The Philosophical Ledger · Agentic AI Strategy · April 2026
80% of enterprises are building AI agents. 6% are getting results.
The gap is the equation: what three important strategy documents reveal when you read them together.
Effective deployable capacity = architecture × governance × knowledge.
The Pattern
Here is what I keep seeing. An enterprise signs a platform deal. Spins up a centre of excellence. Builds RAG pipelines, vector stores, agent prototypes. Shows impressive demos to the board. Gets budget for phase two.
Then nothing scales.
The agents work beautifully in the sandbox. They stall in staging. The governance team cannot sign off because there is no audit trail for agent decisions. The compliance team cannot approve because the agent’s scope is too broad to define clear accountability. The finance team cannot attribute costs because inference spend is buried in a shared cloud bill. The knowledge the agent needs — why this customer gets this treatment, why this exposure connects to that limit — lives in a senior banker’s head, not in any system the agent can access.
The prototype impressed everyone. The production path does not exist.
of enterprise agentic AI projects will be cancelled by 2027 because of escalating costs, unclear value, and inadequate risk controls. Not model capability.
Gartneraverage AI governance maturity across roughly 500 organisations. Governance and agentic controls lag data and technology across every region.
McKinsey AI Trust Survey, March 2026This is not a technology problem. The models work. The platforms work. What does not work is the space between the platform and the production environment — the layers most enterprises have not built.
What Three Strategy Documents Reveal Together
In the last month, three pieces of thinking landed that — read separately — each illuminate part of the picture. Read together, they reveal the whole problem and point toward the equation that governs it.
The finance layer
Sokolov and Meir-Huber apply private equity economics to data architecture. Their core argument: architecture efficiency is a financial variable, not a technical one. Platform choices are capital allocation decisions that should be governed like M&A.
The “clean data for AI” narrative is a consultancy comfort blanket. Machines do not need clean data. They need addressable data: locatable, interpretable, actionable, traceable, and cost-attributable.
From hierarchy to intelligence
Dorsey and Botha’s document is the most provocative corporate strategy memo in years. The argument: the org chart is a 2,000-year-old information routing protocol built for human bandwidth limits.
AI changes that constraint. Block replaced middle management with two “world models” — a continuously updated representation of the company and a real-time picture of customers — and cut 40% of its workforce.
The knowledge layer
McKinsey calls it the semantic layer that turns data into knowledge. Alation calls it the knowledge layer. Gartner D&A 2026 called it critical infrastructure.
The organisations pulling ahead are not the ones with the most agents deployed. They are the ones that built the layer most have not even scoped yet.
The consumer of enterprise infrastructure is shifting from humans to machines. When it does, three things break simultaneously: your economics, your org chart, and your controls.
Most enterprises are fixing them in silos. That is why 80% are experimenting and 6% are getting results.
The Three Layers Most Enterprises Get Wrong
Capital markets have already figured out part of this. The February 2026 “Software-mageddon” — over $800B erased from SaaS — showed value migrating from the application layer, which agents replace, to the layers agents depend on. Databricks raised at $134B during the selloff. The market priced the shift correctly, but incompletely.
Model layer
Foundation models. You are a consumer. This layer commoditises as open source narrows the gap. Value is captured by model providers.
Data platform
Infrastructure providers. You are a customer. Your competitor can sign the same Databricks contract tomorrow. Necessary, not a moat.
Knowledge + governance
Your ontologies, decision history, institutional memory, and agent governance. Cannot be bought. Cannot be replicated. The only layer where advantage compounds over time.
Before agents, institutional knowledge lived in people. Credit officers carried decades of judgement in their heads. Relationship managers knew which counterparty exposures connected across product lines. When humans consumed your data, they brought their own knowledge layer.
Agents do not. They need it architected. For the first time in enterprise history, institutional knowledge must be externalised, structured, and governed as infrastructure. Two banks can sign the same platform contract. They cannot replicate each other’s knowledge layer.
The Equation That Explains The 80/6 Gap
Mathias Frey offered the anchor: capacity scales logarithmically with headcount, shaped by architecture efficiency. AI compresses headcount. Architecture remains the binding constraint. Sokolov and Meir-Huber build their paper on this insight.
