Supply chain AI agents: Where should autonomy begin and end?
AI agents can scan emails, audit invoices, flag billing issues, develop transactional data and negotiate contracts along the supply chain -- but they can't execute the deal (yet).
There could be a time in the future when supply chains completely run on autopilot. Autonomous AI agents would negotiate contracts and complete deals while humans watch for any glitches in the automated process -- but not anytime soon.
AI agents have proven capable of handling low-level supply chain work, such as invoice reconciliation and research. They can even negotiate a contract, but they aren't allowed to execute the deal.
"Tomorrow there is true promise, but it is unproven," said Hal Hallsson, head of enterprise architecture at freight shipper Bison Transport. The data collection along the supply chain is "extremely complex," he added. Some of the hesitancy among transportation companies to rely on supply chain agents for more complex tasks is specific to an industry that typically takes a conservative approach to new technologies.
"I think the concept of trusting anything that is not human is really what we're all grappling with at this point," said Mark Riskowitz, senior vice president of operations at cookware maker Caraway Home.
Although many supply chain tasks are being automated, the concern is how much autonomy agents should be allowed in making decisions. "What decisions do you delegate? What threshold triggers do you need before human intervention is required? How do you measure an AI agent's performance? Who's accountable when it doesn't go right?" asked Abe Eshkenazi, CEO of the Association for Supply Chain Management.
We can delegate a decision to technology. You can't delegate accountability.
Abe EshkenaziCEO, Association for Supply Chain Management
Businesses might let AI agents reroute shipments, change suppliers or approve payments, but only with trusted data and clearly defined bounds, Eshkenazi said. They could also negotiate with suppliers, but under strict limits. What isn't being delegated to agents is responsibility for the outcome. "We can delegate a decision to technology," he said. "You can't delegate accountability."
Humans bring more than accountability, said John Burns, senior director of financial systems and transformation at equipment rental company Herc Rentals. "The one thing an agent cannot do is … read body language," he explained. "It can't tell what the other person's thinking." Routine, pre-negotiated purchasing can be automated, and an agent can research, compare bids and even make a counteroffer, Burns said, "but the human has to sign it."
Real-world agent applications
At Caraway, AI agents from supply-chain platform maker Prysmic AI audit contracts and invoices, flag billing issues and develop data for negotiations, Riskowitz said. While the agents are capable of negotiating directly with suppliers, they're not allowed to perform that function, he said.
With tight guardrails and narrow job descriptions, agents are handling contract disputes and routine negotiations, said Kelly Breakstone Roth, CEO and co-founder of Prysmic. In the next year or two, she believes businesses will increasingly let agents negotiate directly with one another, while humans handle more strategic relationships. "For most of our customers today, there's an approve button," she said.
Bison isn't ready for automated approval, but AI agents scan emails and faxes, assign a confidence score, automatically notify customers about any missing information, enter completed orders into the company's transportation management system and, if necessary, flag an order for human review, according to Hallsson. The order intake process is "very rigorous" and structured, and exceptions can be routed to people, he said.
As for allowing greater autonomy, Hallson said agents are error-prone, don't always understand business context and mistakes can compound faster than humans can catch them. Bison will build its own small language model to handle the acronyms and specialties specific to the transportation industries, he said, because the "big frontier models do not offer me that."
Language barriers to agent autonomy
The language problem isn't specific to the trucking industry. Large language models (LLMs) are probabilistic. They can produce different outputs from one execution session to the next even when the inputs don't change.
We would be cautious about granting open-ended authority.
Carol LongResearcher, Harvard University
Researchers at Harvard, MIT, Purdue and Georgia Tech ran agents through a supply chain simulation with identical data and found that orders varied from run to run. In one test, the variation widened as it moved upstream from retailer to wholesaler to distributor to factory -- a phenomenon they call the agent bullwhip effect.
"Agents can make inconsistent decisions across otherwise identical situations and, on occasion, produce disproportionately large errors. We would be cautious about granting open-ended authority," said Carol Long, Harvard researcher and lead author of the September 2026 paper "Reliability and Effectiveness of Autonomous AI Agents in Supply Chain Management."
Eranova, maker of an agentic AI software for logistics and transportation back-office operations, adds a layer of deterministic tool code between the LLM and ERP system, according to Ethan Barnes, the company's head of engineering and deployment. Guardrails in the tool calls keep probabilistic agents within requirements, he said. For instance, if the model requests "apply cash," the system first runs reconciliation checks. "The tool calls themselves should have the guardrails written in that can protect the use case and give the confidence that's needed," Barnes noted.
What exists in reality today are more transactional commodity-based negotiations with tight guardrails in the long tail.
Andrew Roszko CEO, Jaggaer
"The way that companies, even the world's largest companies, think they operate is very different from the way they actually operate," said Nitin Jayakrishnan, CEO and co-founder of procurement automation platform provider Freehand. The AI platform's agents can analyze about five years of contracts, invoices and transactional data to set the operating policy for customers to approve, thereby controlling agent decision-making.
Jaggaer, a procurement software provider, is using AI agents to negotiate contract renewals, said CEO Andrew Roszko. The system uses historical pricing and supplier performance plus delivery data to renegotiate lower-tier contracts. Otherwise, those contracts might renew automatically, sometimes at unfavorable prices. "What exists in reality today are more transactional commodity-based negotiations with tight guardrails in the long tail," Roszko said.
Spencer Penn, CEO and co-founder of procurement and strategic sourcing platform maker Lightsource, said the platform's AI agents can solicit bids, analyze them, draft responses and suggest counteroffers, but no customer has requested an agent that makes a final award without human approval. As to how much leeway agents should have in negotiating contracts, he said any restrictions are "by design, but not by limitation of technical capability."
Patrick Thibodeau has worked for several decades as an enterprise reporter, focusing on IT and workforce management, ERP, high-skills immigration, tech policy and high-performance computing.
