An AI Session Replay Is Evidence, Not an Audit Trail

A screen recording can show where an AI agent clicked. It cannot, by itself, prove that the click was authorized, that the correct business record changed, or that the result was accepted downstream.

Amazon Web Services’ new identity and session-replay capabilities for desktop agents are useful governance building blocks. They also make an important control distinction visible: replay is evidence, not the complete audit trail.

For operating businesses automating ERP, purchasing, inventory, customer service, claims, finance, or legacy desktop applications, the durable record must connect the person, agent, permission, approval, attempted action, actual system change, and recovery path.

What AWS announced on November 16

Documented: Amazon WorkSpaces for AI agents added OIDC identity-provider support, on-behalf-of sign-in, and session replay. AWS says on-behalf-of sign-in lets an agent work under the signed-in user’s directory identity, permissions, and Group Policies.

AWS says session replay combines screenshots with a plain-language transcript and delivers the resulting evidence to the customer’s Amazon S3 bucket. The product is designed for agents that operate ordinary desktop applications—including ERP-style workflows—through the same kind of graphical interface employees use.

Vendor claim versus documented capability: AWS describes the release as providing employee-like oversight and helping move agents into production. The release documents identity and replay features; it does not establish that every business action is authorized, correct, complete, or reversible.

The documentation adds important boundaries

AWS’s technical documentation says Replay Session is in public preview and is off by default. When enabled, it captures the whole screen immediately before each agent action and records metadata such as the tool call, arguments, response, and timestamp. Amazon Bedrock then produces a video, AI-generated transcript, and HTML report.

Also documented: Replay Session does not redact sensitive data; typed text can appear in metadata and transcripts; processing may occur in other AWS Regions; and delivered artifacts do not receive a signature or checksum. AWS recommends S3 versioning and Object Lock to protect them from modification or deletion.

Unresolved: the public materials do not publish transcript-accuracy measurements, evidence-completeness tests across application types, or proof that a visible click corresponds to the intended downstream transaction.

What replay can prove—and what it cannot

Replay can help establish:

  • What appeared on screen before an action
  • Which control the agent attempted to use
  • The sequence and timing of visible steps
  • Which tool call and arguments were recorded
  • Where an investigation should begin

Replay alone does not establish:

  • That the user authorized that exact action
  • That the agent selected the correct customer, order, item, vendor, or machine
  • That the destination system committed the change
  • That a retry did not create a duplicate
  • That the action complied with a business rule
  • That the evidence has not been altered after delivery

The business needs both AI governance and transaction-level evidence from its systems of record.

Keep three identities separate

On-behalf-of access is useful, but attribution still needs three distinct records:

  • Initiator: the employee, scheduled job, or event that requested the work.
  • Agent: the named agent, deployed version, instructions, tools, and accountable owner.
  • Execution identity: the user or service identity and effective permissions used in the destination application.

A person signing in does not mean that person approved every later action. Record the authorization decision separately, bind it to the exact proposed operation, and preserve the agent version that executed it. Apply least privilege through agent cost and access governance, not merely by inheriting a broad employee role.

Put deterministic controls before the click

Replay helps after an action. Deterministic validation prevents avoidable actions before they occur.

  • Confirm record identity using stable IDs, not names or screen position.
  • Validate required fields, status, quantity, price, account, destination, and authority limits.
  • Block duplicate submissions with idempotency keys or transaction checks.
  • Require human approval for money movement, unusual terms, production changes, customer commitments, deletion, or external communication.
  • Re-read the proposed change immediately before execution if the record could have changed.

The approval must cover the action that executes. If the agent changes the proposal after approval, the approval expires.

Bridge Road’s human approval controls provide a practical pattern.

Create a business transaction receipt

For every consequential action, connect the desktop session to a structured receipt outside the agent process:

  • Run, session, agent, tool-call, and transaction identifiers
  • Initiator, execution identity, effective permissions, and approval record
  • Target system, record ID, proposed before-and-after values, and rule results
  • Request time, response, retry history, and downstream confirmation
  • Replay artifact location and integrity information
  • Final business disposition: committed, rejected, reversed, duplicate, or unresolved

Vercel’s audit guidance makes the same general distinction: an attempted action and evidence of its effect are different records. A timeout can follow a successful write, so blindly retrying can cause a second change.

For commerce, tie the receipt to the order, payment, shipment, refund, or customer message. For manufacturing, tie it to the work order, part revision, inventory movement, or maintenance event.

Govern the recording itself

A replay can capture customer information, employee data, credentials displayed on screen, financial details, confidential prices, product designs, and typed business content. AWS documents that Replay Session performs whole-screen capture and no redaction.

  • Perform a data-flow and vendor review before enabling recording.
  • Exclude workflows that cannot be recorded lawfully or safely.
  • Mask sensitive fields in the application where possible.
  • Restrict replay access to named investigation and compliance roles.
  • Set retention by business, legal, privacy, and incident-response needs.
  • Verify whether cross-region processing matches data-residency obligations.
  • Log every replay view, export, deletion attempt, and policy change.

These decisions belong in the vendor and model risk review before production.

Protect evidence from the system it evaluates

An agent should not be able to edit or delete the record used to investigate it. Store the authoritative receipt and integrity metadata outside the agent’s writable environment.

  • Enable versioning and write-once retention where appropriate.
  • Hash exported artifacts and preserve the hash in an independent log.
  • Separate evidence-write permissions from evidence-read and delete permissions.
  • Alert on delivery failures, retention changes, missing sessions, and broken correlations.
  • Test retrieval with a known transaction rather than assuming collection is complete.

The current IETF agent-audit-trail document is an Internet-Draft, not an adopted standard, but its emphasis on records an agent cannot suppress or alter is directionally useful. Treat that as emerging guidance, not a compliance guarantee.

Design investigation and recovery before launch

Assign exception ownership for ambiguous outcomes. Operations owns the business queue; the application owner confirms the destination record; security preserves evidence; the agent owner evaluates model or tool behavior; and an accountable business leader decides whether to resume.

When a session is suspicious:

  1. Pause new work and preserve the session and transaction evidence.
  2. Check the destination system before retrying anything.
  3. Quarantine affected orders, records, files, or work instructions.
  4. Reverse or correct only through an approved workflow.
  5. Restore the manual process or last known-safe automation.
  6. Restart with reduced permissions and enhanced monitoring.

Connect this workflow to Bridge Road’s AI incident-response controls and the responsible operations owner.

A practical 30-day validation exercise

  1. Select one low-risk desktop workflow in ERP, CRM, inventory, or support.
  2. Map the initiator, agent, execution identity, effective permissions, approvals, and destination records.
  3. Add deterministic validation and an idempotent transaction receipt.
  4. Enable replay only after privacy, retention, access, and residency review.
  5. Run known success, rejection, timeout-after-success, duplicate, and partial-completion cases.
  6. Verify that an investigator can reconstruct the outcome without relying on the agent’s summary.
  7. Practice rollback and manual fallback before increasing autonomy.

Direct sources

Make every consequential action reconstructable

Bridge Road helps operating businesses connect agent identity, permissions, approvals, transaction evidence, monitoring, and recovery around real workflows.

Schedule a 30-minute friction review