Ce = loga(n) × G(t) × K(d)
Effective deployable capacity = architecture × governance × knowledge.
Architecture. Platform efficiency multiplied through headcount. This is where most budget goes. Necessary, but not sufficient.
Governance. The fraction of capacity permitted to operate. A gate between theoretical capability and production capability.
Knowledge. The multiplier from institutional depth: semantic context, decision traces, and the “why” behind work.
The formula reduces to Frey when governance equals one and knowledge equals one. It predicts what Gartner observes.
Organisation A
Excellent platform. Governance maturity at 0.3. No knowledge layer. Effective capacity: 30% of theoretical. Seventy percent of the investment is locked in staging.
This is the group that gets cancelled.Organisation B
Same platform. Governance maturity at 0.8. Deep institutional knowledge. Effective capacity is roughly four times higher than Organisation A.
Same models. Same vendor contract. Different outcome.What The Companies Scaling Have In Common
McKinsey’s strongest finding across eight years of AI research is that the single strongest predictor of enterprise-level AI impact is whether an organisation fundamentally redesigned its workflows. Not model sophistication. Not data estate. Not budget.
They redesigned the work
Not “agent on top of existing process,” but “process rethought around what agents can do.” Replace the routing, not just the router.
They structured institutional knowledge
Ontologies, decision traces, relationship maps, and the “why” behind decisions. This is the investment almost nobody is making.
They built governance into architecture
Runtime policy gates, agent identity, and mandate verification. Not compliance theatre bolted on at the end.
The Dorsey Question
Dorsey’s move demands serious engagement. He did not write a think piece. He cut 4,000 people and restructured a public company. The stock rose 20%. The question is not whether his insight about hierarchy is correct. It is. The question is what determines whether it works for your organisation.
When the model works
Customer signal equals revenue signal. Cost of misrouting is a bad product experience, fixable in the next sprint. The company can ship, observe, and iterate at speed. The world model is the intelligence layer.
When it needs a substrate
Cost of misrouting is enforcement action, capital charges, or reputational damage. You cannot ship-and-iterate with a credit decision. The world model must be paired with a trust model.
The Consumer-Control Matrix
Hierarchy. Quarterly reviews. Analysts interpret. Managers route.
Most banks todayAgents plus static approvals. Agent sprawl. Ghost agents. The 40%.
Where most initiatives landRuntime policy. Dynamic controls. Humans still decide.
Necessary waypointAgents plus runtime policy and knowledge. Trust as infrastructure. Governance at inference.
TargetBlock went Q1 to Q4 directly because the cost of Q3 for a payments processor is manageable. A bank that tries Q1 to Q3 discovers that the cost is not manageable — and retreats to Q1, writing off the investment. The correct enterprise path is Q1 to Q2 to Q4. Build runtime controls first. Then shift the consumer.
The Sequence
What the 6% do
- Build G(t) + K(d)Knowledge layer and runtime controls.
- Redesign workImprove architecture and restructure routing.
- Deploy agentsGoverned, knowledge-grounded, and production-ready.
- ResultEBIT impact that compounds.
What the 80% do
- Buy platformSign deal. Build RAG. Show demos.
- Deploy agentsChase quick wins before controls are real.
- Hit the wallGovernance gap. Knowledge gap. Cannot go production.
- ResultStaging, write-offs, and Gartner’s 40%.
So What
Three things are true simultaneously. Architecture efficiency is a financial variable, not a technical one. Sokolov and Meir-Huber are right. Hierarchy is information routing and AI routes better. Dorsey is right. Workflow redesign is the strongest predictor of impact. McKinsey is right.
What none of them fully see — because none of them are operating inside a regulated enterprise at scale — is that these are not separate insights. They are three consequences of one shift: the consumer of your infrastructure is changing from humans to machines.
The economics, the coordination model, and the controls must change together. Fix one without the others and you arrive in Q3.
The enterprises that compound advantage in the agentic era will be the ones that built the layer no vendor can sell: institutional intelligence — structured, governed, and deepening with every agent deployed against it.
The gap between 80% and 6% is the equation. The question is which side you are building for.