Cold chain AI earns its keep when QA managers never log in

The QA manager PAXAFE wants to build for never asks about the shipment. Speaking in a MarketScale Studio segment, the company describes a quality assurance manager, the person who decides whether a temperature-sensitive shipment can be released, who is never told about a problem shipment, never has to log in and never hears from the carrier, because a software agent closed the loop first. That, PAXAFE argues, is the line between a good agentic system and the rest of the market.

The interviewer asked what question a buyer should put to vendors to tell AI cold chain technologies apart. PAXAFE's answer was a four-rung ladder. Most products, by the company's count, sit on the second rung.

The timing matters because the trade press is now drawing the same line. In a September 30, 2026 article, Supply Chain Management Review reported that agentic AI is being applied to supply chain exception management, the handling of shipments, suppliers and inventory plans that go off course, and that a forecast or alert has limited value if the exception then waits in a queue for a person to investigate and resolve it. The same article set a boundary that PAXAFE's framing leaves implicit: the level of autonomy granted to an agent must match the risk of the decision.

Watch and answer describe most of the market

Watch, in PAXAFE's model, is a thin software layer on top of telemetry, the stream of temperature, location and status readings that sensors on a shipment send back. The company says arguably everyone does that now. It is the baseline a buyer should expect from any monitoring product, not a point of difference.

Answer is where the company places itself. PAXAFE puts most cold chain platforms, its own included, at the answer stage.

The chat layer exists, in PAXAFE's telling, because a single dashboard cannot anticipate every question a quality team, a logistics team and a carrier might ask about the same shipment record. A conversational interface lets each of them ask in their own terms. The company calls it a logical next evolution and says everyone is doing chat, or at least trying to.

Controlant, which monitors pharmaceutical shipments, describes that stage in its own words. In an undated press release on its site, the company said its platform delivers conversational intelligence that lets users interact with supply chain data in natural language, and that it exposes curated shipment, excursion and quality data to AI agents through patterns aligned with the Model Context Protocol, a published way for AI systems to read context from other software. Carsten Lützhøft, Controlant's chief commercial officer, said in the release that the company extends this with the ability to build agentic AI solutions for its customers.

For a buyer, the test at this stage is whether the system can say, in plain language, why a specific shipment went out of range, and whether that answer rests on validated data. Controlant's release ties its answers to GxP-relevant data. GxP is the family of good practice regulations, including good distribution practice, that govern how pharmaceutical product is made, stored and moved.

Suggest turns an answer into a next step

PAXAFE's third stage is the one it says separates a nice conversation from a result. The company frames it as the question that follows any answer: how does the operator put it into practice, and what happens next so that the answer translates into physical revenue. A good answer that ends up on a slide deck for an executive to admire, in PAXAFE's view, does not change what happens to the shipment.

In cold chain terms, suggest means the system proposes a response to an excursion, the period a shipment spends outside its allowed temperature range, rather than only reporting it. Accepting that proposal is still a human's call. That approval step is the whole distance between suggest and act.

Gartner's staged view starts near the same point. In a May 21, 2025 press release, Gartner said agents in the first stage of agentic AI appear as task-specific intelligent assistants designed to aid humans, and gave the example of a procurement agent tasked to buy supplies based on stock levels, projected demand and market conditions. In the same release Gartner announced its prediction that by 2030, 50% of cross-functional supply chain management solutions will use intelligent agents to autonomously execute decisions.

Read against PAXAFE's ladder, Gartner's first stage sits between answer and suggest, and its 2030 prediction describes a market where half of the software has reached act. A cold chain buyer evaluating a platform this year is choosing a vendor against that five-year horizon.

Act closes the loop before anyone asks

PAXAFE describes act as the stage it is actively building toward internally. The company lays out the sequence an agent has to complete on its own.

Three things happen in that chain. The agent detects a condition, decides whether that condition should have occurred, and picks a response by comparing the situation with what it has done before. Then it executes without waiting.

The company's measure of success is what the human experiences, which is nothing.

PAXAFE describes the result as a living, breathing system under the hood that triages shipments for the user without the user ever knowing an agent did the work. By the company's own count, very few products, if any, can do that today.

PAXAFE does not say in the interview which actions its agent would be allowed to take on its own, such as rerouting a shipment, rejecting it, alerting a carrier or opening a claim, or where a human approval sits in that sequence. For a quality team, that list is the specification that matters, because the actions an agent is permitted to take are what a release decision and an audit will have to account for.

project44 and FourKites say they already act

PAXAFE's count of who has reached the act stage runs into other vendors' own claims. project44, in product material currently on its site, markets what it calls AI agent orchestration, powered by 57 agents. The page describes an exception management agent that retrieves reason codes for late deliveries and confirms bookings when ocean shipments are at risk of being rolled, an execution recovery agent that detects carrier cancellations, ranks alternatives from the routing guide, meaning the shipper's ranked list of approved carriers for a lane, and calls carriers for quotes, and a freight procurement agent that negotiates rates and books freight within guardrails the customer's team defines.

