Blog · ← All posts
Proving What Your Agent Emailed: Sent, Received, and Promised — the Mail Record That Ends Arguments
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 6 min read
The short answer: To prove what an agent emailed, record every send as a ledger action carrying the mail platform's message ID, the recipient in masked form, the subject, and a digest of the exact body sent — then reconcile the ledger against the mail platform's own sent log, which is the independent witness. With that pair, claims of the form 'your agent promised X' resolve mechanically: the message either exists with that digest (and its content is producible and authenticated against the digest) or it does not exist in a verified, fully reconciled record — in which case the promise was never sent by your agent.
Replay a recorded agent incident free → Step through a real recorded run in the replay view, or paste ledger records into the public verifier and get a VERIFIED/BROKEN verdict in seconds — no signup.
Email is the agent action customers screenshot. It arrives outside your systems, it reads like a commitment (because it is one), and it survives — in inboxes, in forwards, in dispute attachments — long after the conversation that produced it is forgotten. Agents that email are therefore agents whose words will be quoted back, sometimes accurately, sometimes creatively, always confidently.
The defensive design is not to email less; it is to make every send provable in both directions — what was sent, and what was not. This guide covers the recording discipline for agent email: the fields that matter, the digest trick that authenticates content without storing it in the ledger, and the reconciliation that makes 'never sent' a provable claim instead of a hope.
What is a content digest for email? a SHA-256 fingerprint of the exact message body as sent, stored in the send's ledger record alongside the message ID and a summary. The digest lets any later-produced copy of the message be authenticated — recompute the fingerprint and compare — without the ledger storing the full text: proof of content, without a second archive of every email the agent ever wrote.
The send record: five fields that decide disputes
Message ID first: the identifier your mail platform assigns at acceptance, recorded in the outcome. It is the join key to the witness — the platform's sent log — and the single field that converts an argument about your records into a lookup in a third party's. Recipient and subject in the summary, with personal addresses masked to the domain and a stable token where the ledger needs to group sends per recipient without holding addresses in clear. The body digest, computed over the exact bytes sent (after templating, after personalization — the final form, never the draft). The decision context: the reasoning that produced the email and the record it answers, so the send sits in its causal place — a reply to a specific customer message, a step in a specific workflow — rather than floating as an isolated artifact. And the outcome from the platform: accepted, bounced, or deferred, with the platform's timestamps.
Receives matter too: the inbound half of the story
Promises are disputed in context, and the context is usually the customer's preceding message. Record inbound mail that the agent acts on as cause events: the message ID, the sender token, the timestamp, and a digest of the content the agent consumed — particularly when the agent's subsequent send will be judged against what it was answering. The pair (inbound digest, outbound digest) reconstructs the exchange's evidentiary core without warehousing either text in the ledger: both parties hold the actual messages, and the digests arbitrate between conflicting copies, which is the only role the ledger needs to play.
This matters most in the pasted-content cases: when a customer pastes text purporting to be an earlier commitment, the absence of any matching inbound or outbound digest in a verified history is the measured, scoped way to say so.
The bounce and the retry: where email incidents breed
Email failure handling is a duplicate-action factory: timeouts on send (the classic lost acknowledgement — accepted by the platform, unconfirmed to the agent), retries that double-send, and fallback paths that quietly switch providers mid-incident, splitting the record across two platforms' logs. Record attempts as attempts, with the same discipline as money actions: the attempt, the ambiguity, the platform check by message ID or idempotency reference where the platform supports it, and the recovery decision — all in sequence. Reconcile per platform when fallbacks exist; coverage measured per witness is what keeps a provider switch from becoming a completeness hole with a marketing name.
Producing the answer: format for the dispute
When the claim arrives, the response assembles from records in minutes: the send in question (or the enumerated sends to that recipient in the window, when the claim is about a promise that does not appear), each with message ID, timestamp, subject, and digest; the mail platform's matching entries, confirming the sends in a system neither party controls; the chain verification for the period and the coverage figure for the mail witness, which together support the absence claim; and, where the body itself is contested, the produced message authenticated against its recorded digest. Keep the tone factual and the packet short — email disputes are won by the party whose evidence has the fewest moving parts, and a digest plus a message ID is as few as moving parts get.
