An AI Agent Shared Responsibility Model for Deployers

Table of Contents

Executive Summary

For organisations deploying AI agents on third-party platforms, our AI agent shared responsibility model shows who runs each control, who owns it and what evidence supports that allocation.

We build on existing shared responsibility models with four allocations – platform-inherited, builder-owned, shared and deployer-owned – and six rules for applying them to each workflow.

Before you rely on a provider’s control, you need evidence and a contractual basis; for shared controls, you need a documented split, while a named person in your organisation remains accountable.

Your choices when building or changing an agent can also make you a provider under the EU AI Act, so we show how to test for role migration and record each allocation in your Statement of Applicability.

Our Agentic AI Readiness Assessment, Agentic AI Risk Control Framework and hands-on support can help you assess readiness and allocate controls; contact us to explore your options.

Purpose

In the first paper in this series, Agentic AI Assurance at the Point of Deployment (August 2026), we proposed twelve objective criteria for assuring the organisations that deploy agentic AI.

Criterion 5, shared-responsibility allocation, asks you to allocate each control explicitly across the AI agent supply chain, and to recognise that building agents on a third-party platform can change your regulatory role.

That matters because one organisation may now build an agent on a platform run by a second, using a model trained by a third, for a workflow in which a fourth holds professional or regulatory duties. In UK financial services, third-party systems already make up 53% of high-materiality AI systems (Bank of England and FCA, October 2026). When something goes wrong, you will need to show whose control failed and who owned it.

Three bodies have already published shared-responsibility models for AI: the Cloud Security Alliance, in its AI Controls Matrix (July 2025) and Agent Shared Security and Safety Responsibility Model (June 2026); the Coalition for Secure AI, in its AI Shared Responsibility Framework (May 2026); and Singapore’s Infocomm Media Development Authority (IMDA), whose agentic AI governance framework distinguishes platform providers from system providers (May 2026). These models describe who runs which security controls.

In this paper, we build on them from your side as the deployer. We set out the evidence you will need before you should rely on a provider’s control, a test for when building on a platform changes your role, and a way to record each allocation in your Statement of Applicability – the register of which controls apply to each workflow. We start with the precedent these models share: cloud computing.

The precedent: shared responsibility in cloud computing

Cloud computing faced the same question two decades ago, and its answer gives us a starting point.

Cloud providers such as Amazon Web Services and Microsoft publish shared responsibility models. Each states which controls the provider runs (physical security, the virtualisation layer, service availability) and which the customer always keeps (data classification, identity, configuration). The split varies by service: a customer renting raw infrastructure carries more controls than a customer using finished software. The Cloud Security Alliance’s STAR programme lets providers publish evidence that they run their share.

The precedent teaches three lessons:

  • Allocate responsibility control by control.
  • Let the split follow how much of the technology the provider runs.
  • The customer always keeps the controls that depend on its own context: its data, its users and its purpose.

The AI agent supply chain

To apply these lessons to agents, you first need to know who is in the supply chain. Your agents depend on external providers that sit outside your direct control, and your own organisation can play up to two roles inside it. One organisation can hold several roles, each with its own duties.

External providers fall into the six groups we set out in our blog, Managing AI Agent Vendors and External Dependencies: 

  • Model providers. Train and serve the foundation models your agents use, including open-weight models and versions modified from them. They own model integrity, update control and baseline safety behaviour.
  • Agent platform providers. Run the platforms you build agents on, including orchestration, tool integration, memory, sandboxing, logging and guardrails. Each platform provider is itself a customer of a model provider and relies on that provider’s controls as you rely on the platform’s.
  • Tool and connector suppliers. Provide the APIs, plugins, connectors (such as MCP servers) and other agents your agents call. They own the integrity and security of each tool and give notice when its behaviour changes.
  • Data and knowledge sources. Supply the external data feeds, retrieval stores and vector databases your agents read. They own the quality, lineage and security of what they supply.
  • Cloud and infrastructure providers. Host your models, agents and data. They own physical security, availability and the infrastructure controls in their own shared responsibility models.
  • Existing suppliers adding agents. Add agents to software you already use. They own the agent features they add, which you assess as you would any new provider.

