The platform demonstration is excellent. Identity works, retrieval is fast and the evaluation dashboard is green. Six months later, the second workflow still needs a new permissions model, different evidence standards, another review queue and a separate owner. What, exactly, was reused?

The answer cannot be “the platform”. That describes a node without proving an edge. The strategic claim is stronger: a capability already operating for one bounded purpose has lowered the marginal cost, elapsed time or operating risk of a named second decision population. That claim is testable. It can be supported by delivery records and operating evidence, or rejected when context, rights, controls and service conditions fail to transfer.

An adjacency graph replaces the feature roadmap with a network of evidence-bearing relationships. Capability nodes include governed identity, context retrieval, evaluation assets, tool contracts, audit evidence, escalation routines and bounded action controls. Consumer nodes are decision populations, not application names: resolving an account query, preparing a credit file, investigating an alert or authorising a payment. Each edge states which asset transfers, under what contract, to whom, with what economics and within which failure boundary.

This keeps business value in view. A reusable retrieval service has no adjacency until a workflow owner can show that it improves a decision or outcome and can capture the benefit. Teece's analysis of profiting from innovation emphasises the role of complementary assets around technology.1 The inference here is practical: the model or service is only one part of an operating capability, and possession of that component does not establish downstream value.

01 · Graph semanticsDefine the node, consumer and edge before drawing the arrow

A capability adjacency exists when a proven operating prerequisite transfers to a new decision population through a stable, authorised and measurable contract. “Proven” means exercised under representative conditions. “Operating” includes people, process, controls and recovery as well as software. “Transfers” means the new consumer can use the asset without recreating its meaning, permissions or acceptance evidence. “Measurable” means the portfolio can compare the later consumer with a credible first-consumer or independent-build baseline.

A capability node must therefore have an owner, version, supported purpose, service conditions, failure modes and change path. Governed identity is not merely sign-on; it can include actor, role, delegation, purpose and decision authority. Context is not a database; it is selected, permissioned, provenance-bearing material whose meaning survives use. Evaluation is not a dashboard; it is a portable case set, failure taxonomy, acceptance threshold and regression routine. Bounded action is not tool access; it is a contract linking allowed act, preconditions, approval, record, fallback and kill path.

A consumer node starts with a decision population. It describes who or what decides, the cases encountered, consequence, current baseline, affected people, service level and benefit-capture owner. This prevents a graph from connecting vague labels such as “finance copilot” and “customer agent”. The same interface may serve different decisions with incompatible evidence. Conversely, different interfaces may consume one genuine capability contract.

An edge is directional. If credit-file preparation can reuse customer identity and document provenance, the reverse is not automatically true. It is also versioned. A change in data rights, model, input distribution, control owner or service level can expire the receipt. The graph should display only accepted and deliberately tested edges, while speculative links remain in a separate hypothesis register.

The receipt is compact because it points to evidence rather than replacing it. It names the consumer decision and baseline; identifies the reused asset and version; describes interface stability, context transfer, rights and portable tests; records isolation, fallback and exit; measures incremental cost and time; and assigns an owner and expiry. The receipt can be audited after delivery to see whether the promised reuse occurred.

Receipts also make rejected edges reusable knowledge. If context failed to transfer, record the exact semantic or jurisdictional mismatch rather than the verdict alone. If a service objective conflicted, preserve the demand and recovery evidence. Another consumer may fall inside the proven boundary, while a future capability revision may repair the mismatch. This produces a disciplined hypothesis backlog: rejected routes remain searchable evidence, but they do not appear as accepted dependencies or inflate the business case.

Milgrom and Roberts model complementarities among technology, strategy and organisation, cautioning against treating technology as an isolated change.3 Empirical work by Bresnahan, Brynjolfsson and Hitt associates information technology with workplace organisation and skilled labour in a system of complements.4 These sources do not propose this graph. They support the narrower inference that reusable technology may require coordinated organisational assets before its downstream economics appear.

