Key Contract Issues for Agentic Managed Services

How Does the Evolving Delivery Model Affect Governance?

Governance is how the managed services model addresses complexity, uncertainty, and change. The contract establishes parameters for an effective working relationship to address situations that go beyond the contract terms. Those governance provisions include structures common in managed services agreements, including operational meetings, management committees, executive escalation, reporting, change control, and continuous-improvement processes.

Governance is central to agentic managed services because the provider is operating AI agents whose behavior and performance may evolve over time as models, data, prompts, workflows, and operating conditions change. That challenge is already evident in the market: in a recent ISG study, only 21% of enterprises said they have the internal skills to govern AI-enabled BPO services, while two-thirds expected their providers to lead AI adoption. The governance provisions are thus a value driver in agentic managed services deals.

For agentic managed services, the contract can set the guardrails within which the provider may deploy and operate agents. Those guardrails might include, for example, permitted use cases, agentic tools, models, systems and data uses; human-review requirements; prohibited actions; and applicable company AI governance and risk-management policies and directions. Outside those guardrails, changes would require company approval through a defined process. Recurring governance may include review of model or workflow changes, retraining or other material modifications, emerging drift or performance degradation, and changes in the level of human oversight required.

Operational governance provisions can also provide processes for exigent circumstances. For example, the contract can specify when the provider must suspend an agent, when the company may direct suspension of an agent or cut off its access, how services continue during the interruption, and how the parties decide whether to remediate, revert to a prior version, increase human review, or return temporarily to manual processing.

Observability is the practical foundation for that governance. The contract may require the provider to maintain and make available logs, decision traces, tool-use records, model and version identifiers, prompts or system instructions where appropriate, evaluation results, exception records, and human-intervention records. The required information should be tailored to the company’s oversight, audit, regulatory, and incident-response needs without turning ordinary governance into unrestricted access to the provider’s proprietary operating methods.

What Staffing Provisions Address Human Oversight Obligations?

Behind an agentic AI managed service is a human workforce. Even where AI agents perform much of the operational work, human reviewers, escalation personnel, subject-matter experts, prompt and workflow engineers, operational analysts, and incident responders remain responsible for monitoring, correcting, and supporting those AI agents. As AI agents perform a greater share of the work, fewer people may be needed overall, but those people may supervise a much larger volume of operational activity and their roles may become more important in controlling risk and achieving objectives.

Staffing provisions under the agentic managed services model should therefore be revisited. Depending on the service, the contract may need commitments including, for example, regarding level of human oversight, staffing levels or other capacity measures, coverage hours, escalation timing, qualifications and training (including for agentic technologies), language capabilities, geographic or citizenship limits, background checks, key personnel, and turnover and knowledge continuity. These provisions advance the company’s interest in the provider having sufficient qualified capacity to meet its monitoring, intervention, review, and recovery obligations.

The company may also want transparency into material changes in the division of work between agents and people, including the extent and location of human support, material subcontracting arrangements, and other operational dependencies. That may include a negotiated right to approve material changes to the provider’s staffing plan, although providers commonly resist disclosure of their proprietary operational methodologies. Where pricing, risk, compliance, or service quality are based on a level of human review, reductions in that review may require approval through an agreed governance or change process.

Conversely, the provider will reasonably want flexibility to improve its delivery model. A well-structured contract can preserve that flexibility by defining the outcomes, oversight requirements, and guardrails that matter to the company, while allowing the provider to change the particular mix of people and agents within those boundaries.

Who Determines the Agentic AI Technology Stack?

The provider’s delivery discretion may extend not only to the mix of people and agents, but also to the technology underlying the agents. In managed services, the provider typically supplies or selects some of the systems and tools used to deliver the services, while the company supplies other systems and tools, including its systems of record. With agentic managed services, those decisions are consequential because model choice, orchestration tools, monitoring platforms, system integrations, and agent configurations can materially affect service quality, security, data use, cost, and regulatory posture.

It may be useful to expressly distinguish between provider-selected and company-provided or company-mandated AI systems and platforms and, generally, provide that the party that selected or required the technology is responsible for the consequences of that choice. For example, if the provider selected the technology as part of its own managed-services solution, the provider would remain responsible for meeting its service commitments and for appropriately selecting, integrating, configuring, managing, and mitigating dependencies in that solution. Where the company mandated a particular model, platform, or other technology, the provider would be excused for its failures to the extent they are caused by that mandate, provided that the provider identified the concern before implementation and used reasonable efforts to work within or mitigate the constraint. Such an allocation of responsibility, however, would not by itself mitigate all risks arising from upstream third-party AI models, whether provider-selected or company-provided. As described in our prior Legal Update on agentic AI supply chain risk, the parties should first determine which of those risks can be reduced through architecture, access controls, monitoring, human review, and other solution engineering, then use the contract to assign responsibility for those measures and allocate the risk that remains.

The direct agentic AI technology stack is not the only question, because the provider’s solution may depend on the company’s environment. AI agents generally need integration points and API access, identities and entitlements in company systems, adequate data quality, documented and standardized processes, and retirement of the manual workarounds that a human workforce previously absorbed. If the company’s environment is not ready to support the provider’s solution, the work to make the company’s environment becomes a separate pre-deployment project and a contingency for deployment, with pricing and other adjustments if delayed. Identifying and expressly contracting for any such project reduces the risk of a later dispute over whether a performance shortfall was caused by the provider’s implementation or the company’s environment.

