WW/SUPPLYCHAI
Five barriers stand between AI pilots and trusted supply chain planning
ketteQ CEO Mike Landry identifies five barriers to trusted AI agents in supply chain planning, from data readiness and ERP fragmentation to approvals, accountability and time-zone delays.
- By
- Tom Whitfield
- Filed
- Length
- 1,021 words
- Read
- 5 min

Key points05
- Five barriers separate AI-agent pilots from trusted daily planning work.
- ketteQ says its software runs thousands of supply-chain scenarios automatically.
- A five-minute cross-time-zone task can take two or three days under human approval workflows.
- Authority can expand in stages from approved supplier emails to automated messages with daily reports.
- Quintus explores disruptions and their effects on customer orders through natural-language questions.
By Tom Whitfield
Five barriers separate a promising AI agent pilot from dependable daily supply chain planning, according to Mike Landry, CEO of planning, automation and orchestration software provider ketteQ.
Landry’s framework offers no freight-rate, TEU, tonnage or capacity benchmark. It addresses the operating layer that determines whether software can recommend material orders, identify shipments needing attention or contact suppliers.
For shippers, carriers and forwarders, an agent’s value comes from shortening the path from a disrupted plan to a supplier response or customer commitment. Its risk comes from bad records, missing context or permissions that let an uncertain recommendation travel too far.
That distinction matters. Speed without control creates operational noise. Control without speed leaves the decision waiting in an inbox.
What does it take to trust an agent?
“AI can’t be a black box,” Landry says. Planners need to inspect the data, assumptions and logic behind each recommendation, especially when an order change could alter material commitments.
During a pilot, teams should retain access to the systems and analyses they already use. They also need time to compare the agent’s answer with their own work and investigate discrepancies before granting more authority. Independent checking should form part of adoption, not a final compliance gesture.
This approach lets responsibility expand only as users see that recommendations hold up under scrutiny. It also gives the business a practical test: can the planner identify what the software used, what it assumed and why it proposed a specific action?
Can the data support an agent?
Capacity, sales and supply data often sit in separate systems, while poor record maintenance leaves gaps and conflicts. Those defects do more than frustrate reporting. They give an agent an incomplete or contradictory view of the plan it must assess.
Landry recommends starting with the data itself. Teams can ask AI to identify disagreements between records, show the evidence behind each finding, explain the planning effect and specify what requires correction. Users can then add context or inspect the underlying entry.
This gives a pilot a defined job before the agent receives permission to act: improve the information supporting decisions and test whether the system recognizes uncertainty. That matters when a missing field, conflicting record or hidden assumption can turn a plausible recommendation into the wrong order or missed exception.
Can agents work across fragmented systems?
ERP systems hold transaction records that planning tools need. Companies running several ERPs must first assemble a consistent view, but Landry does not expect them to replace those platforms to adopt AI. He says the technology can organize information across multiple ERPs and work with unstructured documents alongside structured records.
The test is not whether an agent can open one application. Planners need it to use the information required for a decision wherever that information sits. That may mean crossing an ERP boundary, reading a document or carrying a recommendation into another vendor’s planning platform.
Separate agents create a further coordination problem. Companies must define how those agents exchange information and instructions, then decide how to resolve conflicting recommendations. A useful pilot follows one task through every dependent system and handoff, exposing access gaps or broken coordination before expansion.
Who approves what?
Before expansion, management should specify:
- which actions the agent may take without approval;
- which decisions require human approval;
- who reviews its actions and can intervene; and
- when incomplete information or changed circumstances require a person to step back in.
A governance board or certification does not remove that operating work. Landry rejects the idea that AI can carry accountability on its own, placing responsibility for judgment with people, particularly when an agent has incomplete information or the wrong context.
Authority can grow in stages. A planner might repeatedly approve an email to a supplier under defined conditions, then authorize similar messages to send automatically with a daily report. The company still needs a clear trigger for human intervention when circumstances fall outside those conditions.
What changes when execution takes days?
An agent’s speed helps little if the next step waits in an inbox. Approvals and handoffs can stretch a short task across several working days, particularly when suppliers and customers operate in different time zones.
Landry points to North American companies using manufacturers in Asia. A question can arrive while U.S. employees sleep; by the time the right people respond, the recipient may have finished work. Another exchange can add a day.
“What might take five minutes takes two or three days,” he says. Companies adopting agents should measure those waits, identify routine exchanges suitable for automation and preserve human review where judgment matters. They should also test whether teams can act on new information as it arrives instead of waiting for the next scheduled planning meeting.
For shippers, carriers and forwarders, that workflow delay can erase much of the analytical gain. Faster detection matters only if the organization can change an order or answer a supplier before the relevant working window closes.
How does ketteQ put this into practice?
ketteQ combines supply chain planning with Quintus, an AI capability that lets planners ask about potential disruptions and their effects on customer orders. The company says its software runs thousands of supply chain scenarios automatically and uses Quintus to explore results through questions about late deliveries, requests for earlier shipments or supply orders needing action to protect future deliveries.
The product links those questions to execution. It helps planners decide how to respond, then coordinates actions among suppliers, manufacturers and customers. Recurring tasks can move from human approval toward limited agent execution, with reporting that lets the user review what happened.
Landry describes the design as combining machine speed with human judgment and context. The practical test for shippers, carriers and forwarders is not whether an AI agent can produce an answer. It is whether controlled permissions, cleaner data and faster handoffs let that answer change a material order or supplier communication without losing accountability. Landry’s model points to an incremental governance path: pilots must establish cleaner data, explicit handoffs and bounded permissions before agents earn broader authority as supply conditions change.
Source: supplychainbrain.com
More from Tom Whitfield
Show full bio
Market editor covering companies, deals and strategy in Ocean and air freight, container shipping at Waybill Wire.
955 articles
Related05
AI agents move into fleet operations, raising control questions
ShipStation Global Puts 250 AI Agents Into Live Freight Operations
57% expect autonomous AI for over a quarter of warehouse decisions
57% of Firms Expect AI to Run a Quarter of Warehouse Decisions by 2028
Flexport opens freight platform to Claude, ChatGPT, and Copilot