The first AI we put into our own dispatch operations was very good at noticing problems. That turned out to be the problem.
It flagged delayed trucks, missing tracking signals, vehicles running off route and loads at risk of missing a delivery window. Every alert was accurate. On a busy day there were hundreds of them, and a small team of planners had to read each one, judge whether it mattered and work out what to do. We had automated the noticing and left the deciding exactly where it was.
Most operational AI is still stuck at that stage. It’s sold as intelligence, but it behaves like a smoke detector: useful, loud and entirely dependent on someone being there to act. The step that changes the economics is letting the system act. That step is also where most teams hesitate, and they’re right to. The question isn’t whether an AI agent should make decisions. It’s which ones?
Alerts don’t scale
An alert costs the system nothing to send and costs a person time to process. Someone has to read it, check it, decide and act. In a freight network running a few hundred trips a day, even a small exception rate produces more alerts than a planning desk can handle well. The team triages. Easy alerts get cleared, ambiguous ones wait, and some important ones get lost in the noise.
Adding more AI to that setup makes it worse, not better. A smarter model finds more problems earlier. Unless something else changes, it just hands people a longer list.
What has to change is who acts. For a large share of routine exceptions, the right response is known in advance and low-risk to carry out. Those are the decisions an agent should take off the list entirely.
Four questions that set the boundary
We’ve settled on four questions for deciding where any given decision sits.
Can it be undone? Re-sequencing drops on a route can be reversed in minutes. Awarding a lane contract isn’t possible. The harder a decision is to reverse, the more human involvement it needs.
What does a wrong call cost? A slightly worse route costs a little fuel. A missed delivery for a key customer, or a payment released against a bad document, costs far more. Autonomy should shrink as the cost of error grows.
How fast does it need to happen? Some decisions lose their value if they wait for a person. Re-slotting a shipment before a dispatch cutoff is only worth doing before the cutoff. Time pressure is a reason to give an agent more room, provided the first two answers allow it.
How sure can the agent be? An agent working from clean, complete data can be trusted with more than one working from patchy signals. A truck that disappears from GPS might be delayed or might just be in a dead zone. If the agent can’t tell the difference, it shouldn’t act as though it can.
Three tiers of authority
Those questions sort decisions into three tiers. The examples come from freight, but the structure carries over to any live operation.
| Tier | What the agent does | Freight examples |
| Act and report | Makes the decision, logs it and tells people what it did | Re-sequencing drops after a delivery slot changes, chasing a missing tracking signal with the driver, and passing a proof of delivery that matches the order. |
| Propose and wait | Prepares the decision with its reasoning and waits for approval | Moving a load to another transporter; changing the dispatch plan close to a cutoff; rejecting a proof of delivery with discrepancies |
| Escalate with context | Hands the decision to a person with the data and options already gathered | Anything that changes contract terms, customer commitments, payment disputes or safety |
The third tier matters as much as the first. An agent that escalates well saves time even when it decides nothing. A planner who receives a case with the evidence attached, the likely causes ranked and two or three options laid out can decide in a minute what used to take twenty.
The boundary should move
These tiers aren’t fixed. A new deployment should start cautious, with most decisions in the propose-and-wait tier, and widen the agent’s authority as it earns trust.
The best evidence for when to move the line is how often people override the agent. If planners approve a type of proposal unchanged almost every time, that decision is a candidate for the act-and-report tier. If they keep editing it, the agent is missing something, and the edits show what. We’ve written about how planner overrides feed back into AI planning to improve it week by week. The same data tells you when an agent is ready for more responsibility.
The line can move back, too. A new customer, a new region or a festive-season peak can make yesterday’s safe decision riskier today. Boundaries deserve the same regular review as any other operating policy, not a one-time setting at launch.
What an agent owes the people around it
Giving an agent authority doesn’t remove accountability. It changes what accountability requires.
Every action should be visible: what the agent saw, what it did and why. Every autonomous action should be reversible by a person, quickly and without raising a support ticket. And every decision should leave a record someone can audit later, whether that’s an operations head reviewing a bad day or a finance team checking why a payment went out.
This is the difference between an agent people trust and one they work around. A system that can’t explain itself loses a team’s confidence after its first bad call. One that shows its working and lets people correct it gets better over time, because every correction becomes data.
How we applied it at Libera
When we built AI agents into Libera’s Freight TMS, we gave each agent one kind of decision instead of building a single general-purpose agent. There are separate agents for capacity planning, bid evaluation, load planning, trip monitoring, proof-of-delivery verification and freight intelligence. Narrow scope makes boundaries easier to set and easier to check. You can say exactly what the trip-monitoring agent is allowed to do; in a way, you can’t for an agent that does everything.
Every AI-driven action in the platform keeps a human in the loop, but the loop looks different in each tier. For routine actions, it means visibility and the power to reverse. For bigger ones, it means approval before anything happens.
We also built and ran these systems on ElasticRun’s own distribution network before offering them to anyone else. That gave us something most teams don’t have at the start: real override data showing where the boundaries should sit.
Beyond freight
None of this is specific to logistics. A customer-service agent issuing refunds, a finance agent matching invoices or a security agent isolating a device faces the same four questions. Can it be undone? What does a mistake cost? How fast must it happen? How sure can the agent be?
Teams that answer those questions explicitly, decision by decision, will get more from agentic AI than teams that argue about autonomy in the abstract. The useful debate isn’t whether AI should decide. It’s which decisions, on what evidence, and who checks. Get that right, and the alert queue finally starts to shrink.
About the Author:
Sheetal Kumar Ajamera is Senior Principal Architect at Libera, where he leads the engineering behind the platform’s freight procurement, planning, execution, and invoicing modules. He has spent his career architecting large-scale supply chain and ERP systems, with a focus on turning fragmented logistics processes into connected, data-driven platforms. At Libera, his work centers on the AI agents that power real-time rate benchmarking, load optimization, and billing reconciliation for shippers across India.
Connect with Sheetal on LinkedIn
