Project44 Splits Into Two, Caps Staff at 600, and Bets on 'Infinite Labor

The Split: Two Companies, One Bet on Infinite Labor

Project44 split into two companies on July 14, betting that AI agents can replace the human coordination layer that has always run global freight. The original entity keeps the Project44 name serving shippers: manufacturers, retailers, and brands moving freight. A new company, LSP44, takes the logistics service provider segment: brokers, 3PLs, and forwarders. Each gets its own engineering team, sales organization, and roadmap. Autopilot, launched in May 2026, is a no-code platform that lets supply chain teams deploy AI agents across workflows without writing prompts, building integrations, or hiring prompt engineers. Founder and CEO Jett McCandless calls the thesis "infinite labor": agents absorb the repetitive, high-volume coordination work that has always required human bodies.

This is not a reorganization. It is a bet that the economics of logistics software have inverted. For years Project44 built "Signal," a data graph connecting 259,000 carriers across 186 countries, processing 1.5 billion shipments a year, as project44's PR Newswire release reports, and ingesting 700 million logistics events daily. On top of that, the company built "Trigger," an exception engine released in 2022 that turned the firehose of raw shipment data into prioritized work items. Those two foundations took a decade. A model upgrade cannot replicate them.

Autopilot: The Execution Layer

Autopilot sits on that foundation. It gives users a visual workflow canvas where AI agents respond to real-time logistics signals — a late shipment, a missing PRO number, a container dwelling too long at the port of discharge — and act autonomously. The pre-built library covers roughly 40 workflows today: collecting truck milestones, late-shipment reason codes, missing PRO numbers, carrier outreach for ETA confirmation. Project44 ships two to three new workflows a week. Customers configure conditional branching to target specific carriers, lanes, or shipment types, and define post-agent operations: send notifications, create tasks, add notes to shipments, or route work to a human operator. A draft-and-publish model lets teams iterate safely. Movement's Collaboration Center logs and audits every agent action.

Early results suggest the thesis has teeth. Agent-driven outreach and resolution interactions grew 60-fold, from 500 per week to more than 30,000. Agents have completed nearly one million automated carrier communications, improving carrier data quality by 30 percent and cutting data-issue resolution time by three-quarters. Eastman Chemical Co., an early Autopilot customer, said the platform let its team expand into APAC and onboard less-technical carriers without adding operational complexity. "The right work reaches the right people without manual intervention," said Josh Moss, Eastman's process lead for global supply chain. Project44's own figures show a 4 percent reduction in freight spend, a 70 percent drop in manual coordination, sourcing cycles up to 75 percent faster, and a 40 percent reduction in disruption-related costs.

McCandless argues that the agentic AI startups crowding logistics conferences are not competitors. They are blank-canvas tools without the underlying data, system of record, or distribution to deploy at scale. "They don't have any context," he said. "So all these companies have to create data lakes, normalize the data, sync it with their TMS and CRM and all these other systems. Then they have to hire prompt engineers. They have to connect into Project44 and extract the data. And then piece agents together to communicate and hand information back and forth. Most of the projects are failing." The moat, he says, has shifted to three things: distribution into the largest shippers and LSPs, proprietary data, and a culture that can absorb constant change.

Notably absent from Project44's positioning is the phrase the rest of the category has spent five years trying to own: control tower. McCandless rejects the metaphor. Control towers evoke air traffic: tens of thousands of human controllers operating a system built on 1960s technology. That is the opposite of what Autopilot does. The company seriously considered calling itself an "intelligent operating system" when it rebranded to a decision intelligence platform last summer, but held off until it could ship product that delivered on the framing. Autopilot is that product.

The pricing model signals the same shift. Autopilot is currently free. "We actually don't charge for it at all right now," McCandless said. The implied model is outcome-based: Project44 will share in the savings or efficiency Autopilot delivers rather than billing on seats, tokens, or per-agent fees. Internally, the company lives the transition. Headcount peaked at 1,200 and now sits at 582. McCandless said the company will never exceed 600 people. The reason: AI now writes roughly 85 percent of Project44's code. The organization that sells infinite labor runs on it.

