounding invisibly | Comprehension debt surfaces: velocity decay, incident rates, failed technical diligence | Remediation tooling improves sharply; the cost of fixing bloated estates falls |
The dated prediction from the companion report stands: enterprise codebases bloated on undisciplined AI adoption reach a structural bottleneck by approximately mid-2027 — not as a cliff, but as velocity decay, rising incident rates and failed diligence, surfacing roughly twelve months behind the 2025–26 adoption wave.
One honest consequence cuts against our own interest, so it belongs in print: models are getting materially better at repairing what earlier models bloated. The window in which disciplined remediation is scarce — and therefore expensive — runs roughly 2026 to 2028. After that the fix gets cheaper. What never compresses is the organisational half: decision rights, boundary pricing, accountability for the number, and the governance that makes an estate auditable. Install the discipline while the fix is scarce; keep the discipline because it is the part no model erodes.
The Governed Model
Any team can ship AI-written code fast. Very few can ship code the next hire will maintain, the next auditor will sign off on, and the next model release will not render obsolete. The difference is five guard-rails and a gate.
Why the Category Sells the Wrong Layer
The adoption gap persists for a structural reason, not a technical one: almost everyone selling into it is paid on the wrong layer. Model vendors monetise seats and tokens — their unit of success is usage, not earnings impact. Implementation agencies monetise builds and pilots — layer 7 is the fastest thing to demo and the easiest to invoice, while the data and integration layers stretch timelines and thin margins. Consultancies are paid for the roadmap, and then for the next roadmap. Nobody is acting in bad faith; everybody is acting rationally.
Nobody in the category is compensated on whether the workflow was redesigned. That is why 89% became 37% became 6% — and why the numbers are flat while spend rises.
The disciplined approach inverts the economics. A defined scope and an end date, so there is no annuity to protect. Data and integration before agents, because that is where the earnings sit. Governance built to stay with the client, not with us. And a gate on everything that ships, because the estate has to survive the next auditor and the next model release — not the demo.
| Failure mode | Guard-rail |
|---|---|
| Models recommend deprecated infrastructure | A curated golden-path infrastructure context |
| Teams default to the most expensive model | Tiering rules with spend visibility |
| Agents re-implement functions that already exist | Enforced reuse and token-efficiency gates |
| Knowledge stays locked in long-tenured staff | An AI-mapped codebase; every change carries its plan |
| Ungoverned, invisible LLM spend | One managed control surface for models, prompts and tools |
Sequence matters more than tooling: data and integration first, governance before scale, agents last — the order the public-record example in Chapter 2 ran, and the reverse of the order failed programmes buy. Every output that reaches production passes a gate; this report attaches no per-engineer multiple and no operator-to-agent ratio, because cadence is the number a buyer can verify.
| Engagement | Outcome |
|---|---|
| Regulated trading platform (firm engagement) | €600,000 a year of planned team expansion avoided, against a €1M+ annual budget (realised, cost avoided against a plan) |
| Prior operating engagement (group unnamed) | Flagship platform modernised and released at month 18 after three failed rebuild attempts (realised); release cadence 6x against the sector norm — two production releases a month, four development releases a day (operational); monthly hosting down ~80% on a comparable load, on a greenfield product (operational) |
The European Overlay: Attribution, Obligations, Funding
Three things a European deployer’s advisers keep asking for in one place: how to make the impact attributable, what the regulation already requires, and where the subsidy sits. This chapter is written to be forwarded. Regulatory and funding positions are stated as at September 2026 and verified per country at engagement.
Making It Attributable
The reason two-thirds of organisations cannot find AI in the accounts is that nobody set up the accounts to find it. The method is not sophisticated; it is merely disciplined. A baseline measured before deployment, per use case, on one named P&L line. An owner for each deployment, by name, accountable for that line. The line measured, not the hours — self-reported time saved is testimony, and finance does not book testimony. And a quarterly review that retires what moved nothing.
Hours returned is a feeling. A margin movement has a line number.
