Fraud, service, credit and compliance appear to diversify an AI portfolio. Then one foundation-model gateway fails, the common identity control blocks fallback and the same scarce reviewers are needed for recovery. Four different themes stop together.

A ranked list of stand-alone business cases is not a portfolio. Each case may clear its own value, risk and adoption tests while the funded set concentrates exposure to one provider, data estate, platform, legal interpretation, control, customer population, transformation team or macro assumption. The same set may also share genuine fixed cost, evidence and learning that no individual case can justify alone.

Portfolio economics must model downside dependence and productive interdependence together. Ignoring the first mistakes thematic variety for diversification. Ignoring the second can starve a foundation that enables several options, repeat control work and sequence change badly. Summing project values neither reveals common failure nor allocates shared benefit.

The article's original construct is a dependency-adjusted AI portfolio ledger with three connected records: a bet record, a dependency record and an interaction record. It reconciles base, common-shock and opportunity scenarios while retaining ranges, exposure bands and hard concentration limits.

01 · Three connected recordsDescribe the bets, dependencies and interactions before optimising the set

The bet record names the consequential decision, value pool, evidence range, capital at risk, reversibility, time horizon, option or commitment stage and accountable owner. Boundaries matter because sponsors can make projects appear independent by drawing them narrowly while leaving a common platform or outcome assumption outside every case.

The dependency record names shared model or vendor, cloud, data source, platform, control, legal basis, workflow, scarce role, customer population, outcome and macro assumption. For each factor it records exposed bets, shock scenario, substitutability, time to recover, concentration limit and owner. A dependency without an operational mechanism is not labelled correlation.

The interaction record classifies common-mode downside, resource collision, sequencing dependency, mutual exclusion, shared fixed cost, reusable evidence, learning spillover or value synergy. This grammar prevents every relationship being represented by one ambiguous score. Two bets may share a harmful provider exposure and a beneficial evaluation asset at the same time.

Records are linked through stable identifiers. A bet can depend on several factors, and one factor can expose several bets. An interaction names the particular pair or set it affects. This topology prevents the same dependency being copied into project prose with different names and allows a factor owner to see total exposure before approving another connection.

Capital at risk is not only approved spend. It includes non-redeployable integration, committed vendor terms, customer transition, control reliance and the portion of shared infrastructure that cannot survive a stop. Reversibility and expiry show how quickly the portfolio can change course. Two equal-value bets can deserve different positions because one preserves a near-term exit.

Value pools also need identity. If fraud and collections both claim the same avoided loss, or service and operations claim the same released time, overlap is removed before aggregation. The portfolio owner states which outcome owns the primary receipt and which bet contributes causally. Contribution is not permission to book the whole value twice.

Evidence ranges carry provenance and date. A field result, benchmark, management estimate and contractual commitment do not deserve equal confidence. When a dependency changes, only affected ranges reset. The ledger avoids discarding all evidence after a provider update while preventing stale project estimates from surviving a material architectural change.

Markowitz established the foundational insight that portfolio choice depends on relationships among components, not their isolated properties.1 AI initiatives are not traded assets: they lack liquid prices, stable return distributions and abundant histories. The analogy stops at dependence. The ledger does not infer market-style covariance from pilot scores.

Project-selection research explicitly considers interdependencies and incomplete information.234 It supports structured interaction reasoning, but the organisation still needs local causal evidence. The unit of analysis is a funded combination and sequence under stated constraints.

This is a reconciliation identity, not a valuation formula. Benefits that depend on the same outcome cannot be counted in full for every child. Shared fixed cost appears once with an allocation rule. A risk passed from a project to a platform remains in the portfolio unless the platform truly reduces or transfers it outside the boundary.

The rival: novel portfolios cannot support precise correlation estimates

The criticism is correct. Small pilots, shifting models and evolving workflows do not provide stable correlation estimates. A colourful matrix can create a more sophisticated form of false precision. Use named dependencies, ordinal exposure bands, stress scenarios, robust dominance and concentration limits where statistical estimation is not credible.

