The audit trail agency buyers ask for before trusting AI send
Agencies will not hand a client domain to an AI sender that cannot explain what it sent, why it sent it, and who approved it. Marketing pages that say "compliant" are not evidence. Buyers ask for an audit trail they can export per client workspace when a prospect complains, a security questionnaire lands, or SOC 2 reviewers want control proof.
GTM keeps a compliance gate in the database: no send without a verdict and an unsubscribe URL. Field provenance tracks where prospect data came from. Those pieces matter. The send audit trail is the chain that joins trigger, draft, approval, message ID, and delivery outcome inside one client boundary.
What "audit trail" means for AI outbound
MailerToGo's implementation guide for AI-sent email draws a clean line. Standard application logs prove a function ran. An email audit trail must prove which agent decided to send, which model version drafted, whether a human approved, which message ID the provider returned, and what delivery webhooks later confirmed.
Clarm and ARMO, cited in that guide, push the same idea: make logging platform-default, not opt-in. The workflow that skipped logging is always the one an auditor asks about. Append-only storage and tenant partitioning are not nice-to-haves. They are how you hand Client A's package without dumping Client B into the zip file.
GTM is US-resident today with SOC 2 Type II and ISO 27001 in progress, not certified. Buyers should treat "in progress" as a timeline conversation, not a shortcut past an exportable send trail.
The fields buyers should demand at send time
Ask vendors to show sample records, not slides. A practical minimum at send time:
- Client workspace ID (tenant)
- Agent or system identity distinct from the human seat
- Triggering input (scheduled step, manual send, reply follow-up)
- Model and version used for the draft, when AI wrote it
- Human approver and timestamp when supervised mode required one
- Compliance gate verdict and unsubscribe URL present
- Recipient reference (raw or hashed per policy)
- Provider message ID
- Send timestamp
- Sequence and step IDs
Then join webhook events on message ID: delivered, bounced, complained, and engagement events you actually store. Without the join key, you can prove intent to send and still fail to prove what the mail system confirmed.
StoryPros' consent-ledger guidance for AI BDR outreach adds the consent and suppression side. Opt-outs and "do not contact" states need timestamps and survival rules so a purge of prospect PII does not resurrect a suppressed address. GTM's product notes already treat suppression as something that survives purge. Your audit export should show that those writes happened.
What the trail is for (and what it is not)
Replay. Reconstruct one bad send for one recipient on one date.
Control evidence. Prove the compliance gate and human QA checkpoint were on during the window under review.
Per-client export. Security and legal teams want a package scoped to one workspace.
It is not a vanity open-rate dashboard. Observability tools sample and aggregate for ops health. Audit storage needs completeness and tamper resistance. Use both. Do not pretend Datadog answers a client DPA request.
Agency-specific failure modes
Cross-client bleed. An export that includes another client's message IDs fails the bid even if the product works.
Missing approval rows. If supervised mode was required and the log has no approver, auditors treat that as a control failure, not a "logging gap."
Model version omitted. Two versions of the same model can draft different claims on the same profile. Version is evidence.
Webhook never joined. Send succeeded in the app, bounce arrived at the provider, and nobody can prove the link.
Questions to put in the RFP
- Can we export an append-only send history per client workspace?
- Is agent identity separate from human seat identity?
- Do you store model version, compliance verdict, and unsubscribe presence on every send?
- How do delivery webhooks join to the agent-action record?
- Do suppression and opt-out events survive prospect purge?
- Who can access the trail, and is that access itself logged?
If the answers are vague, keep the AI in draft-only mode until they are not.
GTM is built so agencies can run find, qualify, write, send, reply, book with a gate that blocks incomplete sends and provenance on outbound fields. Ask for the trail during evaluation. Join the waitlist at https://trygtm.com/waitlist when you want that conversation with sales.
Sources
- https://resources.mailertogo.com/faq/build-audit-trail-emails-sent-by-ai-agents
- https://storypros.io/how-to-build-a-contact-level-consent-ledger-for-ai-bdr-outre
- /workspace/trygtm/knowledge/product.md