Why Dashboards Aren't Enough

Logistics software has spent two decades perfecting the art of showing you the problem. Control towers, dashboards, and transportation management systems made supply chains more transparent than ever. A dispatcher could see a delayed shipment, a missed SLA, a sudden demand spike, all in real time. But visibility alone does not drive outcomes. A dashboard tells you a delivery is delayed; it cannot resolve the delay. When disruptions occurred, the system flagged the issue. Resolution still required human intervention. This gap between insight and action is where traditional systems fall short.

Agentic AI closes that gap by replacing repetitive decision-making loops altogether. Instead of generating recommendations for a planner to approve, agentic systems make decisions and act on them autonomously, within defined business rules. They do not wait for human approval at every step. They continuously evaluate conditions, choose the best course, and execute in real time. The distinction is subtle but profound: traditional AI supports human decision-making; agentic AI replaces the decision-making loop for routine execution, freeing humans for oversight and strategy.

The highest-value logistics architecture combines both: Generative AI interprets and explains. Agentic AI decides and executes. Generative AI closes the communication loop.

The workload being automated is not trivial. Logistics is a complex constraint optimization problem. Every decision must balance cost efficiency, SLA commitments, vehicle capabilities, driver availability, regulatory requirements, and customer preferences, variables that often conflict. Traditional systems simplify this complexity, often leading to suboptimal decisions. Agentic AI systems evaluate multiple constraints simultaneously. They navigate trade-offs in real time, deciding whether to prioritize cost savings or SLA adherence based on business priorities. This ability to process complexity at scale enables truly intelligent logistics execution.

Consider what gets replaced across the workflow. Agents autonomously tender loads to spot markets when contract carriers are unavailable, negotiating rates with digital freight brokers within pre-set margin guardrails. They continuously monitor traffic, weather, and delivery windows, re-routing vehicles and updating ETAs in real time to minimize fuel and mileage. SLA management shifts from reactive monitoring to proactive enforcement: agents detect "at-risk" orders and autonomously upgrade service levels or split shipments to meet the delivery promise. Instead of a planner firefighting each disruption, agents detect anomalies, query real-time conditions, and act autonomously for routine decisions. Nearly one in four truck journeys in Europe runs completely empty. Agentic routing engines combat this by continuously running predictive capacity algorithms, co-mingling pickup and drop-off loads within the same geographic zones.

The architecture required is not a thin AI layer on top of existing systems. It demands an execution architecture that can sense operational change, evaluate constraints, act through enterprise systems, and maintain governance at every step. A practical reference architecture includes source systems, an event ingestion layer, a unified data layer, decisioning agents, optimization engines, an orchestration layer, execution systems, monitoring and observability, and governance controls. Agentic logistics systems are only as strong as the data environment they operate in; they need connected, trusted, timely information across the network: order data, shipment data, inventory data, fleet and driver data, carrier data, telematics and IoT data, customer promise data, cost data, historical performance data.

Early adoption data signals the shift is underway. PwC's May 2025 AI Agent Survey found 79 percent of companies already adopting AI agents in some capacity, with 88 percent of executives planning to increase AI budgets over the next 12 months specifically for agentic AI. The 2026 State of AI Agents Report notes 57 percent of organizations now deploy agents for complex, multi-stage workflows rather than simple isolated tasks. Measurable outcomes track through deployment-specific KPIs: cost per delivery, SLA adherence, route productivity, fleet utilization, dispatch automation, exception-resolution time.

The human role doesn't disappear; it moves up the stack. Agents handle routine monitoring and exception detection. Humans focus on complex judgment calls and relationship management that keep accounts growing. The new partnership lets supply chain leaders focus on resilience strategy and network design while agents autonomously manage the flow of goods and information. The real question is not whether your systems can predict. It is whether they can act.