A relationship earns a place when the causal link can be described and a scenario can test it. Shared provider credentials create technical common mode. One specialist pool creates a resource collision. One regulatory interpretation creates policy dependence. A jointly reusable test corpus creates positive spillover. Mere similarity of project labels is insufficient.

Unknown dependence remains explicit. A sponsor may not know whether two application vendors use the same model service or region. That is missing evidence with a potential concentration consequence, not independence. Procurement, architecture and operational-resilience owners must resolve the chain before the board credits diversification.

Ordinal bands describe exposure using observable conditions: number and consequence of dependent bets, substitutability, recovery time, tested fallback and concentration against tolerance. The band does not pretend that “high” is a probability. It is a governance state that selects investigation, cap, exercise or rejection.

Scenario movement can be reported without a single expected portfolio value. Show the low and high range under the base case, the affected range under a named shock and the opportunity range after evidenced reuse. A hard-risk breach remains a constraint even if another bet's gain would offset it arithmetically.

02 · Interaction grammarSeparate dependence, collision, sequence, exclusion and synergy

Common-mode downside means one shock impairs several bets. Exposure depends on affected value, capital, reversibility, detection and recovery, not the number of project names. A shared service can be economical while its concentration remains above tolerance. The response may be a cap, fallback, substitution path or staged authority.

Resource collision occurs when individually feasible plans require the same scarce people, test environments, customer communication or review capacity in the same period. Adding business-case headcounts does not reveal the queue. The interaction record places demand on a common calendar and identifies which deployment moves or loses value.

Sequencing dependency means one result, control or asset changes the feasibility or value of another. A shared evidence spine may need funding before regulated uses can scale. Mutual exclusion means two approaches compete for the same workflow or architecture and should not both receive full rollout economics.

Positive interdependence includes shared fixed cost, reusable evidence, learning spillover and value synergy. These are distinct. Cost sharing reduces duplication. Reuse transfers an asset. Learning changes a later decision. Synergy creates an outcome unavailable to either bet alone. Each has an owner and receipt.

A shared platform can improve control consistency and lower operating cost while making outages or design errors common to the portfolio. The ledger reports both columns. Treating centralisation as purely efficient suppresses concentration; treating it as purely risky can force wasteful fragmentation.

Dependencies can be nested. An application relies on a model gateway, which relies on a cloud identity service and a common data pipeline. Substituting the application vendor may leave the lower layers unchanged. Recovery evidence traces the complete service path so a superficial alternative does not receive hedge credit.

Interactions also change over time. Early projects may collide for scarce evaluators, then create reusable assets that reduce later demand. A platform may initially diversify models while later standardisation increases concentration. The record includes the commitment stage and horizon, allowing the board to fund an option now without assuming the mature portfolio has the same topology.

Mutual exclusion deserves explicit closure. Competing architectures, policy interpretations or customer journeys can each look viable in isolation but create switching and operating conflict when both scale. Fund parallel learning only while the option value exceeds duplicated cost and before contradictory workflow commitments become difficult to reverse.

Synergy requires a joint counterfactual. The question is what outcome the combination creates beyond the sum of independently deliverable outcomes. Shared customer context might improve both service and fraud intervention, but the claim needs a permitted data path, action mechanism and receipt. Enthusiasm about a common platform is not evidence of synergy.

03 · Negative controlsCatch diversification theatre and duplicated economics

Thematic diversification illusion

Fraud, service, credit and compliance use different labels and value pools. All depend on one foundation-model gateway and identity control. A provider policy change or control failure creates a common stop. Diversification is assessed against failure factors, not business-function names.

Shared-enabler double count

Every child case claims the full benefit of one data-quality improvement, while none owns its cost or delivery risk. The portfolio ledger removes the repeated benefit, records the enabler once and allocates downstream incremental value only where a causal path exists. The enabler can then be funded on portfolio evidence.