02 · Levels of reuseCopied code is not yet an operating capability

Teams commonly report reuse when they copy a repository, call the same foundation model or buy seats on one platform. These acts may reduce a narrow engineering task. They do not show that context, evidence or operating accountability transfers. The distinction matters because most delay and risk can sit outside the model call: agreeing purpose, securing rights, mapping exceptions, creating tests, integrating identity, training reviewers, monitoring failures and maintaining a fallback.

The first rung is copied code. It may save implementation hours, yet each consumer owns a fork and future changes diverge. The second is a shared technical service with a supported interface and service objective. It can amortise operation, but consumers still need local evidence. The third is transferable control and evaluation: permissions, logs, cases and acceptance routines retain meaning across a defined class of consumers. The fourth is an operating capability whose technical service, control path, staff routine, economics and change process have all been exercised by later consumers.

Movement up this ladder should be claimed only for the scope proved. A portable red-team harness does not prove portable business acceptance tests. A shared audit event schema does not prove that each decision retains sufficient evidence. A retriever returning provenance may transfer technically while document rights do not. Each rung can be valuable; the error is representing lower-rung reuse as if it had already removed downstream discovery and operating work.

Teece, Pisano and Shuen describe dynamic capabilities through integrating, building and reconfiguring competences in changing environments.2 The relevant portfolio implication is not that every shared service becomes a dynamic capability. It is that repeatable integration and reconfiguration require an organisational routine. An edge receipt, compatibility test, consumer registry and controlled retirement path are such routines. They convert one project's undocumented memory into evidence that another team can inspect, challenge and, where conditions match, deliberately reuse.

The false-platform control

Consider a central retrieval and agent platform surrounded by twelve proposed workflows. Its diagrams show abundant arrows, yet no workflow names a decision population, current baseline, benefit owner or accepted service boundary. The platform has capabilities and interested teams, but the graph has no valid consumer edges. Demand expressed as workshop attendance, prototype requests or licences allocated is not operating demand.

This is a negative control because the centre can be technically excellent. The absence of edges does not imply that the platform is bad. It means the portfolio cannot yet claim adjacency economics. The appropriate work is consumer discovery: define a bounded decision, expose its current cost and consequence, and test exactly which platform assets remove work. A platform team should not convert untested aspirations into an internal revenue forecast.

False adjacency also appears through naming. Two workflows both “summarise documents”, but one prepares a low-consequence internal brief while another assembles regulated evidence for a consequential decision. Their source authority, omission tolerance, record retention, reviewer competence and appeal implications may differ. Surface task similarity does not establish transferable context. Meaning and rights are part of the capability.

03 · Incremental economicsMeasure the second consumer against a credible counterfactual

The financial promise of adjacency is amortisation. A first consumer funds discovery, integration and controls that later consumers should not fully repeat. The edge receipt tests whether that promise survives. Record the later consumer's incremental engineering, data, evaluation, legal, risk, change, training and operating work. Compare elapsed time from authorised start to controlled operation. Track new shared-service costs and local residual work. The result may be cheaper, faster, safer, some combination of the three, or none.

Do not divide a platform budget by the number of ideas in a roadmap. Only validated consumers belong in the denominator, and their consumption should be measured. Nor should the first use case carry every shared investment forever. Separate sunk discovery, genuinely reusable assets, consumer-specific work and continuing shared operation. The portfolio can then decide whether to recover cost, subsidise exploration or retire an underused node without manufacturing unit economics.

The useful curve is empirical rather than guaranteed. The first consumer may be expensive. A second in the same rights, context and service family may arrive faster because identity, contracts, tests and routines transfer. A later consumer with incompatible jurisdiction or latency can move the curve back up. Learning creates an option, not a law. Pindyck's treatment of irreversible investment under uncertainty supports staged commitment when information can change the investment decision.9

Risk reduction belongs in the economic record even when it is not converted into a single monetary estimate. Evidence might show fewer unowned dependencies, faster detection, a rehearsed fallback or narrower incident scope. State the unit and comparator instead of assigning a confidence-free price to avoided harm. The edge can be justified by lower cost, shorter time or better-controlled exposure, but the sponsor should see which claim carries the decision and who captures the benefit. This prevents operational resilience from becoming an unexamined plug in a platform return calculation.

