In our experience, logistics automation that disappoints rarely does so because the software was weak. It disappoints because too many decisions were handed over at once. When results slipped, nobody could say which decision was responsible or whether any of the others could still be trusted. The usual response is to pull everything back to manual, which leaves the team further behind than where it started.
The teams that avoid this work differently. They pick a single decision, let the system handle it while people watch closely, and move to the next only once the first has proved itself. At first that reads as timid. Twelve months in, they are usually ahead of the teams that automated everything at once, because they never had to unwind a rollout.
Automation is a set of choices, not a switch
When people talk about automating logistics, they usually picture one dial running from manual to autonomous. In practice there are many dials, one for each kind of decision.
Discussion about the automation of the logistics function usually concerns a single dial that is turned from manual to autonomous. However, in reality there are many dials that need to be turned for different decisions.
There is a wealth of human-factor research that describes complex tasks in terms of the following four stages: information-gathering, analysis, decision selection, and decision-execution. The level of automation that is cost-effective for each of these stages can vary.
To map this to the example of logistics, the system could monitor all the trips of vehicles. It could then detect if a vehicle is going to be late for its time slot and come up with three alternatives for it. The system could then decide on the best of these three for the vehicle. Or it could even book this alternative for the vehicle. The same problem but with four levels of trust that the team would need. They might be able to handle the first two decisions on every lane but not the fourth on the lane to the largest customer of the team.
When viewed this way, the single question “Should we automate?” is replaced by a list of much more manageable questions.
The risk with automation is scale more than speed. Software that makes a decision makes it the same way a thousand times, which is the whole point of it. It also means a flawed decision spreads as quickly as a good one. (Changed from here to the end of this paragraph) Analyst forecasts reflect the risk. Gartner expects over 40 percent of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. That’s a cross-industry projection, not a logistics one, but the last phrase is worth noticing. In our reading, inadequate risk controls is what happens when a system gets a lot of room before anyone can easily check or reverse what it does. The practical question is how much any single action is allowed to touch before a person can step in.
There are four additional benefits to making one decision at a time.
Your data problems will surface early as automation exposes the lack of vehicle data, incorrect rate cards, and suboptimal delivery times. Finding one such error while processing one delivery instruction can cost a week to resolve, whereas finding similar errors for a quarter of all your delivery instructions could have cost much more time to surface.
Your rules get written down. Unless you have already documented all of the planners’ unwritten rules, such as that this carrier is unreliable on hill routes in the monsoon and that the customer never accepts a split delivery, then automating one decision will force those people closest to the decision to state the rule that they are following.
Evidence builds trust quickly. After six weeks of good decisions, someone whose work enables that system has a good reason to stop checking on it. Someone told to trust it from the start has a good reason to quietly verify everything as it proceeds to save no time at all.
Decisions are reversed and kept small in error until you have a process around reversing error for that one decision.
What gets lost when everything is automated at once
Another reason for hand dealing decisions is the people involved in making them. When I talk about the lessons of aviation for use on land, I am not talking about decisions involving life and death. The very experienced pilots who fly planes for a living have long since seen the negative impacts of automation on their skills as hand flying pilots. Writing in The Atlantic several years ago, Nicholas Carr described how even routine hand flying has eroded in recent years as more and more aircraft are equipped with automation to fly the plane. The pilots are left to handle the unusual and most difficult situations with less than optimal skills for doing so because they are not regularly practicing their hand flying skills on routine flying.
The same is true in logistics. We have all of this software that automatically handles delays. As a result, planners for shippers and for freight forwarders (brokers) never see ‘most’ of the delays that occur with carriers. And thus they have never had to depend on their ‘gut feeling’ to keep a bad morning contained. They would guess which carrier’s people would actually answer their phone call. And now that a very unusual and very expensive case has occurred, the individual responsible for handling it has been out of practice.
Handing off one decision at a time to be handled by software still does not remove this problem, but it does reveal it in time for us to take action. We can have our planners review a sample of the automated decisions, have different people handle the exceptions on an irregular basis, and strive to keep their judgment sharp.
Choosing what goes first
A decision is a good first candidate when it has most of these traits:
In freight, that points to decisions like re-tendering a load a carrier has just rejected, within a rate card both sides have already agreed, or moving a dock appointment inside a window the customer allows. Assigning vehicles on routine, repeat routes is another, and it’s the kind of planning work Libera’s capacity and route planning engine handle. Awarding routine lanes on a scorecard is a fourth. Libera’s freight transport management system scores incoming bids on cost and service level together, drawing on execution history captured as trips run, which is the kind of rule-driven, checkable choice that suits an early handover.
What should stay with people longer: anything expensive or public, anything with compliance or safety consequences, new lanes with no history, and customer relationships where the right answer depends on context a system can’t see. These aren’t permanent lists. They’re the places where you’ll want a longer track record first.
A ladder, not a switch
The rollouts that work tend to move each decision up a short ladder.
Shadow. The system makes its choice and records it, but the planner decides as usual. Nobody acts on the software’s output. You’re only comparing what it would have done with what people did.
Suggest. The recommendation appears in front of the planner, who accepts or changes it. Now you can measure how often people agree and learn why they don’t.
Act with a veto. The system acts, and a planner can reverse the decision within a set window. This is the first rung where time is actually saved, and the reversal window keeps it safe.
Act and report. The system acts, and people review a sample and the exceptions afterward.
Few decisions ever need to reach the top rung, and that’s fine. What matters is that moving up is a deliberate choice backed by evidence and that moving down is just as acceptable. If a festival season, a monsoon, or a new carrier changes the conditions the rules were written for, stepping a decision back to “suggest” for a few weeks is routine maintenance, not failure.
What it looks like in practice
Let's look at an illustrative example. A shipper decided to re-tender all loads rejected by carriers on 10 routes. The choices of the system in shadow mode agreed with those of the planners most of the time. But the cases where they did not agree were the most valuable. Several of the disputes were on the same carrier on the same hill routes during the rains. The planners had bad experiences with that carrier on those routes during the rains, and one of the planners even explained the reasons in detail for including this rule. The rule was never part of the system, and so it got included. One month later the agreement rate had even increased, and the planners who had suggested the rule even started to care about the performance of the system, as their knowledge had been incorporated into the system.
That last effect is easy to undervalue. Each override carries information: either the system missed something, or the planner knows something it can’t see yet. A system that records overrides and learns from them gets better with use, and the people who fed it stay involved. We wrote about how that loop works in the feedback loop that makes AI planning better every week.
When to move up a rung
Three measures are enough to start. The first is agreement rate: how often the planner would have made the same call. The second is the cost of the disagreements, because ten harmless ones matter less than one expensive one. The third is how long it takes to catch and reverse a bad decision, which tells you how much room you can safely give. A decision earns the next rung when agreement is high, the disagreements are cheap, and reversal is quick.
None of this depends on how impressive the technology is. The pace is set by whichever of three things is slowest: the quality of the data, how clearly the rules are stated, and how much the people involved trust the output. A more capable model doesn’t change that.
Where Libera fits
Start with one of the decisions above. Run it through in shadow mode and let the resulting evidence dictate the pace. Decisions made at the award stage can be checked against the resulting TMS tracking information and settled invoice data. We can work out which decision is best to start with for your specific network. Our team is happy to help work out which decision should go first for your network.
The pace is the strategy
The pressure to automate everything quickly is real, and it usually comes from a good place. A logistics network is a series of promises, some thousands of times a day. The teams behind such networks expand their automation at the rate of their evidence. They may start slow but go faster every quarter as they create an operation that the people around it watch to the point where they stop checking the system because it has earned their trust.