Brokers and forwarders still place a large share of road freight by phone and email: a shipper asks for a price, a dispatcher scans a carrier list, negotiates, books and later chases the driver for status. Digital freight matching combines a rate-prediction model, a matching/optimisation layer and increasingly LLM and voice agents to quote in seconds, book carriers automatically and cut empty kilometres. The human focus shifts to exceptions, relationships and pricing policy.
Problem & value
- Today: fragmented market of many small carriers, load boards, tender portals and email. Quoting relies on dispatcher experience; carriers are chosen by habit; check calls and status chasing consume much of the day. Empty running and waiting time are structural cost drivers.
- AI value: consistent and fast spot quoting, better carrier-to-load fit (lane history, equipment, reliability), automated booking and tracking follow-up, margin protection through rate-risk awareness, fewer empty legs through back-haul matching.
- Evidence: Eurostat road freight statistics report a considerable share of empty running in EU road transport [1]; vendors such as Uber Freight and Transporeon publish product claims on automation and matching [2][3], but figures are vendor-reported and not independently audited — validate with a pilot on your own lanes.
Variants
- Digital broker / marketplace: two-sided platform (shipper app, carrier app, instant price).
- Traditional forwarder or broker: AI copilot for the dispatch desk (quote inbox, carrier ranking).
- Shipper-side procurement: spot-tender automation from a TMS with carrier waterfall (also AGENT-024).
- Small carrier / owner-operator: load search, back-haul finder and rate benchmarking.
- Contract vs. spot hybrid: routing guide for contracted carriers, spot fallback when rejected.
- Private use: not applicable (B2B).
How it works
- load tender (TMS/EDI/email/portal)
- document & message parsing (LLM/OCR: lane, dates, equipment, weight, special terms)
- rate prediction (gradient boosting / time-series on lane, season, fuel, index, lead time)
- carrier matching (score by price, reliability, equipment, position, preference; solver for multi-load / back-haul combinations)
- quote to shipper (auto within guardrails, dispatcher approval above limits)
- booking: tender waterfall or bidding to carriers (API / email / voice agent)
- tracking & check calls (telematics / visibility APIs, LLM or voice agent fallback)
- POD capture, invoicing, feedback into rate and carrier-score models
- Rate model: tabular ML (gradient boosting) predicts price and acceptance probability per lane and lead time; time-series features capture seasonality and tight-capacity periods. Output is a range, not a point, so margin rules can be applied.
- Matching: scoring plus assignment/set-partitioning solver (OR-Tools, HiGHS, Gurobi); for multi-stop combination see OPT-001 and OPT-012.
- Agents: LLM reads quote-request emails and writes replies; voice agents (ASR + LLM + TTS) place routine check calls; agents call tools (TMS API, rate engine) rather than inventing prices.
- Human-in-the-loop: price floors/ceilings, new carriers, high-value or hazardous loads and any negotiation outside the guardrail go to a dispatcher; policy owner reviews model drift weekly.
Data preparation
Quality of historic lane and quote data decides whether rate prediction beats a dispatcher's gut feeling. The data typically lives in the TMS or ERP (orders, invoices, freight cost), in shared mailboxes (quote requests and carrier offers), in Excel rate sheets, in carrier portals and as PDF or paper CMRs and proof-of-delivery scans.
Typical gaps: lane definitions that differ between systems (postcode versus city names), quotes without a recorded outcome (won or lost), duplicate carrier master records, expired insurance certificates, scanned documents, and rates held only in individual dispatchers' mailboxes.
Preparation steps: 1. Build a data inventory and gap analysis (PREP-001, PREP-002) for TMS, mail and Excel sources. 2. Export orders, costs and quotes via API or database extract (PREP-005) and harmonise lanes to postcode areas. 3. Clean carrier master data, deduplicate and add performance flags; tidy rate sheets (PREP-011). 4. Harvest win/loss labels from booked versus declined quotes (PREP-020). 5. Digitise CMR and POD scans with OCR (PREP-006, DOC-014).
