Portfolio Concentration Governance requires organisations to detect and manage excessive dependence on any single model family, provider, task type, or governance assumption across their agent portfolio. When an organisation builds 30 agents on one model family from one provider, it has created a single point of failure that can materialise as a simultaneous outage, a pricing shock, a capability regression, or a regulatory prohibition affecting all 30 agents at once. This dimension requires continuous measurement of concentration across defined axes and enforcement of diversification thresholds or documented risk acceptance when thresholds are exceeded.
Scenario A — Provider Concentration Triggers Simultaneous Outage: A retail bank has deployed 24 AI agents across customer service, fraud detection, credit decisioning, and regulatory reporting. All 24 agents use the same foundation model from the same provider, accessed through the same API endpoint. The provider experiences a 14-hour outage due to a data centre incident. All 24 agents fail simultaneously. The bank's customer service channels go silent, fraud detection pauses (during which £2.1 million in fraudulent transactions pass through), credit decisions queue for 14 hours causing 890 mortgage application SLA breaches, and two regulatory reports miss their filing deadline.
What went wrong: The organisation never assessed concentration risk across its agent portfolio. Each agent was approved individually (per AG-249), but no portfolio-level view existed to identify that all agents depended on a single provider. The shared dependency was invisible at the individual use-case level. Consequence: £2.1 million fraud losses, regulatory investigation for missed filing deadlines, FCA supervisory letter citing inadequate operational resilience, £340,000 in customer compensation for mortgage SLA breaches.
Scenario B — Model Family Governance Assumption Invalidated: A healthcare organisation deploys 15 agents using the same model family for clinical decision support, patient triage, appointment scheduling, and resource allocation. The agents share a common governance assumption: that the model's outputs are calibrated — that is, when the model assigns 80% confidence to a recommendation, the recommendation is correct approximately 80% of the time. A model update changes the calibration characteristics. The organisation's governance controls, designed around the previous calibration profile, are no longer appropriate. All 15 agents are simultaneously affected, but the organisation's monitoring detects the calibration shift only in the clinical decision support agent (which has the most rigorous monitoring). The other 14 agents operate with degraded governance for 6 weeks before the issue is identified across the portfolio.
What went wrong: The shared governance assumption (model calibration) was never identified as a concentration risk. Each agent was monitored independently, and the cross-portfolio dependency on a single calibration assumption was not tracked. Consequence: 6 weeks of degraded governance across 14 agents, potential patient safety impact from miscalibrated triage recommendations, mandatory clinical audit of all agent-influenced decisions during the affected period.
Scenario C — Pricing Concentration Creates Budget Shock: A consulting firm builds 18 agents on a single provider's API, consuming approximately 4.2 billion tokens per month at a negotiated rate of $2.50 per million input tokens. The provider announces a pricing restructure: the negotiated rate will increase to $8.00 per million tokens in 90 days. The firm's monthly AI operating cost jumps from $10,500 to $33,600 — a 220% increase. Because all 18 agents depend on the same provider, the firm has no leverage to negotiate and no ready alternative. Migrating 18 agents to a different provider would take an estimated 6 months of re-engineering, testing, and revalidation.
What went wrong: The organisation concentrated its entire agent portfolio on a single provider without assessing the commercial risk of that concentration. No alternative provider capability was maintained, no portability requirements were included in agent design, and no concentration threshold triggered a diversification review. Consequence: £276,000 annualised cost increase, 6-month migration timeline if diversification is pursued, reduced margin on client engagements during the transition period.
Scope: This dimension applies to any organisation operating three or more AI agents in production. Concentration risk is inherently a portfolio-level concern — it becomes meaningful only when multiple agents exist. For organisations with fewer than three agents, concentration risk should be noted in the use-case approval (AG-249) but portfolio-level governance is not yet required. The scope covers concentration across five defined axes: model family, provider/vendor, task type, governance assumption, and infrastructure dependency. The scope extends to agents operated by third parties on the organisation's behalf if the organisation depends on their continued operation.
4.1. A conforming system MUST maintain a portfolio registry that records, for each deployed agent, the model family, provider, task type, governance assumptions, and infrastructure dependencies.
