Freight broker and carrier communications data for AI agents

Why load-linked threads beat raw broker inboxes

A raw mailbox export teaches an agent how brokers write, but not what a message should cause in the TMS. Brokerage agents already triage inbound carrier email and phone inquiries, qualify carriers and capture quotes [1], and some run carrier conversations across many programmatic inboxes at high volume [2]. Training and evaluating that behavior needs each message tied to the state change it produced: a quote entered, a carrier assigned, an appointment moved, a check call logged.

The join key usually already exists. Brokerages keep a transaction record per load, and rate confirmations, bills of lading and PRO numbers carry the identifiers that let a supplier link email, portal and SMS threads to a load ID without guesswork. Ask how many messages could not be linked and what happened to them, because unlinked residue is where agents learn the wrong lessons.

What one record should contain

The useful record unit is one load with its full communication timeline, each message annotated with the load event it triggered. Ask suppliers to export threads from the TMS (McLeod, Aljex, Revenova, Turvo, Tai and similar) rather than from the mail server alone, because the TMS holds the status history and the bill of lading reference. Where shippers tender electronically, X12 EDI 204 (tender), 990 (response), 214 (status) and 210 (invoice) messages give machine-readable anchors that pair well with the human email around them; see our guide to carrier shipment status (EDI 214) data.

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "load_id": "L-000481",
  "mode": "FTL dry van",
  "lane": {"origin_3digit_zip": "606", "dest_3digit_zip": "752"},
  "pickup_window": "2025-03-04T08:00/12:00 local",
  "messages": [
    {"seq": 1, "channel": "email", "role": "shipper", "type": "quote_request",
     "linked_event": "quote_created", "ts_offset_min": 0},
    {"seq": 2, "channel": "email", "role": "broker", "type": "quote",
     "rate_band": "B4", "linked_event": "quote_sent", "ts_offset_min": 22},
    {"seq": 3, "channel": "edi_204", "role": "shipper", "type": "tender",
     "linked_event": "tender_received", "ts_offset_min": 310},
    {"seq": 4, "channel": "email", "role": "carrier", "type": "load_offer_reply",
     "linked_event": "carrier_assigned", "carrier_id": "CARRIER_7F2A"},
    {"seq": 5, "channel": "pdf", "role": "broker", "type": "rate_confirmation",
     "linked_event": "rate_con_signed"},
    {"seq": 6, "channel": "call_transcript", "role": "driver", "type": "check_call",
     "linked_event": "eta_updated", "location": "generalized_to_county"},
    {"seq": 7, "channel": "email", "role": "carrier", "type": "detention_request",
     "linked_event": "accessorial_submitted"}
  ],
  "outcomes": {"quote": "won", "tender": "accepted", "pickup_appt": "met",
               "delivery_appt": "missed_late_45min", "detention": "paid_partial",
               "fraud_flag": "none"}
}

Ask for rate confirmations as both the PDF and the extracted fields (linehaul, fuel, accessorials, payment terms, TONU and detention clauses). That pairing gives you document-extraction training data and a ground truth to check what an agent claims was agreed.

Outcome labels that make the data trainable

Outcome labels turn a correspondence archive into supervised and evaluation data. The minimum set for a quoting, booking and tracking agent:

Fraud cases such as double brokering and carrier identity theft are rare but disproportionately valuable as labeled negatives; ask whether the supplier flags them and how the flag was confirmed. Disputed loads also overlap with cargo claims, covered in our page on freight claims data for AI. Policy-aware labels matter as much as outcomes: research on task-oriented dialogue shows value in recording the actions agents take under business rules, not only the slots they fill [7].

Check-call audio and consent

Check-call recordings are licensable only if consent to record was captured in a way that covers every party. California, for example, requires the consent of all parties to record a confidential communication [3], and separately requires all-party consent for calls involving cellular or cordless phones [4], which describes most driver calls. Ask suppliers for the recording disclosure script, the dialer or telephony platform settings and the states where drivers were located, and verify the current list of all-party consent states with counsel before licensing audio.

Voice adds a second issue. Illinois BIPA treats voiceprints as biometric identifiers with retention and consent duties [5], so ask whether any speaker-identification features were derived and whether you may train speaker embeddings at all. If you only need text, license transcripts with diarization labels and leave audio out. For audio specifications such as sample rate, channel separation and transcript alignment, use our contact-center audio requirements template.

Privacy, carrier identity and rate sensitivity

Driver names, cell numbers, live GPS positions and home terminals must be removed or generalized before delivery, and the supplier should document how. If a supplier claims the data is deidentified under California law, the CCPA definition also expects reasonable technical measures, a public commitment not to reidentify, and contractual commitments from recipients not to reidentify [6], so expect that clause in your license. Employee-authored broker email raises workplace-communication questions covered in employee communications in training data.

Carrier MC and DOT numbers are public registration identifiers, yet many carriers are owner-operators, so a number can point to one person. Replace them with stable pseudonymous carrier IDs and keep only derived attributes such as fleet-size band or authority age. Rates inside threads are commercially sensitive to the broker and its shippers; expect banding, time lags or lane generalization to 3-digit ZIP, and check that bands remain fine enough to evaluate quoting.

Buyer request checklist

Use this when describing what you need; it maps to the questions a supplier and its counsel will ask.

Illustrative example: invented to show structure; it does not describe an available dataset.

• Scope: modes (FTL, LTL, reefer, flatbed, drayage), lane regions, date range, approximate load volume.
• Channels: email, TMS notes, portal messages, SMS, EDI 204/990/214/210, check-call transcripts or audio.
• Linkage: load ID join method, event types, percentage of messages linked, handling of multi-load threads.
• Labels: quote outcome, tender outcome, appointment adherence, accessorial disposition, fraud flags and their source.
• Privacy: driver and contact redaction method, location generalization, carrier pseudonymization, rate banding.
• Consent: recording disclosures, states covered, whether audio or transcripts only.
• Use: SFT, preference tuning, offline agent evaluation, held-out test splits, and whether outputs may ship in a commercial product.
• Eval hygiene: frozen test set never used for training, plus injected adversarial emails; see prompt injection evaluation sets.

If your agent also acts across the TMS, load board and accounting system, pair this with cross-system workflow records. For the broader landscape of industry-specific operational data and the full AI data buyer guide library, see the hubs.

How SourceX handles brokerage communications requests

SourceX sources operational datasets, including support and sales histories and documents, from US companies on request and manages the licensing process; nothing is held in stock and a request does not guarantee a match. You describe the data, such as load-linked broker threads with outcome labels, and SourceX looks for US businesses that hold it, with every release approved by the supplying company. Each dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery. Names, emails, phone numbers and account numbers are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. Related context sits on our pages for freight brokerages, logistics buyers, supply chain and logistics datasets, workplace email and chat, dispatch logs and email and calendar agents. You can describe your brokerage data need at any stage.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Requesting freight brokerage email data for AI agents

Describe the loads, channels and outcome labels your agent needs, and SourceX will assess data and licensing permissions with suppliers before anything is agreed. Pricing and allowed uses are set per deal in a license, and nothing is contracted until a supplier agrees. Start a freight brokerage data request.