Brynjolfsson and Hitt argue that organisational transformation and intangible investment accompany information-technology returns.5 Cohen and Levinthal's absorptive-capacity argument connects prior related knowledge to recognising and applying new knowledge.6 Applied carefully, these ideas explain why a later workflow may benefit from trained operators, codified cases and experienced owners. They do not justify assuming that all later consumers share the same context.

04 · Worked graphA regulated-finance spine with bounded consumers

Take a financial institution that has proved four capabilities for a customer-service decision: governed identity, provenance-bearing policy context, a portable evaluation harness and bounded action controls. The next proposed consumers are account-query resolution, commercial-credit file preparation and payment-exception investigation. These are not one “financial-services agent”. Their decisions, consequences, rights, latency and owners differ. The graph begins with separate consumer definitions.

For account-query resolution, identity can transfer if the actor, customer, account, channel and delegated authority retain the required meaning. Policy retrieval can transfer if sources, jurisdiction, effective date and citations are supported. The evaluation harness may reuse routine and adversarial cases, but acceptance thresholds must reflect the query consequence. A bounded adjustment tool needs limits, approval and a complete record. The edge passes only for the scope that these contracts cover.

Credit-file preparation can reuse some identity and evidence-provenance assets without inheriting customer-service authority. Its consumer is an analyst preparing an official file, not an assistant deciding approval. Credit documents, policies, periods, entities and exceptions require new cases. The evaluation harness transfers as a mechanism and schema; many test instances do not. The operating receipt measures whether file preparation becomes faster or more complete without shifting review and evidence requests downstream.

Payment-exception investigation may reuse the identity and audit-event spine while rejecting the retrieval service's latency or availability objective. It may demand stricter segregation of duties, transaction-specific evidence and a fallback that operates under outage. If the shared service cannot offer isolation, an incident in customer queries could disable payment operations. The right graph may reuse selected contracts and decouple execution.

This mixed result is a strength. It prevents the portfolio from choosing between two slogans, “one platform” and “every team builds its own”. A capability can be shared at one layer and isolated at another. The graph exposes where reuse actually changes delivery and where heterogeneity remains economically or operationally decisive.

The counter-case is therefore specific. Suppose the institution's second and third bounded workflows both reuse the same actor-role-purpose identity claims, provenance schema, evaluation runner and release evidence. Delivery records show that teams avoid recreating those assets, local acceptance cases plug into the runner, consumer faults remain isolated and elapsed time falls against comparable independent work. Operators use one change route and can invoke their own fallback. Those observations validate two edges. They do not prove that a fourth decision, a different jurisdiction or an automated action boundary will transfer.

05 · Strongest rivalIndependent systems can be safer and cheaper than forced sharing

The strongest rival says each use case should remain independently engineered. Shared services impose coordination overhead, queue consumers behind a central owner and create systemic failure. A security defect, model regression or identity outage can spread across every connected workflow. Consumers with different service objectives must negotiate compromises. Local teams may move faster when they control the whole stack and its change calendar.

This rival is correct under several conditions. Decouple when contexts are semantically different, rights do not transfer, acceptance cases are not portable, service objectives conflict, or the shared dependency creates an unacceptable common-mode failure. Decouple when coordinating the shared contract costs more than maintaining local capability. Decouple when a consumer cannot exit or fall back without waiting for a central team. A graph with fewer honest edges creates more decision value than a dense speculative map.

Technical debt reinforces the warning. Sculley and colleagues describe entanglement, hidden feedback loops, undeclared consumers, data dependencies and changes in the external world as sources of machine-learning system debt.14 The adjacency graph should expose rather than hide dependencies. Every accepted edge increases the importance of versioning, consumer registration, compatibility testing and a controlled way to refuse or retire consumption.