Tools: the TMS exports, a small ETL job or n8n workflow, and an OCR component. Expect roughly half of the total project effort to go into data work, mostly lane harmonisation and label recovery.
Implementation paths
Which path for whom
- smb (A): mature freight-matching and TMS products with EU hosting exist at SMB prices, and spot brokerage is a fairly standard process.
- enterprise (B): rate models and carrier scoring built on your own lane history and TMS are a competitive asset, and volume makes per-seat or per-load fees high.
- developer (B): cloud APIs for LLMs and OCR plus your own rate model let you integrate with the customer's TMS faster than training or hosting everything.
Path A — Buy (recommended for most SMBs)
- Digital brokers / marketplaces: Uber Freight (US, global), sennder (EU, Germany), Timocom and Trans.eu (EU freight exchanges, with Timocom pricing/market-insight tools), Transporeon (EU, now part of Trimble) with spot-tender and carrier-network modules, Teleroute/Alpega (EU). Check current product lines before purchase.
- Brokerage automation: TMS vendors with quote and tender modules, e.g. Alpega TMS, Transporeon, Trimble, Descartes, project44 / FourKites / Sixfold (EU) for visibility and automated check calls.
- Choose A when: you want quick start and can accept network lock-in and platform fees.
Path B — Build on cloud APIs
- Quote/mail agent: frontier LLM with structured outputs (Claude, Gemini, GPT, Mistral — EU region or EU provider) extracting lane/date/equipment into JSON and drafting replies.
- Rate and match services: managed ML (Vertex AI, Azure ML, SageMaker) on your quote history; or a library model served as an internal API.
- Integration: TMS REST/EDI (EDIFACT IFTMIN), visibility APIs, freight-exchange APIs where licensed, orchestration with n8n or a workflow engine.
- Voice: managed speech APIs plus an LLM for scripted check calls; always disclose that it is an AI caller.
Path C — Self-host (open source)
- Models: LightGBM/XGBoost/CatBoost for rates; Chronos or similar time-series foundation models for index forecasting; OR-Tools or HiGHS for matching; open-weight LLMs (Mistral, Qwen, Llama-family) via vLLM for email parsing; Whisper-class ASR for call transcripts.
- Hardware: HW1 for the ML/solver part, HW2 for a small LLM and speech stack.
- Effort: needs data engineering and ML staff; pays off for brokers with high volume.
Path D — Custom training
Core of competitive advantage: train rate and acceptance models on proprietary quote/booking history (typically tens of thousands of historic loads across lanes), including lost quotes (censored data). Carrier-reliability and ETA models need telematics history. Use champion/challenger testing before automating.
State of the art (2026)
- Agentic brokerage: large players and start-ups advertise AI agents that quote, negotiate within limits, book and perform check calls. The market is early; sustained straight-through rates differ strongly by lane complexity.
- Pricing: gradient boosting still dominates tabular rate prediction; foundation models help mainly on unstructured inputs (emails, PDFs), not on the price itself.
- Visibility: real-time transportation visibility via telematics/ELD-style APIs and the EU eFTI Regulation (electronic freight transport information) digitising transport documents [4].
- Limits: thin data on rare lanes, sudden market shocks (strikes, border closures), adverse selection (carriers accept only underpriced loads), freight fraud and identity theft on carrier onboarding, LLM mistakes on terms such as Incoterms or ADR. Never let an agent confirm prices outside guardrails.
Shared resources & combinations
- Same stack: tabular/time-series forecasting with PRED-001 and OPT-004 (pricing); solver components with OPT-001, OPT-012 and OPT-008.
- Documents and operations: DOC-006 (CMR, delivery notes, POD) and AGENT-024 (freight operations agent) are natural companions; voice/ASR stack shared with contact-centre use cases.
- Data assets: lane rate history, carrier scores and visibility data are reusable across all.
