An ops team with engineers.

You are not buying extraction, you are buying the part that makes it safe to write into the TMS. The deliverable your team keeps is the eval set, because that is what tells you when it quietly stops working.

• The model reads each carrier email and returns a fixed structure your code can check, with anything it is unsure about marked as unsure instead of guessed.
• Writes into the TMS go through a narrow allowlist of fields. Every write is logged next to the email that caused it, so any record can be traced back.
• The first run is review-only. A person approves each write until the accuracy on your own mail is measured, then the safe field set goes automatic and the rest stays with a person.
• An eval set built from your real emails, including the ugly ones. Prompt and model changes get scored against it before they ship.
• A spend cap and an alert per carrier format, because a format change shows up as cost and as silence before it shows up as an error.

If your carriers send genuinely structured documents, check Rossum or your TMS vendor's own parsing before building anything. Parsing a stable format is a solved product.

Custom earns its place in the messy half: a dozen formats that change without notice, and rules that are specific to your lanes. That is where a bought parser stops and an agent with a measured error rate starts.

• Carrier formats change without telling you. The system has to notice a drop in confidence rather than keep writing confidently.
• Your engineers own this afterward. The repo, the prompts and the eval set live in your account from day one, and prompts get versioned like code.
• Anything that moves money or commits to a customer stays behind a person. A wrong status update is cheap to fix, and a wrong invoice is not.

Pick one carrier and one field set, measure it, and expand from the numbers. Free conversation first.