— AI & agents

User delegation or service identity: whose authority should an AI tool use?

Choose user delegation or workload identity for AI tools with an authority worksheet covering resource ownership, downstream access and offboarding.

4 min readUpdated 2026-11-05By AptixLabs studio
Requester: Human or schedule; Acting principal: User or workload; Application scope: Verified tenant access; Downstream resource: Intended credentialsAptixLabs · 2026-11-05

Original explanatory figure of a synthetic or proposed design; not a deployed product screenshot. View full-size diagram

Choose an AI tool's acting principal before choosing its credential mechanism. A delegated user operation should preserve that user's permitted authority; a scheduled workload may use a dedicated service identity with its own bounded permissions. Record the resource owner, execution principal and revocation path together. A valid cloud token or managed login does not answer which customer account a business action may affect.

Draw the authority chain in ordinary terms

For a synthetic report, name the requester, workspace owner, report worker and storage reader. Under user delegation, removal of the user's membership should affect later operations according to the application's documented revocation checks. Under a workload identity, the worker's permissions exist independently of a particular human session. The product must therefore record the authorizing workflow and permitted tenant scope rather than borrowing the most recent user's identity.

Google Cloud's MCP authentication guidance distinguishes authentication choices by runtime and principal. It recommends granting the agent identity only the required permissions. That cloud guidance applies to Google Cloud MCP resources; it does not automatically govern access to a separate SaaS application's records. Keep the cloud IAM boundary and the application's membership boundary visible in the diagram.

Google Cloud MCP authentication and principal choices

Compare two report contracts

In a delegated workflow, Mira asks for her workspace's permitted report now. The server verifies Mira, resolves membership and performs the read within that authority. In a scheduled workflow, a report worker runs nightly for explicitly enrolled workspaces. Its service identity can call the required dependencies, but application code must still select only enrolled accounts and apply delivery rules. Neither design should trust an arbitrary tenant ID supplied by the model.

Synthetic report authority comparison
DecisionDelegated userScheduled workload
Who acts?Verified memberDedicated service identity
Account scopeCurrent permitted membershipExplicit enrollment contract
Removal eventMember/grant lifecycleWorkflow and identity lifecycle
Audit fieldsRequester and tenantAuthorizer, worker and tenant

A useful audit record has both requested_by and executed_as where those identities differ. Include the tenant, operation identity and grant reference. Do not copy an incoming token into an unrelated downstream Authorization header. The receiving service must use a credential intended for that resource through an approved delegation or workload mechanism. Keeping token audiences distinct avoids turning the adapter into an implicit authority bridge.

Managed authorization still has an issuer and offboarding policy

Supabase's November 3, 2026 announcement describes Claude-issued enterprise-managed authorization for eligible Supabase organizations. It preserves a member-bound identity rather than turning every request into a universal service account. The published flow has specific organization, SSO and administrative prerequisites, and its offboarding differs from an identity-provider-issued flow. Review those conditions with the organization administrator before changing the trusted issuer.

Supabase: Claude-issued managed authorization, availability and offboarding

This provider example is a reminder to put issuer choice in the design worksheet. A login UI can disappear while the underlying member identity and removal procedure remain critical. Ask which systems must remove the person, when already issued credentials expire, and whether changing the issuer affects other tools. Do not generalize a managed provider feature into a capability of a custom MCP server.

Test the principal model you selected

Use two fictional workspaces and one removed member. For delegation, verify permitted access, crossed-workspace denial and the post-removal behavior promised by your token and membership checks. For a workload, verify the allowed schedule, tenant enrollment and destination restrictions. Also remove a human administrator and prove the scheduled task does not accidentally depend on that person's cached session.

No provider deployment or offboarding experiment is executed here. The table is a proposed authority worksheet. Review logs for identities and operation outcomes, while excluding credential values and report contents. If operators cannot explain who authorized a report and which principal executed it, adding another authentication library will not resolve the ambiguity. Bring that single journey to an integration review before expanding the tool inventory.

MCP server authentication: trace the identity all the way to the record

Evaluate the tool action, not just the answer.

Bring one tool workflow, supported hosts and a sanitized failure trace for a scoped AI integration review.

AptixLabs ai agent development

Should an AI tool use user delegation or a service identity?

Use the principal model that matches the task: member-bound authority for delegated actions, or an explicitly scoped workload identity for independent scheduled work.

Does managed authorization turn users into a service account?

Not necessarily. Inspect the provider’s member mapping, issuer and removal policy rather than inferring the principal from the absence of a consent screen.

Can a service identity ignore tenant checks?

No. Cloud permission to call a dependency does not grant access to every tenant’s application records.

Need help with ai agent & automation development?

Tell us about your product, existing stack and the problem you need to solve. We can discuss integration scope, implementation tradeoffs and a practical first milestone.

Start a conversation