FourKites puts a number on it. On its site, the company says its carrier operations agent, named Tracy, monitors every shipment, contacts carriers automatically and updates stakeholders, and it reports a track record of 75% of exceptions resolved autonomously. The same page runs a counter showing more than 3.5 million supply chain issues acted on by its agents.

Both figures are the companies' own, published on pages built to sell the products. Neither page walks through a single exception from detection to closure with no person involved, which is the evidence a quality team would need before trusting an agent with a temperature excursion.

The overlap with cold chain is also partial. The exception types project44 lists are freight events: late shipment reason codes, rolled bookings, route deviation, missed cut-offs at the port of loading. FourKites names pharma among the industries for its compliance and documentation agent, which collects proof of delivery and validates customs paperwork, rather than for Tracy. PAXAFE's bar, an excursion handled before a QA manager hears of it, is a different kind of loop from rebooking a canceled truckload, and the interview's claim that very few can close it is specific to that loop.

Gartner counts about 130 real agentic vendors

The analyst numbers argue for doubting every vendor's self-assessment, PAXAFE's included. According to a June 25, 2025 Gartner statement, over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. Gartner described a practice it calls agent washing, the rebranding of existing products such as AI assistants, robotic process automation and chatbots without substantial agentic capability, and estimated that only about 130 of the thousands of agentic AI vendors are real.

Anushree Verma, a senior director analyst at Gartner, said in the same statement that current models do not have the maturity and agency to autonomously achieve complex business goals or follow nuanced instructions over time. That is a direct bound on the act stage: the reasoning step PAXAFE describes, choosing an action from what the system has done in the past, is the step Gartner says today's models handle least well.

Executive surveys say the industry is spread across the ladder. In a sponsored report on the IBM Institute for Business Value site, produced by IBM with Oracle and Accelalpha in 2025 and based on a survey of more than 300 chief supply chain officers and chief operations officers run with Oxford Economics, IBM wrote that 31% of supply chain executives were executing and scaling proofs of concept for self-sufficient AI agents, with another 22% still developing a proof of concept. IBM's report also describes the engine: large language models paired with fit-for-purpose small language models, which it says can be tuned to a domain such as dynamic logistics.

Put those pieces together and the technical path to act looks like a stack. Controlant's release describes the data layer, validated shipment and excursion records handed to models through the Model Context Protocol. IBM's report describes the model layer. What PAXAFE adds is the memory layer, a record of past actions and their outcomes that an agent can reason over, and that record is something a vendor has to accumulate shipment by shipment before an agent can act on it.

Autonomy has to match the risk of the shipment

Supply Chain Management Review's September 2026 article suggested where to start: routine, high-volume exceptions with clear rules and a measurable cost of delay. Gartner's May 2025 release made the same point as a recommendation, telling chief supply chain officers to define operational parameters so agents work within a specified scope and do not take incorrect actions with negative consequences. A missed dock appointment fits that description. A biologic sitting two degrees warm for an unknown number of hours does not, which is why the cold chain is a harder place to reach act than general freight.

Controlant's release names the rules that govern the records an agent would act on. It said its platform complies with FDA 21 CFR Part 11, the U.S. regulation that sets the conditions under which electronic records and electronic signatures are accepted as equivalent to paper, and with EU Annex 11, the European guideline on computerised systems used in regulated manufacturing. Controlant said that means audit trails, access controls and data integrity safeguards so that AI-driven insights and automation align with GxP expectations.

Those rules make an agent's actions traceable. They do not say who carries the cost when the action is wrong. Who answers for an agent's wrong call on a pharmaceutical shipment, the shipper, the carrier or the software vendor, is a question neither PAXAFE's interview nor Controlant's release settles, and a quality team should expect to write that answer into its own procedures before any agent is allowed to release or hold product.

The practical move for a buyer is to ask PAXAFE's question and then two more. Ask the vendor which stage its product has reached, then ask for the list of actions the agent takes without sign-off and for the record of past actions it reasons from. A vendor at the answer stage will have a chat window to show; a vendor at act should be able to show one excursion, start to finish, that no person touched.

The signal PAXAFE points to is the QA manager who never logs in. The signal Gartner points to carries a date: in its June 2025 statement it predicted that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024. If cold chain platforms are closing excursion loops on their own by then, the four-stage question will have become the only one worth asking.