Your map should also include every subcontractor (sub-processor) beneath these providers. Your own organisation then plays one or both of the remaining roles, and serves a third party:

  • Builder. Composes agents on a platform: defines goals and boundaries, writes prompts, selects tools, connects data sources and grants permissions. These composition choices give an agent its purpose and much of its risk.
  • Deployer. Puts the agent to work and owns run-time governance, human oversight, monitoring, incident response and organisational readiness, including the duties the EU AI Act places on deployers of high-risk AI systems (Article 26). Your firm may be both builder and deployer; each role keeps its own duties.
  • End user. Uses or is affected by the agent. End users carry no allocation in this model; every other party owes its duties to them.

Your role depends on what you build

These roles can change. When you build on a platform, your choices can turn you from a deployer into a provider – a shift we call ‘role migration’.

Under Article 25 of the EU AI Act, you take on a provider’s obligations if you put your name or trademark on an existing high-risk AI system, substantially modify one that remains high-risk, or change the intended purpose of an existing AI system so that it becomes high-risk.

Our Agentic AI Risk Control Framework (‘the Framework’) applies this test through its substantial-modification assessment (control 5.07). The assessment checks drift, composition changes and purpose changes against the threshold, and records whether you have taken on a provider’s duties.

So, if you build on a platform, your role depends on what you build. Assess whether your composition choices meet one of Article 25’s three conditions for taking on a provider’s duties. At the platform layer, you remain the platform’s customer.

Both descriptions will be true at once, at different layers, and an assurance approach that asks for a single answer to ‘provider or deployer?’ will miss one of them. Your role will also shape your liability: the EU’s new Product Liability Directive extends product liability to software, including AI systems, placed on the market or put into service after 8 December 2026.

Our control allocation model

Because your role differs by layer, our model allocates responsibility one control at a time, as the cloud precedent suggests. For each workflow, it gives every applicable control one of four allocations:

  1. Platform-inherited (P). A model, platform, tool, data or cloud provider runs the control. Your duty as deployer is to verify it: you hold named provider evidence (a certification, audit report, trustmark or contractual attestation) and a contractual basis for each inherited control.
  2. Builder-owned (B). You discharge the control through the choices you make when you build the agent: its goals, permissions, tools and data connections.
  3. Shared (S). You and a provider each run part of the control, and you document where one part hands over to the other. For example, the platform generates telemetry, and you retain it, monitor it and act on it.
  4. Deployer-owned (D). You discharge the control through how you run the agent: registers, oversight, monitoring, drills, incident response, training and board reporting.

Six rules you’ll need

On their own, these four labels would leave room for gaps and disputes, so we use six rules to avoid this. The first three ensure each allocation is complete and provable, the fourth keeps it current as your role changes, and the last two keep it on record and keep a person accountable.

  • Rule 1 – Complete allocation. Give every control that applies to a workflow exactly one primary allocation, so that every control has an owner.
  • Rule 2 – Inheritance requires evidence. Mark a control platform-inherited only when you hold named provider evidence and a contractual basis for it. If either is missing, the control reverts to you. For EU financial firms, DORA (Articles 28–30) already sets out the contract terms they must agree with ICT providers.
  • Rule 3 – Shared means specified. Record a shared allocation only with a documented split that states which party performs which part, and how they hand over.
  • Rule 4 – Test for role migration. Assess your composition choices against the substantial-modification threshold (control 5.07). If you cross it, record the provider duties you have taken on and reset the allocation.
  • Rule 5 – Allocation lives in the Statement of Applicability. Record each allocation beside the applicability decision in the workflow’s Statement of Applicability. ISO/IEC 42001 clause 6.1.3(f) requires a Statement of Applicability identifying necessary controls and justifying their inclusion or exclusion. Reset it whenever your platform, tool catalogue or control framework changes.
  • Rule 6 – Accountability stays with you. A named person in your organisation stays accountable for every control, whatever its allocation. A contract binds only the parties who sign it (IMDA legal working group, May 2026), so accountability stays with you even when a provider does the work. This follows the Framework’s principle that only people are held morally responsible for actions taken by processes that include AI, and it matches the Cloud Security Alliance’s Agent Owner.

