What foundation-plus-automation looks like in real operations

The challenge

Shipment data lived in the transportation management system, carrier invoices arrived by email and EDI in inconsistent formats, and accessorial charges were tracked in a shared spreadsheet maintained by one senior clerk. Every night, the settlement team manually matched shipments to invoices across three logins, and mismatches piled into a shared inbox nobody owned. By month-end, roughly a third of the team’s week had gone into reconciliation and chasing carriers for paperwork.

Our approach

We began with a two-week diagnostic: tracing a shipment from tender to settlement alongside the clerks who actually did the work. That surfaced the real problem — not missing data, but three systems that each held part of the truth and no agreed definition of a "matched" shipment. We built the foundation first: a governed shipment fact table ingesting TMS events, carrier invoices, and accessorials into one warehouse model, with quality checks that flagged missing documents at ingestion rather than at month-end. Only then did we add automation: document classification and extraction for inbound carrier paperwork, confidence-scored so that anything uncertain routed to a small human review queue instead of proceeding silently.

Implementation timeline: 13 weeks end to end: 2-week diagnostic, 3-week design, 8-week build delivered in two-week increments.

Outcomes

• 62%

  less manual reconciliation time

  measured against the diagnostic baseline over the first full quarter after launch

• 9 → 1

  exception queues, with a named owner

  previously nine shared inboxes and spreadsheets across terminals

• ~70%

  of carrier paperwork auto-classified above the confidence threshold

  the remainder routes to human review — by design, not by failure

The challenge

Referrals arrived by fax, portal upload, and email in dozens of payer formats. Intake staff keyed each one manually, frequently two or three times when the same referral arrived through multiple channels. Duplicate records created downstream scheduling conflicts, and when auditors asked how a referral had been processed, the answer lived in individual staff memory. Turnover in the intake team meant recurring retraining on undocumented process knowledge.

Our approach

Working closely with intake staff and the compliance lead, we documented the real intake process first — including the informal rules experienced clerks had never written down. We then built structured intake: inbound documents are classified by payer and referral type, key fields are extracted with confidence scores, and duplicate detection matches against existing records before anything is created. Anything below the confidence threshold — or any potential duplicate — routes to a review queue designed with the intake team, showing exactly what the system extracted and why it was unsure. Every action, extraction, and human correction is logged for audit reconstruction. We deliberately kept final referral acceptance with human staff: the automation prepares and de-duplicates; people decide.

Implementation timeline: 16 weeks: 3-week process documentation and design, 10-week build in two-week increments, 3-week parallel-run with the intake team before cutover.

Outcomes

• 41%

  fewer duplicate intake records

  measured over the first quarter after duplicate detection went live

• 100%

  of automated actions reconstructable from the audit log

  verified with the client compliance lead during a dry-run audit

• Confidence-based review

  low-certainty extractions never proceed without a human

  review queue volume stabilized at roughly one in five intakes — by design