CargoWise's Fortress Shows Cracks

CargoWise's fortress is showing cracks. For more than a decade the WiseTech Global platform served as the default operating system for large and mid-sized freight forwarders. Its breadth and global customs footprint — built through acquisitions of country-specific customs vendors — created what one forwarder described as "ripping out a person's circulatory system while the person was still alive" to replace. That moat held until December 1, 2025, when WiseTech moved roughly 95 percent of customers to new Value Pack pricing. CEO Zubin Appoo called it "not a pricing change, but a totally different commercial model." Forwarders called it something else: cost increases of 20 to 50 percent, "unprecedented user unrest" per The Loadstar, and a breach of the implicit contract that made CargoWise the platform of record.

The revolt created an opening. Kareem Naouri, co-founder and CEO of Logistaas, said the idea that CargoWise had no credible competitors was outdated. "That might have been true 10 years ago, but it certainly isn't the case today," he said. "There are other serious providers in the market, operating globally, integrating across the industry, and using AI, all at a lower cost." Logistaas publicly committed to a maximum annual price increase of 3 percent. Descartes, Germany's Riege, Yojee's Mosaic, and GoFreight, citing 4- to 8-week implementation versus 6 to 12 months for CargoWise, all appear in forwarder shortlists now. Magaya bundles forwarding with a built-in WMS starting around $200 per user per month, generally cited as more predictable and faster to implement. TechnBrains' audited sample of 133 practitioner observations found pricing, billing, contracts, and renewals accounted for 43 percent of negative feedback, the largest category by far, with 90 percent of negatives occurring after selection, concentrated in ongoing operation and renewal.

Project44's split into LSP44 and its Autopilot launch arrives into this fracture. Rather than demand a rip-and-replace, Project44 positions its platform as an execution layer connecting transport management systems, ERP software, carrier networks, and external data sources into a single logistics graph. "Value migrates to the execution layer," the company argues. "When AI does the work and captures not just the data but the reasoning behind it, the question of which database stores the record matters less than which system does the job and holds the intelligence." That model — AI orchestrating across fragmented estates rather than inside a single application — mirrors what 5U AI, a startup that recently raised $3.2 million in pre-seed funding, is building. CEO Yagiz Abik said replacing a TMS was "a two-year migration project," while the immediate opportunity lay in automating the work itself. "Our workers treat the TMS as the system of record, whichever one it is," he told The Loadstar. "We've integrated with systems most Silicon Valley companies have never heard of, because that's the reality of European freight."

FreightSuite has taken the boldest position: an AI-native platform from the ground up, arguing legacy systems were never designed for autonomous operations. Robert Petti, founder of Prompt Global, believes the industry's real differentiator will not be whether AI sits inside or outside the TMS, but whether it has access to high-quality operational data, business context, and robust governance. In his view, companies risk becoming distracted by architecture when the harder challenge is giving AI the information it needs to make reliable decisions. Gartner forecasts SCM software with agentic AI capabilities will grow from less than $2 billion in 2025 to $53 billion in spend. Fifty-five percent of 509 supply chain leaders surveyed expect agentic AI to dominate initiatives in 2026. Yet 56 percent cite integrating AI with legacy systems as a major challenge, and half lack the internal talent to implement it.

The legacy vendors are not standing still. CargoWise's customs moat remains real — integrated with around 20 countries where competitors have one, two, or none — and the Freight & Trade Alliance negotiated a Transitional Pricing Protection mechanism with WiseTech for affected members. But the pricing revolt proved that lock-in cuts both ways: it keeps customers, but it also fuels the search for alternatives when the commercial model shifts. The next battle won't be won by the deepest customs library. It will be won by the system that makes the decisions.

The Rebuild: Skills, Architecture, Operations

The split at Project44 signals a broader rupture: freight technology teams can no longer treat AI as a feature layered onto dashboards. The shift demands new skills, new architectures, and new operating models, all at once.

