Email parser with human review

Most business email is easy to read. A small share is not: a total written as "same as last time", a handwritten delivery note scanned at an angle, an order where two dates could both be the delivery date. A parser that treats every email the same either pushes uncertain values into your ERP or forces you to check everything by hand. Neither is acceptable for invoices, orders or anything that moves money or goods.

The review queue splits the stream. Emails that are clear go straight through by webhook or API. The ones that are not wait for a person. Because ParsedMail scores every field, as described on the AI extraction page, the split is based on the actual values, not on guesswork about which senders are difficult.

On Growth and Scale an email gets status needs_review when any of these is true:

• Any field has a confidence below the threshold of its inbox. The workspace default is 0.80, and each inbox can have its own value.
• A required field is missing. Missing required fields always score 0.
• The inbox classifies email across several schemas and no schema fits the email.

Pick the threshold per workflow. Supplier invoices that go to payment justify a strict threshold such as 0.90. Marketing leads that a salesperson will call anyway can run at 0.70. Emails in review are not delivered to your webhook until someone acts.

The review screen shows the original email, including attachments, next to the extracted fields. Fields under the threshold are highlighted with their confidence, and warnings, for example a date that could not be converted, appear on the field they belong to. The reviewer edits values in place, using the same types as the schema: a money field stays an amount plus currency, a table keeps its columns.

Then there are two actions:

• Approve marks the email as reviewed and delivers it to your webhook, with status delivered.
• Reject stops the email. It is not delivered, and stays in the log with its status for your records.

After approval the webhook payload and GET /emails/{id} return the corrected values, not the original extraction, and the email carries "reviewed": true. Your downstream code does not need a second path for reviewed mail. If you poll the email parsing API, filter by status to skip items still waiting in needs_review.

Workspaces have four roles: owner, admin, member and reviewer. The reviewer role is for people whose job is checking data, for example an accounts payable clerk or an order desk agent. Give them the reviewer role and they work the queue alongside the people who set up inboxes and webhooks. Growth includes 5 seats and Scale 20. Two-factor authentication is available for every user, and Scale adds SSO with SAML and an audit log.

If the same field keeps landing in review, the schema usually needs a clearer description. "Date" on an order can mean order date or delivery date; renaming it to "requested delivery date" and adding a short note often fixes it. Every schema edit increments the version, and you can replay emails from the log against the new version to check the effect. Replays of already counted emails are not counted again. The AI email parser page explains why clearer descriptions raise confidence.

The review queue and classification are part of Growth ($149 a month, or $74 a month billed yearly) and Scale ($499 a month, or $249 a month billed yearly). Starter gives you confidence scores on every field, so you can build your own checks in code. Compare plans on the pricing page.

The queue only contains mail that is worth a person's time. Duplicates (same Message-ID, or same sender, subject and body within 30 days in the same inbox) are marked duplicate and never reach review. Mail from senders outside the allowed list is logged as blocked_sender. Gmail and Outlook forwarding confirmation messages show in the log without being parsed. That keeps reviewers on real decisions instead of noise.