The Obligations Already Running
The AI Act’s duties arrive in phases, and the 2026 omnibus agreement moved some of them — which is precisely why a dated summary is worth quoting. As at September 2026:
| In force | What it requires |
|---|---|
| Since Feb 2025 | Prohibited practices banned; the AI-literacy duty — staff using AI systems must be trained for the use, and it applies to deployers, not only vendors |
| Since Aug 2025 | General-purpose model obligations on providers; national authorities and the penalty framework in place |
| Since 2 Aug 2026 | General application, and the Article 50 transparency duties: users told when they are interacting with AI, synthetic content labelled. Enforcement powers live — fines to €35M or 7% of turnover at the top of the scale |
| Deferred by the omnibus | Substantive high-risk obligations moved to Dec 2027 (Annex III) and 2028 (product-embedded). Deferred is not cancelled: implementation runs 12–18 months, so the clock is effectively now for anyone running AI in hiring, credit or essential services |
Two regimes sit alongside and do not defer: GDPR still governs the data every model touches — legal basis, impact assessments, transfers — and the Data Act has applied since September 2025. The first compliance task is not paperwork; it is classification. An estate map that cannot say which systems fall where under the Act is the same accidental estate that cannot find its earnings impact. One exercise answers both.
The Money on the Table
European deployers fund AI adoption twice — once in the P&L and once in taxes already paid — and most claim only the first. The routes, in the order a mid-market company should check them:
| Route | What it provides |
|---|---|
| European Digital Innovation Hubs | Subsidised advisory and test-before-invest across every EU region, co-funded nationally — the free first step before any build |
| Digital Europe deployment grants | 50% of eligible costs on standard grants, up to 75% on SME-support topics; mostly consortium-based, with AI-focused calls running through 2027 |
| AI Factories | Subsidised GPU compute for SMEs and startups, no consortium required |
| National R&D instruments | Rolling tax credits on development work — examples in force: 36% payroll credit (NL), 25–35% (DE), 30% plus a subsidised AI diagnostic covering 40% of its cost (FR). Every market differs; check before the project starts, not after |
One rule governs all of it. The grant should fund the layers nobody can see — the data foundation, the integration, the governance — and the audit that maps them. A grant that buys another interface subsidises the 37%. Funding follows the same sequence the estate does.
The Starting Move: The Estate Audit
Do not begin with a model comparison or a pilot. Begin by exposing what the estate actually is. It takes a day.
Map everything currently running against the seven layers of Exhibit 2, and mark each layer deliberate or accidental. Then answer three questions in writing. What sits underneath each interface anyone names? Do the systems of record agree with each other — and if not, where do they disagree? And for every workflow an agent is meant to enter: who owns redesigning that workflow, by name?
Most companies can complete the layer map in an afternoon and cannot answer the third question at all — which is itself the finding. An estate with an accidental data layer and no workflow owner will convert AI spend into the 37%, not the 6%.
Alternate entry — for AI and studio readers, this paper stands in for step 01.
Precision beats power. Alignment beats scale. Truth beats urgency.
About the figures. Every figure carries one of three classes on this report’s claims: delivered (verifiable against records), modelled (calculated against a stated baseline), or external (published third-party data, source named). The classes are never summed. Market figures in Chapters 1, 3 and 4 are external and self-reported by their respondents; the revenue bands in Exhibit 4 are a modelled mapping onto the published stage definitions. Delivered figures in Exhibit 7: the two firm engagements are attributed to Re-In-Essence; the prior operating engagement is stated without naming the group, its sponsor or its customers. Chapter 5 is forecast. Chapter 7’s regulatory and funding positions are stated as at September 2026, summarise the published instruments without constituting legal or tax advice, and are verified per jurisdiction at engagement. Full bases: a separate evidence note, available on request.
A Reinessence Group company.
Sources. Market data: McKinsey, The State of AI 2026 (fielded 4 May–8 June 2026; 1,719 respondents, 97 countries). Estate example: Heineken N.V. annual reporting; Microsoft customer story, December 2025. Code structure: GitClear, The Maintainability Gap: AI Code Quality in 2026 (623M analysed changes). Developer productivity: METR, July 2025 (arXiv:2507.09089) and 2026 revision. Scaling stages after Acquisition.com’s published $100M scaling roadmap, re-expressed; revenue bands modelled, not sourced. Regulation: Regulation (EU) 2024/1689 (AI Act) and the 2026 omnibus agreement on high-risk applicability dates; GDPR; the Data Act. Funding: Digital Europe Programme work programme and EDIH network materials; national scheme parameters as published August–September 2026.
Important information
This report is for general information only. It is not investment, financial, legal or tax advice, and it is not a recommendation about any particular company. Market figures are published third-party data, cited as their publishers report them. Modelled figures illustrate stated assumptions; they are not forecasts. Results from past engagements do not predict results elsewhere. Nothing in this report is a promise or guarantee of any outcome. Take advice from your own advisers before acting on it. Full terms.