Central capability also creates political congestion. The shared team can become an internal gatekeeper with priorities unrelated to the consumer's outcome. A stable interface helps, but ownership still matters. Publish the supported contract and change window, allow consumers to observe service health, and make escalation explicit. If a consumer needs a permanently exceptional configuration, separation may be clearer than an expanding central service full of flags.

There is also a legitimate exploration argument for duplication. Two independent approaches can reveal which abstraction deserves to be shared; premature convergence can freeze the first team's assumptions into every later design. Time-boxed duplication is sensible when evidence is weak and learning value is explicit. It becomes waste when teams unknowingly repeat settled work or when no one compares the results. The portfolio should name the experiment, retain both outcome records and decide whether to converge, preserve diversity or close the route.

Contain the blast radius before celebrating the graph

An accepted edge needs isolation at four levels. Demand isolation prevents one consumer exhausting capacity. Data and rights isolation prevents context crossing boundaries. Change isolation uses versioned contracts, staged release and consumer-specific acceptance tests. Decision isolation ensures that a failure cannot silently expand authority. The consumer also needs a safe fallback, a kill path, recovery ownership and enough evidence to reconstruct affected actions.

NIST's AI Risk Management Framework organises work across Govern, Map, Measure and Manage, including context, risk measurement, monitoring and response.11 Its playbook provides suggested actions rather than a universal checklist.12 For adjacency, the inference is that reuse cannot mean inheriting another team's risk judgement. Each consumer maps its context and consequence while reusing only evidence and controls that remain applicable.

A circuit breaker is more than a shutdown button. It defines the signal, authorised actor, scope and safe state. A retrieval-quality alarm may route one consumer to manual search while leaving others available. An identity-integrity event may require a broader stop. A model change may suspend only consumers whose acceptance cases fail. The graph should therefore carry dependency class and isolation boundary, not just an arrow.

06 · Executable instrumentMake the next edge earn its line

The Adjacency Edge Test evaluates one proposed relationship, not the entire platform. The team enters evidence for the consumer, asset, contract, context, rights, isolation, incremental economics, portable tests, owner and exit. Each field can be absent, conflicting, testable or proven. Conflicting transfer conditions direct decoupling. A complete proof set supports reuse. Partial but bounded evidence creates a test. An unnamed consumer or unproved asset produces no edge.

Run the test before allocating a consumer to the shared roadmap, then rerun it after controlled delivery. The first run authorises evidence acquisition. The second compares promised and observed cost, time and risk. Keep both receipts. If later consumers repeatedly fail the same contract, the capability boundary may be wrong. If they pass with stable economics, strengthen the node and its operating capacity rather than merely adding features.

07 · Governance cadenceValidate, measure, isolate, strengthen or retire

Review the graph as an operating portfolio, not an annual architecture poster. Every proposed edge begins as a hypothesis with a consumer owner and a discriminating test. Accepted edges record observed economics, service health and expiry. Changes to purpose, context, rights, interface, model, control or service level trigger impact analysis. Consumer-specific regressions can suspend one edge without necessarily condemning the node.

Track four graph measures. Coverage is the number of bounded consumers with accepted receipts, not the number of integrations. Realised reuse is avoided or shortened work evidenced after delivery. Dependency exposure is the consequence and recoverability of each shared failure. Congestion is the delay imposed by the shared team's queue and change process. These measures reveal whether the graph compounds capability or accumulates coupling.

The OECD's review of AI adoption in firms discusses complementary assets and organisational conditions around adoption.10 A portfolio should translate that broad observation into local evidence rather than borrow an economy-wide average. Record staff routines, data preparation, process change and management work that each consumer still requires. If these complements must be rebuilt every time, the reusable node may be narrower than assumed. If the same trained routine and evidence contract operate across consumers, their avoided work belongs in the edge receipt.

Deliberate learning matters. Zollo and Winter distinguish experience accumulation, knowledge articulation and knowledge codification in capability development.7 An edge receipt makes learning explicit: teams record which cases transferred, which assumptions failed and which routine changed. March's exploration and exploitation framing cautions against overcommitting to existing knowledge at the expense of alternatives.8 The graph should therefore preserve room for independent experiments where adjacency is uncertain or diversity is valuable.