Allocation is not value creation. Charging each bet a share may improve accountability but does not prove the platform produced all claimed outcomes. The receipt separates fixed cost avoided, asset reuse and incremental child value. If the enabler fails, the shock scenario removes dependent benefits together.

Change-capacity collision

Individually attractive deployments need the same operations specialists, reviewers and customer communications in one quarter. The calendar reveals infeasible concurrency. Sequencing can preserve total value, but delay may expire one opportunity. The board chooses the order using value decay, consequence and learning, not sponsor volume.

False independence through vendors

Two application vendors can depend on the same cloud region, model service or identity provider. Separate contracts do not establish substitution. The dependency record traces fourth-party services, credentials, data paths and recovery assumptions. A fallback counts only after its technical, contractual and operating route has been tested.

04 · Scenario movementStress ranges without fabricating an efficient frontier

The base scenario reconciles each bet under its current evidence and adoption-adjusted range, with shared cost and benefit removed from children. The common-shock scenario activates one or more dependency failures and applies recovery time, substitution and residual value. The opportunity scenario adds realised enabler reuse, learning and synergy only where a receipt exists.

The stand-alone sum is shown as a reference, not the portfolio answer. A set can appear valuable in that sum yet become dominated after one common factor and change collision are included. Another set with lower isolated value may preserve more under shock and create a reusable spine.

The common-shock scenario retains unaffected value. A provider failure may stop automated recommendations but leave a manual baseline, deterministic rules or already captured learning. Conversely, it may create recovery cost beyond project downside through customer communication and investigation. The scenario names both the surviving route and the additional burden.

Substitution reduces loss only after elapsed time and transition consequence are included. A technically compatible provider may require new assessment, data movement, contractual approval, prompt changes, staff instruction and outcome validation. The board records time to recover, not the existence of a second logo.

Opportunity scenarios are equally disciplined. Funding a shared evaluator can reduce later evidence cost, but only for compatible populations and controls. Learning from one bet can narrow another without guaranteeing success. Realised reuse is recorded when the later team consumes the asset and changes cost, time or decision quality.

Shock severity and concentration are tested separately. A moderate outage affecting every high-consequence bet may be less acceptable than a severe but isolated failure in one reversible pilot. The board uses affected capital, value, consequence and recovery, then applies the declared common-factor limit. It does not rank scenarios by revenue alone.

Concentration limits can be expressed as affected capital, number of consequential decisions, customer exposure, control dependence or maximum concurrent recovery demand. The chosen denominator matches the failure mechanism. A percentage of project count is inadequate when one bet holds most of the consequence.

Ranges widen when evidence quality is weak. Dependency confidence is recorded separately from effect severity. A high-severity provider link with uncertain likelihood can still breach a hard concentration limit. A low-quality synergy claim is excluded from committed value until its mechanism and owner are proven.

Robust dominance asks whether one combination performs no worse across credible scenarios and better in at least one important respect. When no set dominates, the board uses exposure limits, reversibility and staged learning. It does not resolve uncertainty by choosing the midpoint that favours a sponsor.

Scenario design names the causal event, affected services, response, recovery clock and evidence. “Market downturn” is too broad unless the common value or demand assumption is explicit. Stress remains actionable because an owner can change concentration, substitute a dependency, reschedule demand or stop a bet.

05 · Counter-caseFund a shared spine while preserving substitution and staged change

A regulated-finance portfolio can fund one evidence and control spine before scaling fraud, credit and service bets. The spine provides evaluation datasets, outcome receipts, access controls, incident evidence and model-change tests. Child cases use it without claiming its full value separately.

The portfolio preserves a provider substitution path with compatible interfaces, validated fallback behaviour, data portability, contractual exit and named recovery owners. It caps how many high-consequence actions may depend on the primary provider until the fallback has passed a live exercise.

Operating change is sequenced. The first bounded deployment tests the spine and produces learning for later bets. Specialist and review demand stay within calendar capacity. A negative result can narrow later exposure before capital and workflow commitments harden.

