AI-Powered Carrier Selection in Digital Freight Platforms
AI-powered carrier selection means running a load through a scoring engine that weighs price, capacity, and carrier history, then deciding, with a human somewhere between fully in the loop and fully out of it, who actually hauls the freight. The market estimates around this can't even agree with each other. Grand View Research puts the global digital freight brokerage market at $5.87 billion in 2024, heading to $24.36 billion by 2030 at a 27.3% compound annual growth rate. The Business Research Company, working from a narrower scope, starts at $3.56 billion for that same year.
That gap isn't sloppy research so much as an argument about what counts as "digital" brokerage versus full freight matching. And it matters, because it's the same fuzziness vendors lean on when they describe what their tools actually do. Whether the real base number is $3 billion or $7 billion, the growth rate says something the dollar figure can't: manual carrier selection is getting priced out, one lane at a time. MHI's 2026 Annual Industry Report found 41% of supply chain companies now using AI, up from 30% the year before. That's a real jump in a single year. So the question worth asking isn't whether AI is spreading. It's what, specifically, the AI is doing that used to require a human on the phone with a dispatcher.
The phrase gets slapped on everything from a rules-based routing guide with a scoring column bolted on, to a fully autonomous system that books a load without anyone reviewing it. Those are different products wearing the same marketing copy, and I've sat through demos where I genuinely could not tell you which one I was looking at.
Underneath the label there are usually three jobs bundled together, and they don't always come from the same vendor or work the same way. Carrier scoring and ranking figures out which carrier fits a given load. Tender automation sends the offer out and manages acceptances and rejections without a person watching each one. Procurement optimization rebids or renegotiates contracted lanes as conditions shift, instead of waiting for the annual RFP cycle to roll around. A platform might do one of these well and the other two barely at all, and the sales deck will not tell you which is which.
There isn't one algorithm you point to and evaluate on its own. Most production systems stitch together several model types and several data pipelines, and it's usually the seams between them where things go wrong, not the models themselves. A platform that recommends a carrier is a fundamentally different thing from one that executes the booking, and that difference carries real weight for liability, for compliance, and for how much control a shipper keeps. Gartner's 2025 survey found only 23% of supply chain leaders have a formal, organization-wide AI strategy. Most buyers, in other words, are shopping for these tools without a clear internal answer to the most basic question: what do we actually want the system to own, and what do we keep for ourselves?
The data inputs that make carrier scoring possible
No model beats the data feeding it. It's still the best filter for sizing up any carrier selection claim someone puts in front of you.
The primary inputs are mostly structured, mostly familiar to anyone who's run a TMS: historical lane-level performance (on-time rates, claim rates, tender acceptance by carrier and lane), real-time capacity signals like truck location, hours-of-service availability, deadhead distance, spot rate benchmarks, contracted rate indices, fuel costs, surcharge schedules. It's plumbing. Nothing exotic about it.
What's more interesting is what leading platforms have started layering on top. Weather and traffic data get mapped against predicted transit times. Some platforms pull in carrier financial health signals to flag capacity risk before it turns into a service failure. And there's a genuinely useful move happening with unstructured data: C.H. Robinson's Orders Agent reads emailed tenders, the ordinary messy inbox kind, and builds a complete order from them in about 90 seconds. Freight has run on email and phone calls for decades; getting a model to parse that reliably enough to trust is its own engineering problem, separate from the scoring math entirely.
Garbage in, garbage out applies here with unusual force. Platforms sitting on more lane-level history have a head start newer entrants can't shortcut around. Scale isn't just a sales talking point in this market, it's the raw material the model needs to be any good. Project44's AI Agent Orchestration system has reportedly initiated close to a million automated carrier communications over the past year and claims that automation improved data quality by up to 30%. Automation feeding better data back into the model that runs the automation, that's the loop worth watching, because it's how these systems build a lead over time that gets harder to close. When you're sizing up a platform, ask what data it actually has on the lanes you run. Not what data types it can theoretically ingest someday.