Retire an edge when the consumer no longer uses it, realised economics vanish, rights expire, the contract diverges or isolation becomes inadequate. Retire a node when too few consumers justify its fixed operation, unless it remains necessary for control or resilience. Strengthen a node when repeated edges demonstrate stable transfer and the capacity investment has an identifiable consumer base. Governance is the process of keeping these claims current, not keeping every shared service alive.

ISO/IEC 42001:2023 describes requirements for an AI management system and continual improvement.13 It does not certify that a particular edge creates value or that context transfers. The adjacency receipt complements management-system evidence by attaching reuse to a consumer outcome and incremental economics. Organisational governance can then see both the control system and the business relationship it is meant to enable.

08 · ConclusionGrow by validating edges, not announcing platforms

The adjacency graph changes the portfolio question. Leaders stop asking how many use cases a platform could support and ask which later decision population can reuse which proven prerequisite. They require a stable contract, transferable context and rights, portable tests, isolated failure, accountable ownership, an exit route and observed incremental economics. An arrow becomes a claim with consequences.

This discipline is neither centralisation nor fragmentation by default. It accepts shared identity, context, evaluation and operating routines when the receipts prove that they shorten or de-risk later delivery. It accepts independent systems when heterogeneity, conflicting rights, congestion or common-mode failure dominate. The correct topology is discovered relationship by relationship.

The executive artefact is a sparse, current graph. Nodes state what capability operates and for whom. Edges state what actually transfers and when the evidence expires. Circuit breakers show how failure is contained. Economics show whether reuse materialised. Missing or conflicting evidence is visible instead of hidden behind a roadmap percentage.

One AI capability opens the next only when the next consumer can cross an evidence-bearing edge. Until then, there is a promising node and a hypothesis. That distinction protects capital, keeps ownership legible and turns platform ambition into a sequence of falsifiable operating decisions.

SourcesReferences

1. Teece, D. J. (1986). Profiting from technological innovation: Implications for integration, collaboration, licensing and public policy. Research Policy, 15(6), 285–305.
2. Teece, D. J., Pisano, G. and Shuen, A. (1997). Dynamic Capabilities and Strategic Management. Strategic Management Journal, 18(7), 509–533.
3. Milgrom, P. and Roberts, J. (1990). The Economics of Modern Manufacturing: Technology, Strategy, and Organization. American Economic Review, 80(3), 511–528.
4. Bresnahan, T. F., Brynjolfsson, E. and Hitt, L. M. (2002). Information Technology, Workplace Organization, and the Demand for Skilled Labor. Quarterly Journal of Economics, 117(1), 339–376.
5. Brynjolfsson, E. and Hitt, L. M. (2000). Beyond Computation: Information Technology, Organizational Transformation and Business Performance. Journal of Economic Perspectives, 14(4), 23–48.
6. Cohen, W. M. and Levinthal, D. A. (1990). Absorptive Capacity: A New Perspective on Learning and Innovation. Administrative Science Quarterly, 35(1), 128–152.
7. Zollo, M. and Winter, S. G. (2002). Deliberate Learning and the Evolution of Dynamic Capabilities. Organization Science, 13(3), 339–351.
8. March, J. G. (1991). Exploration and Exploitation in Organizational Learning. Organization Science, 2(1), 71–87.
9. Pindyck, R. S. (1990). Irreversibility, Uncertainty, and Investment. NBER Working Paper 3307.
10. OECD (2025). The Adoption of Artificial Intelligence in Firms. OECD Artificial Intelligence Papers.
11. NIST (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0).
12. NIST (2023). AI RMF Playbook.
13. ISO (2023). ISO/IEC 42001:2023, Artificial intelligence management systems.
14. Sculley, D. et al. (2015). Hidden Technical Debt in Machine Learning Systems. NeurIPS.