Contracts can smooth the path to modifying the agentic technology stack and the customer environment. For example, foundation models and AI platforms may change versions, prices, features, terms, and security characteristics during a multi-year outsourcing term. To avoid technical and operational problems, the contract can provide a process for monitoring and evaluating those changes, maintaining compatibility, proposing substitutions where appropriate, and implementing changes subject to applicable company approval and technical-change-control requirements. The contract should also distinguish ordinary-course technology lifecycle management, which is part of the provider’s responsibility for operating the services, from more fundamental changes that require company approval or may be separately chargeable projects.

Regardless of which party supplies the technology, data-use terms should apply across the operational chain, including prompts, outputs, logs, embeddings, telemetry, evaluation data, and other artifacts generated through service delivery. If company data may not be used to train or improve provider or third-party models, the agreement should say so expressly and require the provider to configure and procure the technologies used in its solution consistently with those restrictions.

How Will the Potential Liabilities and Other Risks of Agentic Managed Services Be Contractually Allocated?

Those distinctions in control over the delivery solution also inform the broader allocation of liabilities and other risks. A useful starting principle, familiar to outsourcing lawyers, is that the provider generally remains responsible for the resources that it uses to perform the services. Replacing a human processor with an AI agent would not, by itself, convert an outsourced service into a software license or SaaS arrangement or shift ordinary service-performance risk back to the company. In the managed services model, that principle generally extends to the provider’s selection, configuration, integration, deployment, and oversight of the resources used in its managed-services solution, subject to specifically negotiated dependencies and limitations.

More generally, responsibility follows control in managed services. The party with the right and practical ability to make a decision, manage a risk, or prevent a failure is responsible for the resulting consequences. The provider generally has the control, except for specified material company responsibilities and other agreed dependencies. Where a failure to satisfy an agreed dependency actually causes or contributes to a service failure, the contract provides the provider proportionate relief. That relief is often conditioned on timely notice, demonstrated causation, mitigation, and continued performance of unaffected obligations, and applies only to the extent and for the period the dependency actually affects performance.

Agentic AI can create a difficult blurring of control. Given the non-deterministic nature of AI, a company that decides to allow a managed services provider to use AI agents for a given function arguably accepts residual risks that cannot be mitigated through agreed human oversight services. In addition, failures in agentic managed services may have multiple contributing causes, some in each party’s control, making traditional single-party fault allocation difficult to apply. And where the company exercises control rights through governance, its participation may itself be seen as an assumption of responsibility, further complicating a control-based allocation. Principles of collaborative contracting therefore apply, sharing risk to incent joint risk mitigation. Mechanisms may include, for example, at-risk fees with earn-back for oversight failures, joint root-cause and remediation obligations, and gain sharing that rewards measured improvements in agent accuracy and intervention rates.

The parties should also revisit both indemnities and the treatment of first-party losses in light of the unique characteristics of agentic AI. The risks include third-party claims involving privacy or confidentiality breaches, infringement, regulatory violations, or other harms that should be addressed through the parties’ negotiated indemnities. Agent errors may also cause direct operational losses, such as erroneous payments, refunds, transactions, communications, or changes to systems of record, that do not necessarily involve a third-party claim. The allocation of high-value losses under the applicable liability caps, enhanced caps, exclusions, and other risk-allocation provisions will depend, among other things, on the service and the degree of provider control, but the contract should avoid gaps created solely because the harmful act was performed by an AI agent rather than provider personnel.

How Should Exit Provisions Address Transfer of an Operating Model Rather than a Software Tool?

Exit provisions are designed to facilitate the transfer of an operating function from a provider to another provider or back to the company. Under the agentic managed services model, the knowledge needed to run the function may now reside in agent configurations, prompts, orchestration logic, evaluation sets, monitoring rules, permissions, runbooks, and records of how humans intervene in edge cases.

The transition-assistance provisions should therefore identify the agent-specific artifacts and knowledge the provider must deliver or make available at exit. They should also address continued operation of provider’s AI agents during transition, export of relevant data and logs, cooperation in transferring third-party AI services or replacing provider-proprietary components, and sufficient human knowledge transfer to allow the successor provider and the company to understand both the automated and manual portions of the service.

Where the provider uses proprietary agents or tooling that cannot be transferred, the parties should address that constraint at signing. Possible approaches include replacement obligations, transition licenses, or sufficient advance notice and assistance to migrate the service to a different solution. Otherwise, the company may be locked into a provider solution.

Takeaways

Agentic managed services are a new deal structure within the managed services outsourcing model. The provider still assumes ongoing responsibility for a defined business or technology function; what changes is the mix of human and agentic resources used to perform it and the operational risks associated with those resources.

The principal contracting task is to adapt the managed services model to the agentic delivery model. Governance should control material changes to agent behavior and authority. Staffing terms should preserve sufficient qualified human oversight. Technology provisions should allocate responsibility for the AI stack and its evolution, address the company-side readiness the provider’s agents depend on, and recognize that upstream model risk is mitigated through engineering as well as contracting. Liability provisions should preserve provider accountability for its delivery model while recognizing company dependencies and reserved decisions. Exit provisions should make the automated operating model transferable.