Positive interdependence becomes realised value only when later work reuses the asset or changes a decision. Building a platform in anticipation is cost and option capacity, not captured synergy. The interaction receipt names the child, reused component, avoided work or improved outcome and owner.

This set can outperform both fragmented projects and one fully centralised launch. It centralises reusable control evidence, decentralises decision boundaries and retains substitution. The portfolio is designed for common-shock survival and learning, not maximum simultaneous activity.

The spine itself has kill criteria, versioning and service boundaries. If later bets cannot reuse it, if control evidence becomes stale or if its delivery risk exceeds the duplicated work it replaces, the portfolio narrows or stops the enabler. Foundational status does not grant permanent funding.

Provider substitution is exercised under realistic load and consequence. Teams verify data rights, identity, logging, latency, output behaviour, monitoring and rollback. The exercise produces elapsed time and residual exposure, not a binary pass. High-consequence bets remain capped until the tested route meets their boundary.

Sequencing uses learning only when the first stage can answer a question relevant to the next. Deploying fraud first merely because it is ready does not justify delaying credit. The record states which dependency, control or adoption uncertainty will be resolved and how the result changes the later commitment.

A negative first-stage result can be productive portfolio evidence. It may stop several dependent bets, justify provider substitution or show that the shared spine is mis-specified. The receipt records avoided exposure and released capacity without relabelling technical failure as success. Decision change is the economic contribution.

06 · Interactive instrumentDependency-Adjusted Portfolio Board

The board accepts up to six bets and diagnoses their combination. Monetary fields use one consistent portfolio unit. Dependency and interaction ratings are operating hypotheses. The result shows base, shock and opportunity ranges, the dominant common factor and missing evidence without inventing an efficient frontier.

07 · Portfolio governanceAssign concentration, interaction and rebalancing authority

Every common factor has an owner who can cap new exposure, require substitution, schedule a recovery exercise or stop a dependent bet. Project sponsors cannot waive a portfolio limit by approving their cases separately. Exceptions record the approving authority, duration and compensating action.

Dependency disclosure is part of funding approval. Architecture, procurement, data, legal, operations and finance review the factors within their evidence. Sponsors attest that shared costs and benefits have not been copied into another case. Unknown fourth parties and unowned enabling work remain visible conditions.

One authority owns the funded set. It can move capital and time across projects, not merely recommend coordination. Without that right, every sponsor optimises locally and common exposure grows through individually rational approvals. Portfolio governance requires power to sequence, cap and withdraw, not a consolidated dashboard.

Receipts distinguish transfer from reduction. Moving model validation to a central team may lower duplication but create a queue. Moving provider risk into a platform may improve contractual leverage but not reduce service concentration. The portfolio records the new owner, capacity and failure mode before claiming a reduction.

Review cadence follows dependency speed. Provider status and specialist capacity may need frequent monitoring; legal interpretation and customer adoption move differently. The board updates the factors that changed and replays relevant scenarios. It does not force every bet through a ceremonial quarterly rescore.

When evidence improves, the board can relax a cap, advance a sequence or recognise reuse. When recovery fails, it can shrink authority even if stand-alone value rises. Rebalancing treats concentration and positive interaction as live portfolio properties rather than facts frozen at annual planning.

Portfolio receipts reconcile realised bet outcomes, shock exposure, shared cost, asset reuse and learning. A kill decision from Paper 020 can remove a bet, but rebalancing also asks what dependency capacity it releases and whether another bet loses an enabler or learning source. Removal can improve or weaken the set.

Rebalancing triggers include provider change, concentration-limit breach, recovery-test failure, evidence decay, resource collision, duplicated benefit, enabler delivery, synergy receipt and option expiry. The board re-runs scenarios rather than preserving the original rank order.

Revised US model-risk guidance emphasises aggregate model dependence and risk-based governance,5 while Basel and FSI sources address third-party, fourth-party, concentration and critical cloud risk.678 DORA establishes operational-resilience obligations in its scope.12

NIST connects lifecycle and supply-chain risk to governance and monitoring,91011 OECD evidence highlights complementary adoption conditions,13 and ISO 31000 supplies a general risk-management frame.14 The three-record ledger and five-route board are this article's synthesis.