How the decision logic turns those inputs into a carrier selection
The dominant pattern in production today is the tender waterfall, and it's not complicated once someone lays it out for you. The shipper or broker sets weighting rules for cost, transit time, reliability, carrier preference. The system scores every qualified carrier against those weights and ranks them. Tenders go out in rank order; if carrier one turns it down, the system moves to carrier two without anyone picking up a phone. The weights and the underlying data refresh continuously, so the same lane can produce a different top pick from one load to the next.
Underneath that waterfall sit a few model types, and they're not interchangeable. Gradient-boosted tree models handle tabular carrier performance and rate data well, which is most of what freight data actually looks like. Transformer-based sequence models pick up patterns across time, like seasonal demand swings on a lane. Reinforcement learning shows up in platforms aimed at optimizing the long-run carrier relationship rather than just who wins today's load. Different tools doing different jobs, not one model wearing three hats.
What separates this from the old rules-based routing guide is the goal itself: the system weighs cost, reliability, and strategic fit at the same time, instead of just chasing the cheapest bid. That's multi-criterion optimization, and it changes what "best" even means here. When a load falls outside the routing guide entirely, or gets rejected all the way down the waterfall, some platforms broadcast it to an AI matching network, collect ranked recommendations, and book the load without a human touching it. Project44's AI Freight Procurement Agent pushes this up to the procurement layer, replacing the once-a-year bid cycle with continuous sourcing, and it can launch mini-bids on its own when a contracted rate drifts noticeably above market. The models retrain constantly on what actually happened: acceptance rates, on-time performance, predicted cost versus real cost. The platform with more shipment volume behind it keeps pulling further ahead of the one with less. Nobody puts that compounding math on the homepage.
Where the efficiency gains actually show up, and how to read the vendor claims
Start with the number that cuts through almost every vendor pitch: empty miles. Roughly a third of Class 8 truck miles driven in the US happen with no freight on board, well above the 20% figure most carriers use internally for planning. AI matching is reportedly cutting empty miles by 30% or more on platforms with enough network density to make good matches. That number holds up because it's measurable and hard to dress up. It isn't a self-reported margin claim you have to take on faith.
The vendor-reported numbers deserve more scrutiny, not because they're fabricated, but because of where and when they get measured. Leading freight intelligence platforms report freight spend reductions in the 8 to 18% range and tender acceptance improvements of 20 to 35% over manual or rules-based methods. Project44's early Freight Procurement Agent deployments reported a 4.1% freight spend reduction, sourcing cycle times cut by as much as 75%, manual coordination effort down 70%. Brokers on these platforms say they're managing meaningfully more loads per week than before, and that recaptured time might be the most believable number in the bunch, even as the spend-reduction figures swing by vendor and by lane.
Treat these figures as ceiling numbers: achieved under favorable lanes, favorable markets, with mature data behind them. Not what to expect your first quarter live. So ask different questions instead. What's the acceptance rate on your specific lanes, not the network average? Does the optimization hold up when spot rates spike, or does the model lag because it's still running on last week's patterns? Over what window were the quoted gains measured, a few good weeks or a full freight cycle that includes peak season chaos? McKinsey's 2025 logistics research puts a possible ceiling on cost-to-serve reduction from AI at up to 30%, a fine number to keep as a reference point. But the space between best-case and median outcome is exactly where your due diligence needs to live, and most vendor decks would rather you not look there too closely.
How major platforms have implemented AI carrier selection in practice
C.H. Robinson has gone all in on scale. More than 30 AI agents run across its network, automating over 3 million tasks. The Quoting Agent turns around customer-specific price quotes in 32 seconds and has processed over a million of them. The Orders Agent, the one reading emailed tenders, builds complete orders in about 90 seconds and handles thousands of truckload orders a day. In October 2025 the company launched what it calls the "Agentic Supply Chain" category, which is as much a flag planted in the ground as a product announcement.
Project44 took a narrower, deeper path, aimed specifically at the procurement layer. Its Freight Procurement Agent benchmarks contracted rates against live market pricing, flags savings opportunities, and runs continuous mini-bids instead of waiting for the next RFP cycle. Smaller slice of the problem than C.H. Robinson's approach, sure, but it's a slice most shippers have historically handled with a spreadsheet and a once-a-year scramble.