Illustrative allocation

To show the model at work, the table below allocates a sample of the Framework’s controls. These allocations are typical for a firm building on a third-party platform. Yours will depend on your platform’s capabilities and your contracts, which is why Rule 5 asks you to record them for each workflow.

Typical allocation Control (Framework ref.) Your remaining duty Typical evidence (examples)
1 – Platform-inheritedSecure communication protocols (8.02)Confirm the mechanisms cover your channels and configurationProvider certification or audit report; log of authenticated inter-agent exchanges
1 – Platform-inheritedModel integrity and update control (15.09)Track update notices; re-test workflows after material model updatesProvider attestations and release notes; regression test results after each model update
1 – Platform-inheritedIsolate and sandbox inputs (13.03)Verify sandboxing covers every input channel the workflow usesIsolation configuration; prompt-injection test results for each input channel
2 – Builder-ownedDefined goals and boundaries (2.01)Version-control goals, metrics, and permitted actionsVersioned list of permitted and prohibited actions
2 – Builder-ownedMinimal access rights (11.01)Grant least privilege with justification and automatic expiryApproved permission specification; access-control tests
2 – Builder-ownedHarness and tool catalogue change governance (1.11)Review every harness or tool-catalogue change as a deployment eventVersioned change log; change approval records
2 – Builder-ownedSubstantial modification assessment (5.07)Assess composition and purpose changes against the threshold; record any provider duties taken onDocumented determination for each material change
3 – SharedAutomatic kill switch and shutdown (2.08)Platform supplies the mechanism; you define stop criteria and drill itStop authority assignment; kill-switch test and drill records
3 – SharedRuntime execution and forensic logs (21.02)Platform generates telemetry; you retain it, monitor it and reconstruct eventsTimestamped action logs; forensic reconstruction test; retention policy
3 – SharedPost-deployment drift detection (5.05)Monitor prompts and agent behaviour against a baseline; act when thresholds are crossedDrift alert history; regression and drift test results
4 – Deployer-ownedCentral AI asset catalogue (1.05)Register every agent, including those built on platformsAgent register covering live and retired agents
4 – Deployer-ownedDependency mapping and fallbacks (18.01)Map every external dependency for each agent componentDependency map, updated at each material change; supplier risk assessments
4 – Deployer-ownedAssigned ownership and legal and ethical accountability (22.01)Name an accountable owner for each workflow (Rule 6)Agent register or responsibility matrix naming one owner for each agent
4 – Deployer-ownedProportionate oversight and oversight quality audit (29.01, 29.09)Design proportionate oversight and audit it for rubber-stampingHuman oversight matrix; oversight quality audit results

Worked example: a law firm building agents on a third-party AI platform

Consider a law firm that builds matter-management agents on a third-party AI platform. The firm plays a different role at each of three layers, and the model allocates controls at each.

At the platform layer, the firm is the platform’s customer. Model integrity, sandboxing and protocol security are platform-inherited, and under Rule 2 the firm holds the platform’s certifications, audit reports and contractual commitments as evidence. It relies on the platform’s security engineering and checks the evidence for it.

At the composition layer, the firm is a builder. It decides which matters the agents touch, which document stores they read, which actions they may take and which prompts direct them. It discharges every builder-owned control itself, including goals and boundaries, least-privilege access to client files and tool-catalogue governance; these sit outside any platform certificate.

At the operating layer, the firm is a deployer, and its profession adds duties that stay with it. Confidentiality, legal professional privilege and supervision obligations attach to the firm and its qualified lawyers personally, as the Solicitors Regulation Authority confirmed in its August 2026 warning notice. These sit within deployer-owned controls such as oversight design, accountability and incident response, and show Rule 6 at work: a professional duty stays with the firm whatever the allocation.