Five actions you must take today
Email provability is a recording configuration, set once and paying out per dispute:
- Record the message ID on every send outcome. No exceptions for 'minor' mail — disputes choose their own exhibits, and the unrecorded send is disproportionately likely to be chosen.
- Digest the final body as sent. Computed after personalization, over the exact bytes — the digest of a template is a fingerprint of a message nobody received.
- Log inbound mail the agent acts on as cause events. Message ID, sender token, digest — the context that makes the outbound record interpretable and the exchange arbitrable.
- Treat send timeouts as lost acknowledgements. Lookup before retry, recovery decisions recorded; duplicate sends are a process defect, not an agent foible.
- Reconcile mail coverage per platform. Ledger sends against each platform's sent log, including fallback providers, with the figure filed monthly.
Worked example: the discount that was never emailed
A customer of fictional outdoor retailer Pinecrest Outfitters produced a forwarded email in which Pinecrest's agent appeared to promise 40% off a group order — a $3,300 commitment. Pinecrest's ledger for the agent showed 5,214 sends in the quarter, chain verified, and the mail platform reconciliation at 100%. Interrogation enumerated every send to the customer's domain: 9 messages, each with ID, subject, and digest. None matched the claimed subject; the claimed message ID's format matched a different provider entirely.
The decisive step was the digest test on the customer's own exhibit: Pinecrest invited them to compare — the body they held hashed to a value present nowhere in the ledger, while every genuine send to their domain authenticated against its record. Their IT team's follow-up found the message's true origin in a phishing simulation their own company had run, which had borrowed Pinecrest's branding and been forwarded internally as if real. The claim dissolved, the relationship survived on the strength of being shown proof rather than accused of forgery, and Pinecrest's postmortem added one line to the runbook: for email claims, lead with the digest — it lets the other side discover the truth without losing the argument first.
FAQ
Should the ledger store full email bodies instead of digests?
No — the digest plus the mail platform's own archive covers both needs better. The platform retains the actual message as a business system; the ledger's digest authenticates any copy of it. Storing bodies in the ledger duplicates sensitive content (email bodies are personal-data concentrate) into a second system with a longer retention, buying no additional provability. Digest for proof, platform for content, ledger for the sequence and the decision.
What if our agent emails through several tools and providers?
Then the recording wrapper belongs at the agent's send action, above the provider choice, so every route lands in one ledger with its platform noted — and reconciliation runs per platform, with each provider's sent log as its witness. Provider sprawl is the commonest completeness hole in email recording precisely because each tool 'has its own logs'; the ledger's job is to be the one history that spans them, and the coverage figure is how you know it succeeded.
Can we prove an agent received an email it denies acting on?
Inbound cause records plus the mail platform's delivery logs give you both halves: the platform shows delivery to the mailbox or webhook, and the agent's records show whether a consumption event followed. A delivered-but-never-consumed message is a routing or trigger defect, locatable in the sequence; a consumed message whose digest matches the customer's copy ties the agent's subsequent actions to that exact content. Both conclusions are record statements, not recollections.
Do digests hold up if the customer edits one word of the forwarded email?
Yes — that is the avalanche property doing evidentiary work. A single changed word produces a completely different digest, so the produced copy either authenticates exactly or it does not authenticate at all; there is no 'close enough' for a motivated editor to hide in. Present the test neutrally — it validates genuine copies as readily as it exposes edited ones, which is why offering it early builds rather than burns trust.
What about emails the agent drafts but a human sends?
Record the draft as a decision artifact and the human's send as the action, linked: the ledger then shows what the agent proposed, what (if anything) the human changed — the sent digest against the draft digest answers that mechanically — and under whose hand it left. Draft-and-send workflows are where accountability most often blurs, and the linked pair of records is the entire cure.
Start logging your agents free → The free tier gives you the full hash-chained ledger for 1 agent and 1,000 actions a month, with chain verification and replay included.