08 · Decision principleDiversify failure factors and cultivate productive interdependence

A collection becomes a portfolio only when the organisation can explain how its components fail together, compete for capacity, depend on sequence and create reusable value. Thematic variety and separate vendors do not answer those questions.

Record the bets, dependencies and interactions. Remove duplicated economics. Stress causal common shocks. Test substitution and recovery. Sequence scarce change. Fund shared enablers where the portfolio receipt supports them, and cap exposure where centralisation makes failure common.

A credible portfolio can be smaller than the approved pipeline. Deferring one attractive bet may preserve specialist capacity, keep a concentration limit intact or allow an earlier test to change later architecture. Unfunded does not always mean rejected; the ledger distinguishes waiting for sequence evidence from losing the business case.

The converse also holds. A modest enabler may deserve funding before a higher-value application because it reduces several future uncertainties, establishes a recoverable control path or removes repeated fixed work. Its case is made at portfolio level, with downstream options and failure exposure visible, not by inflating its stand-alone revenue.

Do not ask which six projects have the highest isolated totals. Ask which combination preserves acceptable value across credible shocks, fits operating capacity and turns shared evidence, controls and learning into realised outcomes without counting them twice. The answer names the concentration cap, substitution exercise, sequence, enabler owner and next rebalancing trigger so that the portfolio remains an operating decision rather than a presentation. It also makes the next decision date and the evidence required for it explicit.

Evidence spineReferences

1. Markowitz, H. (1952), ‘Portfolio Selection’, The Journal of Finance, 7(1), 77–91. doi:10.1111/j.1540-6261.1952.tb01525.x.
2. Santhanam, R. and Kyparisis, G. J. (1996), ‘A Decision Model for Interdependent Information System Project Selection’, European Journal of Operational Research, 89(2), 380–399. doi:10.1016/0377-2217(94)00257-6.
3. Liesiö, J., Mild, P. and Salo, A. (2008), ‘Robust Portfolio Modeling with Incomplete Cost Information and Project Interdependencies’, European Journal of Operational Research, 190(3), 679–695. doi:10.1016/j.ejor.2007.06.049.
4. Rodrigues-da-Silva, L. H. et al. (2024), ‘Project Portfolio Selection Considering Interdependencies: A Review of Terminology and Approaches’, Project Leadership and Society, 5, 100115. doi:10.1016/j.plas.2023.100115.
5. Board of Governors of the Federal Reserve System, Office of the Comptroller of the Currency and Federal Deposit Insurance Corporation (2026), ‘SR 26-2: Revised Guidance on Model Risk Management’. Federal Reserve letter.
6. Basel Committee on Banking Supervision (2025), Principles for the Sound Management of Third-Party Risk. BCBS publication.
7. Basel Committee on Banking Supervision (2022), ‘Newsletter on Third- and Fourth-Party Risk Management and Concentration Risk’. BIS newsletter.
8. Financial Stability Institute (2023), Managing Cloud Risk: Some Considerations for the Oversight of Critical Cloud Service Providers in the Financial Sector, FSI Insights 53. FSI publication.
9. Tabassi, E. (2023), Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. doi:10.6028/NIST.AI.100-1.
10. National Institute of Standards and Technology (2023), ‘AI Risk Management Framework: Core’. NIST AI RMF Core.
11. Autio, C. et al. (2024), Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1. doi:10.6028/NIST.AI.600-1.
12. European Union (2022), Regulation (EU) 2022/2554 on Digital Operational Resilience for the Financial Sector. Official Journal text.
13. OECD, Boston Consulting Group and INSEAD (2025), The Adoption of Artificial Intelligence in Firms: New Evidence for Policymaking. doi:10.1787/f9ef33c3-en.
14. International Organization for Standardization (2018), ‘ISO 31000:2018, Risk management, Guidelines’. ISO 31000 standard record.