Finally, if the firm offers its agents to clients as a service, or repurposes them beyond anything the platform assessed, Rule 4 prompts an Article 25 assessment. The firm becomes a provider under that Article if it puts its name on an existing high-risk AI system, substantially modifies one, or changes an AI system’s intended purpose so that it becomes high-risk. At the same time, it remains the platform’s customer. The model records both roles because it allocates by layer.

What this means for you

Agents depend on chains of providers, and when something goes wrong you will need to show whose control failed and who owned it. This paper gives you four steps:

  1. Map the chain. List the six groups of external provider your agents depend on, and the builder and deployer roles your firm plays.
  2. Allocate every control. Give each one an allocation – platform-inherited, builder-owned, deployer-owned or shared – under the six rules.
  3. Hold the evidence. Keep provider evidence for every control you inherit and a documented split for every control you share.
  4. Watch for role migration. Test composition and purpose changes, and record any provider duties you take on.

In practice, you add four fields to each control in your Statement of Applicability: the allocation (1–4), the provider evidence reference where inherited, the split reference where shared, and a migration flag where your choices approach the substantial-modification threshold.

The Cloud Security Alliance, the Coalition for Secure AI and IMDA have given the industry a shared picture of who runs which controls. We believe their models are the right foundation, and that the next step is to make each allocation provable from the deployer’s side, so that your board, your regulator and your clients can see the evidence behind it.

The four allocations and six rules work on any platform and in any sector, and they are small enough to adopt whole. To put them to work, take four further steps:

  1. Start with one material workflow. Allocate its controls, close the gaps you find, then extend the model to your other workflows.
  2. Use your next contract renewals. Ask each provider for the evidence and contract terms Rule 2 requires before you rely on its controls.
  3. Add a role-migration check to change approval. Make the substantial-modification assessment (control 5.07) a condition of every material change to an agent.
  4. Report allocations to your board. Show who owns each control, what evidence backs it and where gaps remain.

We can help with each of these. Our Agentic AI Readiness Assessment shows where you stand, and the Agentic AI Risk Control Framework gives you the controls and evidence to start from. Or if you would like to move faster, we could take your team through allocating the controls for one of your key workflows, reviewing your provider evidence and contracts against the six rules. Contact us to explore your options.

Frequently Asked Questions

Your agents depend on providers you do not control, and in UK financial services third-party systems already make up 53% of high-materiality AI systems. When something goes wrong, your regulator and your clients will ask whose control failed. This model lets you answer with evidence.

Those models describe who runs which controls. Ours adds what you, as deployer, need to prove it: evidence before you rely on a provider (Rule 2), a test for role migration (Rule 4), a record in your Statement of Applicability (Rule 5) and a named owner for every control (Rule 6).

Under our AI agent shared responsibility model, both have responsibilities for different controls. Providers assure the controls they run; you assure the rest and verify the evidence for what you inherit. A named person in your organisation stays accountable for every control (Rule 6).

You are a deployer, and the platform’s customer. Under Article 25 of the EU AI Act, you also become a provider if you put your name or trademark on an existing high-risk AI system, substantially modify one that remains high-risk, or change the intended purpose of an existing AI system so that it becomes high-risk. Rule 4 tests for this.

A record of which party runs each control: a provider, your build choices, your operations, or a documented split. Our model uses four allocations – platform-inherited, builder-owned, shared and deployer-owned – and six rules, recorded for each workflow in your Statement of Applicability.

Picture of Adam Grainger

Adam Grainger

Agentic AI Risk Management

Template Agentic Risk Appetite and Adoption Strategy download

Fill in this form and get access to our
Template Agentic Risk Appetite and Adoption Strategy for free

Agentic AI Risk Appetite Statement and Adoption Strategy

Enterprise-Wide Agentic AI Risk Control Framework

Fill in this form and get access to the
Enterprise-Wide Agentic AI Risk Control Framework.

Subscribe to our monthly newsletter

Fill in this form and stay up to date

Get in touch