Uber Freight took a different tradeoff. Its LLM-powered network of 30 AI agents, launched in the second quarter of 2025, handles end-to-end shipment execution and includes voice-based agents that negotiate rates and have cut driver hold times on booking calls by a wide margin. It keeps a human in the loop for verification, a deliberate design choice, and probably the more defensible one for anyone worried about accountability down the line.
DAT One went the acquisition route, buying Convoy's automated freight-matching technology and folding it into its existing platform tiers, a reminder that the AI is only ever as good as the carrier network sitting behind it. And Trimble unveiled an AI-powered TMS in November 2025 built around seven modules, each with its own dedicated AI agent covering order acceptance, load building, planning, fleet readiness, and live tracking. It's embedded in the TMS itself, rather than a standalone procurement tool.
Five companies, five different bets. None of them are ripping out the carrier network or reinventing routing logic from scratch. They're all speeding up decisions inside workflows that already existed, which, if you squint, is a less thrilling story than the press releases suggest but probably the more honest one.
The limits and failure modes logistics leaders should anticipate
Thin lanes are where these models get exposed. A model trained on high-volume lanes has plenty of signal to work with. On a low-frequency or one-off lane, that signal thins out fast, and that's exactly the kind of load where a human broker's judgment has always earned its keep. Hand that call to a model trained on someone else's data and the system has no real basis for the routing choice it's making.
Market dislocation is the other weak spot, and it's structural, not incidental. During a sudden spot rate spike, a weather event, a port shutdown, a demand shock, a model trained on historical patterns can hand back a recommendation that's technically optimal given what it learned last month and flatly wrong given what's happening this afternoon. How fast the model updates ends up mattering more than how sophisticated it looks on paper.
There's also a category of judgment these systems can score only poorly, if at all. A carrier who bids slightly higher but has a longstanding relationship with a specific shipper location. A new carrier entrant with zero performance history in the system, who might be excellent but looks invisible to the model. Strategic capacity a shipper has quietly reserved that never shows up in spot or contracted rate data. None of that lives anywhere the model can query it.
That 23% figure from Gartner keeps coming back up because it's the crux of the governance problem. Platforms are getting more capable of automating decisions; what's missing, in most organizations, is a policy on which decisions should be automated and at what confidence threshold. Automate without that policy and you've opened a new gap instead of closing one. Uber Freight's choice to keep humans in the loop on rate negotiation looks, from this angle, less like a limitation and more like an admission that removing judgment entirely creates accountability problems, particularly once carrier disputes start showing up. Underneath all of it sits a boring, expensive fact: most shippers run legacy TMS infrastructure that was never built to receive real-time AI recommendations, and the integration lift needed to bridge that gap gets underestimated in nearly every vendor sales conversation I've sat in on.
What logistics and operations leaders should actually evaluate when assessing these platforms
A handful of questions cut through most of the noise. Which of the three capabilities, scoring, tender automation, procurement optimization, does the platform actually deliver today, and which is still sitting on a roadmap slide somewhere? How deep is its carrier network on your specific lanes, given that network scale shapes model quality more than algorithmic cleverness ever will? Does the system recommend a decision or execute it, and if it executes, what override controls and audit trails kick in when something goes sideways? How does it hold up on thin lanes and during market stress, not steady-state performance, but actual case data from a genuinely bad week?
Integration matters just as much as the model underneath it. How does the platform talk to your existing TMS: through an API, a native integration, or does it expect to replace what you already run? And does your shipment outcome data flow back to improve the model, or is it a one-way street where you feed the system and never get smarter in return?
The build-versus-buy question tends to sort itself by size. Large 3PLs building proprietary agent fleets, the C.H. Robinson model, only makes sense at real scale. Mid-market shippers are more realistically choosing between an embedded TMS approach like Trimble's or a procurement-layer tool like Project44's. Either way, the useful framing stays narrow: which capability actually closes the gap sitting in your current carrier selection process, right now, on the lanes you actually run? Everything past that question is just marketing with better production values.