4.2. A conforming system MUST define concentration thresholds for each axis — the maximum percentage of the portfolio that may depend on a single model family, provider, task type, governance assumption, or infrastructure component.
4.3. A conforming system MUST measure portfolio concentration against defined thresholds at least quarterly.
4.4. A conforming system MUST require either remediation or documented risk acceptance by an appropriate authority when a concentration threshold is breached.
4.5. A conforming system MUST assess concentration impact as part of the use-case approval process (AG-249) — each new agent approval must evaluate whether the addition increases concentration beyond defined thresholds.
4.6. A conforming system SHOULD implement automated concentration monitoring that triggers alerts when thresholds are approached (e.g., at 80% of the threshold value).
4.7. A conforming system SHOULD maintain contingency plans for the failure or unavailability of any provider or model family on which more than 30% of the portfolio depends.
4.8. A conforming system SHOULD include portability requirements in agent design standards to reduce switching costs when diversification is needed.
4.9. A conforming system MAY implement concentration stress testing — modelling the impact of a provider outage, pricing change, or capability regression on the full portfolio.
Portfolio Concentration Governance applies the same principle that has governed financial portfolio management for decades — diversification reduces correlated risk — to the organisation's agent portfolio. The principle is well understood in financial services: a portfolio concentrated in a single asset class, geography, or counterparty is vulnerable to events that affect that single factor. The same principle applies to agent portfolios.
The concentration risk is often invisible because each agent is approved and deployed individually. AG-249 (Use-Case Approval Governance) ensures that each agent is individually approved, but individual approval does not address the correlation risk that emerges when many individually approved agents share a common dependency. This is the distinction between micro-prudential governance (each agent is well-governed) and macro-prudential governance (the portfolio as a whole is well-governed). AG-250 is the macro-prudential complement to the micro-prudential controls in AG-249.
The concentration axes are specifically chosen because each represents a failure mode that affects multiple agents simultaneously. Provider concentration means a single outage, pricing change, or terms-of-service modification affects all dependent agents. Model family concentration means a single capability regression, safety finding, or regulatory action affects all dependent agents. Task type concentration means the organisation has over-invested in automating one function while leaving others unaddressed — creating an unbalanced dependency on agent-mediated operations in that function. Governance assumption concentration — the most subtle axis — means that multiple agents share an assumption (e.g., "the model does not hallucinate proper nouns" or "the model's confidence scores are calibrated") that, if invalidated, undermines governance controls across the portfolio simultaneously.
This dimension connects directly to AG-093 (Supplier Concentration and Exit), which addresses vendor lock-in at the supply chain level. AG-250 extends this concern to the portfolio level — it is not just about whether you can exit a supplier, but about how much of your agent portfolio is affected if you must.
Portfolio concentration governance requires a portfolio-level view that does not naturally emerge from individual agent governance. The implementation must create and maintain this view.
Recommended patterns:
Anti-patterns to avoid:
Financial Services. Financial regulators have extensive experience with concentration risk governance. FCA requirements for operational resilience (PS 21/3) explicitly address concentration risk in technology outsourcing. The same principles apply to AI agent portfolios. Firms should align agent portfolio concentration thresholds with their existing technology concentration risk frameworks. The PRA's expectations on third-party risk management (SS2/21) apply when agent dependencies are on third-party providers.
Healthcare. Concentration in healthcare agent portfolios creates patient safety risk. If all clinical decision support agents use the same model family and that model exhibits a systematic bias (e.g., underdiagnosing a condition in a specific demographic), the bias propagates across all clinical functions simultaneously. Diversification of model families in clinical applications is a patient safety measure, not just a commercial risk measure.
Critical Infrastructure. Concentration thresholds should be more aggressive (lower) for agents in critical infrastructure. A 60% threshold appropriate for commercial agents may be unacceptable for agents controlling power grid operations or water treatment. Consider thresholds of 30-40% for critical infrastructure portfolios, with mandatory contingency plans for any dependency above 20%.
Basic Implementation — The organisation maintains a spreadsheet listing all deployed agents and their provider/model family. Concentration is reviewed manually once or twice per year. No formal thresholds are defined. Concentration is noted but not formally governed — there is no mandatory response when concentration is high. This level provides visibility but not governance.
