Healthcare AI comparison

    Copilot Studio vs Bedrock Agents vs Google ADK: Where Can a Health System Enforce PHI Policy on Tool Calls?

    A health system can enforce PHI policy on a tool call only where a control sits in the call path and can see the inputs and outputs. Access controls such as connector availability, identity roles and network boundaries limit what an agent can reach but do not decide on an individual payload. Code-level hooks can act on each call, but the health system builds and maintains them separately on each platform. Across a mixed estate, native controls rarely add up to one consistent policy and one body of audit evidence.

    Criteria for deciding whether native controls are enough

    • PHI rules must apply to the content of individual tool call inputs and outputs, not only to which tools are reachable.
    • Agents run on more than one platform, so one rule would otherwise be implemented three different ways.
    • Redaction or blocking depends on custom code that each agent team must write, test and keep current.
    • Human approval must be triggered by policy and recorded with the content reviewed.
    • Auditors need one evidence format for allowed, redacted, blocked and escalated decisions across platforms.
    • Retention and data residency requirements differ from what each platform's native records provide.

    Access controls and inline controls are different things

    When an agent calls a tool, connector, action group, API or MCP server, PHI can leave the health system in the request or enter the agent's context in the response. Controls that address this fall into two groups, and the distinction matters more than any feature list.

    Access-level controls decide whether an agent can reach a tool at all. Connector and data policies, identity and role permissions, and network boundaries belong here. They are valuable and necessary, but they act on the tool, not on the content of a given call. An agent that is permitted to reach a scheduling API can still send a full patient record to it.

    Inline controls act on the individual call. They read the input or the output at runtime and can allow it, redact part of it, block it, or hold it for a person to approve. Only controls in this group can enforce a rule such as "no patient identifiers in free-text parameters to this tool."

    Copilot Studio, Bedrock Agents and Google ADK each present control points from both groups. The architect's task is to establish, for each PHI rule, which group the available control belongs to.

    Enforcement points across the tool call lifecycle

    Organizing the comparison by lifecycle stage, rather than by product feature, keeps the question on where policy can actually act. For each platform, apply the same five tests.

    • Pre-call inputs: Can a control read the parameters the agent is about to send and allow, redact or block them? Connector availability or a network boundary does not qualify. A hook or filter that sees the payload does.
    • Authorization: Which identity does the call run as, and which tools can it reach? Data policies, roles and network boundaries sit here. They decide reach, not payload content.
    • Post-call outputs: Can a control inspect returned data before it re-enters the model context or reaches the user? Returned PHI is often the larger exposure, because the agent may receive more than it needed.
    • Human approval: Can a call be paused for a person to approve, and does the approval record who decided and what content they saw?
    • Evidence: Does every decision (allowed, redacted, blocked, escalated) produce a record that names the call, the policy applied and the outcome, with defined retention and export?

    What each platform asks of the architect

    The three platforms present different kinds of control points. Connector and data policies, policy configuration on agent actions, and identity and network boundaries are set by administrators. Controls of this kind typically govern which tools an agent can reach and under what identity. Where a given configuration also inspects call content, it should be mapped to the pre-call and post-call stages explicitly, not assumed.

    Callback and plugin hooks in agent code are a different model. A hook can sit around a tool call and apply any rule the team can express, including redaction and blocking. The tradeoff is ownership. The health system writes the detection logic, tests it against real PHI patterns, versions it, deploys it with each agent, and keeps it current as tools change. A hook that is missing from one agent is a gap for that agent only, which makes coverage hard to demonstrate.

    Neither model is better in the abstract. Administrator-configured controls are easier to operate but may stop at access. Code-level hooks are more expressive but shift the maintenance burden to the customer. Most health systems will use both, and the PHI policy should be written once, then mapped to whichever control point can enforce each clause on each platform.

    Audit evidence and HIPAA 164.312(b)

    The HIPAA Security Rule technical safeguards at 45 CFR 164.312 apply to covered entities and business associates and operate in accordance with 164.306. Paragraph 164.312(b) is the audit controls standard. For agents, the practical question is what record exists of what an agent was allowed to do with ePHI and what it did.

    Access-level controls usually show that a tool was reachable. They do not, on their own, show that a specific call carrying PHI was inspected and what happened to it. Evidence of tool call decisions needs a record for each outcome: allowed, redacted, blocked or escalated. For each platform, establish where that record is stored, who can read it, how long it is retained and how it is exported to the systems the security team already uses.

    Ownership also needs to be clear. Where a business associate operates part of the agent stack, the contract should say who produces and retains this evidence. Mapping platform evidence to 164.312(b) should follow the organization's own security risk analysis and policies.

    HHS has also published a proposed amendment to the Security Rule intended to strengthen the cybersecurity of ePHI. As of November 2026 it should be treated as a proposal. A periodic review of the control mapping keeps it aligned if the requirements change.

    Where runtime policy enforcement in the execution path fits

    A policy layer in the execution path addresses the gaps above by sitting between the agent and the tool, independent of the platform that built the agent. Trussed is one example. It enforces policy in the execution path across models, agents, tools and MCP servers, at the scope of individual interactions and tool calls. Decisions are enforced at runtime, not reviewed after the fact: allow, redact, block, rate limit, or escalate for human approval.

    Each enforcement action produces audit-ready evidence that is mapped to HIPAA, including 164.312(b) audit controls, among other regimes. Trussed deploys self-managed in the customer's cloud, with configurable retention and data residency. It runs in production with healthcare and pharma organizations and carries SOC 2 Type II, HIPAA, GDPR, FERPA, ISO 27001 and NIST AI RMF credentials. In production it has run 100+ models under one policy plane.

    The scope is specific. Trussed enforces policy on interactions and tool calls. It does not provide end-to-end tracing of agentic workflows or transaction-level governance, and it does not replace platform access controls such as identity and network boundaries, which remain part of the design.

    PHI policy points in the tool call lifecycle

    Pre-call inputs

    Inspect parameters before they reach the tool.

    Authorization

    Decide which tools and data an agent can reach.

    Post-call outputs

    Inspect returned data before it re-enters the agent.

    Human approval

    Pause a call for a named person to decide.

    Evidence

    Record each decision for audit review.

    Map your PHI policy to each enforcement point

    Walk through your Copilot Studio, Bedrock Agents and ADK estate with an expert and see how policy enforcement and audit-ready evidence work on individual tool calls.

    Talk to an Expert