Effort, cost & KPIs
- Path A: weeks to a few months incl. integration; transaction or subscription fees.
- Path B: 3–6 months for quote copilot plus rate service; cost drivers are data cleaning, TMS integration, LLM tokens (small versus labour).
- Path C/D: 6–12 months with ML team; ongoing monitoring.
- KPIs: quote-to-book conversion, quote response time, gross margin per load and margin variance, rate-prediction error (MAPE) per lane, share of loads booked touchlessly, carrier acceptance rate, empty-kilometre share, on-time pick-up/delivery, check calls per load, fraud incidents.
People & change
Start with management and the head of dispatch or operations, who own the process and the margin targets. Involve IT and the data owners of TMS and ERP early, because integration and data access gate everything. Carrier relationships are personal, so bring a few experienced dispatchers in as pilot group and let them veto bad rate or carrier suggestions; booking stays a human click at first.
Works council: depends. If your company has one, co-determination under section 87(1) no. 6 BetrVG is likely when the system logs dispatcher activity, response times or call content (check calls via voice agents), and section 90(1) no. 3 requires timely information about AI-supported work processes. Agree that rate and carrier scores are not used for individual performance rating.
Data protection: carrier drivers and contact persons are personal data. Involve the DPO, sign an AVV with each processor (Art. 28 GDPR) and run a DPIA if call recordings or driver tracking are involved. Inform carriers when voice agents call them (Art. 50 AI Act).
AI literacy: support staff with short hands-on training on reading confidence ranges and spotting wrong quotes (see GUIDE-006). Typical concern: fear of replacement; address it by redeploying time to negotiation and key accounts (GUIDE-008). Plan the pilot with GUIDE-009 and the works council steps with GUIDE-004.
Risks, regulation & compliance
- EU AI Act: typically minimal risk — pricing and matching of goods transport is not an Annex III use. Exception to watch: if the system allocates work to or monitors self-employed drivers (algorithmic management of workers), assess Annex III(4) employment criteria. Voice agents must disclose AI interaction (Art. 50). No Art. 5 practice involved; do not infer driver emotions.
- GDPR: driver names, phone numbers, location and call recordings are personal data; apply purpose limitation for tracking (no tracking outside the transport), retention limits, DPA with vendors, consent/notice for call recording, works-council involvement where employees are affected.
- Platform and competition law: EU Platform Work Directive 2024/2831 may apply if drivers are platform workers; avoid price coordination through shared rate models between competitors (Art. 101 TFEU).
- Sector rules: CMR liability, German Freight Forwarder liability (HGB), Mobility Package and cabotage rules, EU Data Act for access to vehicle data, eFTI.
- Mitigations: price guardrails, carrier vetting and fraud checks (insurance, licence, bank details), audit log of every automated quote, kill switch, backtesting on historic loads.
Getting started
- Export 12–24 months of quotes, wins/losses, bookings and carrier performance from the TMS.
- Measure baseline: quote time, margin, conversion, empty-leg share per lane.
- Pilot a quote-inbox copilot (Path A/B) with human approval; keep the dispatcher in control.
- Add a rate model on top lanes; run shadow mode and compare against real outcomes.
- Auto-book only within narrow guardrails (known carriers, standard equipment), then widen.
Sources
- Eurostat — Road freight transport statistics (empty running) — https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Road_freight_transport_statistics
- Uber Freight — Insights AI and product pages — https://www.uberfreight.com
- Transporeon (Trimble) — Freight marketplace and spot tendering — https://www.transporeon.com
- Regulation (EU) 2020/1056 on electronic freight transport information (eFTI) — https://eur-lex.europa.eu/eli/reg/2020/1056/oj
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Directive (EU) 2024/2831 on improving working conditions in platform work — https://eur-lex.europa.eu/eli/dir/2024/2831/oj
- Google OR-Tools — https://developers.google.com/optimization
- Chronos time-series models — https://github.com/amazon-science/chronos-forecasting