Intermediate Implementation — A portfolio registry is maintained with structured data on all concentration axes. Formal thresholds are defined and approved by the governance body. Concentration is measured quarterly and at each new agent approval. Threshold breaches require documented risk acceptance or remediation. Contingency plans exist for high-concentration providers. Concentration metrics are reported to the governance body quarterly.
Advanced Implementation — All intermediate capabilities plus: automated concentration monitoring with real-time alerts at 80% of threshold values. Concentration stress testing is performed annually, modelling the impact of provider failure, pricing shock, and capability regression. Portability-by-design standards are mandatory for all new agents. The organisation maintains validated fallback capability for all providers serving more than 30% of the portfolio. Concentration metrics are weighted by criticality and blast radius, not just count. The portfolio registry integrates with AG-093 (Supplier Concentration and Exit) for end-to-end supply chain concentration visibility.
Required artefacts:
Retention requirements:
Access requirements:
Test 8.1: Portfolio Registry Completeness
Test 8.2: Concentration Threshold Enforcement at Approval
Test 8.3: Quarterly Measurement Execution
Test 8.4: Breach Response Verification
Test 8.5: Contingency Plan Validity
| Regulation | Provision | Relationship Type |
|---|---|---|
| EU AI Act | Article 9 (Risk Management System) | Supports compliance |
| FCA PS 21/3 | Operational Resilience — Important Business Services | Direct requirement |
| PRA SS2/21 | Third-Party Risk Management | Supports compliance |
| DORA | Article 28 (ICT Third-Party Risk) | Direct requirement |
| ISO 42001 | Clause 8.2 (AI Risk Assessment) | Supports compliance |
| NIST AI RMF | GOVERN 1.5, MANAGE 2.3 | Supports compliance |
| EBA Guidelines | Outsourcing Arrangements | Supports compliance |
The FCA's operational resilience framework requires firms to identify important business services, set impact tolerances, and ensure they can remain within those tolerances during disruption. When important business services depend on AI agents, concentration of those agents on a single provider creates a single point of failure that threatens the firm's ability to remain within impact tolerances. Portfolio concentration governance directly supports the operational resilience requirement by identifying and managing these single points of failure. The FCA has explicitly flagged third-party technology concentration as a key risk in its operational resilience supervisory approach.
Article 28 requires financial entities to manage risks arising from ICT third-party service providers, including concentration risk. DORA specifically addresses scenarios where multiple critical functions depend on a single provider. AG-250 implements this requirement for the AI agent portfolio specifically, ensuring that concentration in model providers is governed with the same rigour as concentration in other ICT service providers.
The PRA expects firms to assess concentration risk arising from dependence on individual third-party providers, including situations where multiple services depend on the same underlying provider (sub-outsourcing). Agent portfolio concentration governance extends this expectation to the AI-specific supply chain, including model providers, hosting platforms, and API services.
| Field | Value |
|---|---|
| Severity Rating | High |
| Blast Radius | Organisation-wide — concentration failures affect all agents sharing the concentrated dependency simultaneously |
Consequence chain: Unmanaged portfolio concentration creates correlated failure risk. When the concentrated dependency fails — through outage, pricing change, capability regression, regulatory action, or security breach — all dependent agents fail simultaneously. The blast radius scales with the degree of concentration: an organisation with 80% of agents on one provider faces 80% agent portfolio failure on that provider's worst day. The consequence is not hypothetical — major model providers have experienced multi-hour outages, surprise pricing changes, and capability regressions that affected all users simultaneously. The operational impact includes simultaneous loss of multiple business functions, inability to fall back to alternatives (because no alternatives were prepared), and extended recovery time (because diversification after a crisis takes months, not hours). The regulatory consequence is a failure of operational resilience — an inability to demonstrate that the organisation could maintain critical functions during a foreseeable disruption scenario.
Cross-references: AG-093 (Supplier Concentration and Exit) provides the supply-chain-level concentration and exit governance that AG-250 extends to the portfolio level. AG-249 (Use-Case Approval Governance) is the individual approval gate where portfolio concentration should be assessed. AG-045 (Economic Incentive Alignment Verification) addresses pricing risks that concentrate when the portfolio depends on a single commercial relationship. AG-251 (Strategic Fit and Substitution Governance) assesses whether agent alternatives exist, informing diversification options.