Start with the skill set. Prompt engineering is already obsolescing. Loop engineering has replaced it: designing systems that handle the iterative prompting, verification, and correction cycles themselves. A good goal for an agent is specific, with an end condition and a stopping condition. The human's job becomes writing intent down, keeping it current, and keeping it visible to the agent. Engineers who spent years perfecting UI flows for dispatchers now need to model decision boundaries, define escalation paths, and instrument every agent action so a person can reconstruct what happened, why, and against what data. Comprehension debt — the gap between total code in a system and the degree to which humans understand it — accumulates fast when agents write code and humans check less of it. Intent debt follows: an agent makes a technically sound change that violates a project objective. Cognitive surrender lurks when operators blindly accept loop output. The antidote is a human in the loop who reads and verifies shipped code.

Architecture must change too. The research is unanimous: the agent is only as good as the semantic, structured, real-time data layer beneath it. GS1 EPCIS 2.0 event data, harmonized carrier master data, unified shipment IDs, and clean HS/HTS classifications matter more than the choice of orchestration framework. Modern supply chain agents run on LangChain, CrewAI, and AutoGPT, often orchestrated through multi-agent pipelines that collaborate, share insights, divide tasks, and resolve conflicts dynamically. Logistics workflows are naturally decomposable — carrier communication agent, customs documentation agent, delivery ETA agent — and making them separable makes each easier to test, govern, and replace. Integration is achievable: agents now connect to SAP, Oracle, Manhattan Associates, Blue Yonder, and most TMS/WMS platforms via API. The infrastructure barrier has dropped. The data barrier has not.

Operating models need a graduation ladder. The industry-standard framing is the human-in-command model: every agent starts in shadow mode (recommends only), graduates to approval mode (executes after human sign-off), then to bounded autonomy (acts within named parameters, escalates otherwise). No agent jumps from pilot to full autonomy without real time in the intermediate modes. SAP made this explicit at Hannover Messe 2026: governance controls, audit trails, identity management, cost caps, escalation rules, and kill switches live in the runtime, not in a compliance PDF. ISO/IEC 42001 is becoming the practical benchmark: if your deployment can't answer a conformance question in 2027, you're building on borrowed time.

Metrics must shift from traditional software KPIs (uptime, latency, throughput) to agent-specific measures. Gartner's January 2026 cost-and-value framework introduced four that industry practice is converging on:

These sit alongside SCOR-aligned operational metrics: Perfect Order Fulfillment, Order Cycle Time, Cash-to-Cash Cycle, Supply Chain Cost as % of Revenue. If an agent doesn't move the SCOR needle, it's a demo, not a deployment.

The playbook is concrete. Map high-frequency, multi-step workflows: freight exception management, inventory replenishment triggers, carrier performance reviews, invoice dispute resolution. Audit data readiness honestly: if freight data lives in five formats across three platforms, solve that first. Define human-in-the-loop boundaries with operations and compliance teams: which actions require approval, at what thresholds, how escalation triggers. Start with a bounded, high-value pilot: one workflow, one business unit, one lane. Build cross-functional AI literacy so planners and analysts, not just IT, can define agent goals and validate outputs. Sequence matters: simplify and standardize workflows first, then add intelligence.

Flexport's 2026 Winter Release already ships agents that audit customs entries, calculate tariff refunds, and cut freight costs up to 10 percent. Wayfair is hiring AI agent engineers. Robin handles back-office work end to end. The operators who lead in 2027 aren't the ones with the biggest AI budget. They're the ones who applied SCOR-aligned KPIs to the right initial workflows, built the data foundation before the reasoning layer, embedded governance in the runtime, and had the discipline to keep agents in shadow mode until performance data justified graduation. Skip that discipline, and you'll spend 2027 explaining to the board why the initiative was canceled.

Project44's split was the first major bet on that discipline. The rest of the industry is now racing to catch up.

Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe and Waymo, and the people building the field.