A follow-up on building an AI copilot for freight dispatch
TL;DR: We don't decide whether to automate a task based on how repetitive it looks. We look at four things: how much the right decision varies between real-world cases, the cost of being wrong, whether the task’s value comes from speed or judgment, and how quickly a wrong output would be detected.
Together, those determine whether our AI dispatch copilot should suggest, ask for confirmation, act autonomously, or stay out of the way.
Automation is not the default win condition
There's an assumption baked into a lot of AI product thinking right now: if a task can be automated, it should be. Manual is friction, automation is progress, ship it.
We don't work from that assumption anymore. Not because automation is bad - obviously most of what we build is automation - but because treating "can we automate this" as the only question skips the one that actually matters: should this specific task be automated, given what happens when it's done wrong, and how often "wrong" looks different each time.
That second part - how often "wrong" looks different - turned out to be the real hinge. A task can be low-stakes and still be a bad automation candidate if every instance of it is weird in its own way. And a task can look scary and still automate well if the failure mode is boring and consistent.
The four questions we actually ask
We stopped trying to have one big "automation philosophy" and instead run four concrete questions per feature. None of them require ML expertise. They require someone who understands the actual operational process and is willing to ask honest questions about it.
1. How much does the right decision vary from one instance to the next?
Extracting predefined fields from a rate confirmation can be relatively consistent once the prompt and extraction logic have been properly refined. In practice, reaching that level of consistency may require several rounds of prompt improvement and validation. Even then, the task is not entirely straightforward: the total may combine linehaul, fuel surcharge, accessorials, different currencies, or handwritten adjustments.
High-variance tasks are different. Negotiating detention time with a broker who's annoyed about a missed appointment involves interpreting tone, the history with that broker, and how much leverage the carrier actually has that week. Two instances of what appears to be “the same task” can therefore require completely different judgment calls.
Low decision variance → a potentially good automation candidate, provided the inputs are reliable, the output can be validated, and failures are detectable.
High decision variance often means the task cannot be made reliably automatable through prompt engineering alone. It may require richer context, structured rules, specialized tools, historical data, or human review to handle different scenarios consistently.
2. What does a wrong output cost, and does the cost scale with volume?
This sounds like the reversibility question from our earlier post, “Where Should an AI Agent’s Autonomy End?”, but it's slightly different — it's about whether errors compound. Auto-flagging a possible detention clause mismatch that turns out to be a false positive costs a dispatcher thirty seconds of annoyance. Auto-sending a rate counter-offer that's off by a meaningful margin, at volume, across dozens of loads a week, doesn't just cost thirty seconds each time - it can quietly damage a broker relationship you didn't even know was at risk until the relationship is already gone.
Even small errors create operational costs, and those costs accumulate as volume grows. Some errors scale roughly in proportion to the number of occurrences, while others create disproportionate or delayed damage by eroding trust, reputation, or a counterparty’s willingness to work with you again.
3. Is the value of doing it mostly about speed, or mostly about judgment?
Some manual tasks are slow because nobody's built the tooling yet - that's a straightforward automation opportunity. Others are slow because they require judgment, and the slowness isn't the bug; it's the mechanism by which a human catches something a faster process would have missed.
Confusing these two is probably the single most common mistake in early-stage automation decisions: “this takes a dispatcher ten minutes” gets read as “this is inefficient” instead of “this is where the actual thinking happens.”
4. How quickly and reliably will we detect a wrong output?
A reversible error is only safe if someone notices it in time. Some failures become obvious immediately — through a validation check, a dispatcher catching the issue, or a product metric moving in the wrong direction. Others may surface only days or weeks later, through a damaged broker relationship, a lost customer, or a pattern of operational friction that is difficult to trace back to a specific automated decision.
This matters beyond the product itself. An incorrect output can affect how a dispatcher communicates with a broker and, if it goes unnoticed, influence the relationship between them.
The longer and less reliable the feedback loop, the less autonomy we should give the system.
This is also how we think about agent autonomy: it's not a property we give an agent once. It's something we earn task by task.
Autonomy should be earned per task, not granted to the agent as a whole.
That's why we don't think about an AI copilot as simply being “autonomous” or “not autonomous.” Different tasks can - and should - have different levels of autonomy.
We start with decision variance and failure cost, then adjust the level of autonomy based on how much judgment the task requires and how quickly and reliably failures can be detected.
That's the approach we're taking with LoadConnect, our AI dispatch copilot for freight teams.
A framework, roughly
We use decision variance and cost-if-wrong as the first pass — effectively a 2×2 for determining the initial level of autonomy:
Low variance + Low cost if wrong → Automate freely
Low variance + High cost if wrong → Automate + confirmation
High variance + Low cost if wrong → Draft / suggestion + human finalizes
High variance + High cost if wrong → Keep manual
The matrix is only the first pass. Whether the task’s value comes primarily from speed or judgment acts as a separate filter: if judgment is central to the task, we lower the level of autonomy by one step.
Detection speed is another. Even a low-cost, reversible failure deserves less autonomy if we are unlikely to notice it quickly and reliably.
In practice, these autonomy levels map to different system designs. “Automate freely” still means validation and logging. “Automate + confirmation” means generating a structured action but requiring explicit dispatcher approval. “Draft / suggestion” means the model can prepare the response but cannot execute it. “Keep manual” means AI may retrieve context or summarize information, but the actual decision remains outside the agent.
The more autonomy we give a task, the more explicit its validation, logging, and failure signals need to become.
The goal isn't maximum automation. It's the right level of automation for the task.
The high-variance, high-cost category is smaller than people expect. But it is also where the “automate everything” instinct can do the most damage. These tasks often look repetitive and well-defined from a distance, but become much less predictable once you examine real-world cases and the consequences of getting them wrong.
An example that goes the "wrong" way on purpose
Most of what I write about tends to be examples where we correctly chose not to automate something. It's worth including one where we got it wrong first.
Early on, we built a feature that automatically sent a standard response when a load’s pickup window changed slightly — a templated “here’s the updated appointment, please confirm” type of message. On paper, it looked like a low-variance task: a time changed, so send an update.
In practice, the reasons behind those pickup window changes turned out to be far more varied than the template assumed. Sometimes the change was routine. Other times, it signaled a larger issue with the shipper that the dispatcher needed to address directly rather than paper over with an automated message.
We had measured variance at the surface of the task — a timestamp changed — instead of at the level of the decision underneath it, where the real variation was.
Surface-level variance ≠ decision-level variance.
We walked the automation back to a suggestion the dispatcher reviews before sending, rather than an auto-send.
The lesson wasn't "don't automate messaging." It was: **measure variance at the level that actually matters, not the level that's easiest to measure. **A timestamp is easy to check for variance. The judgment call underneath it isn't, and that's usually the one that counts.
What this costs us
Being disciplined about this has a real cost, and it's worth naming instead of glossing over: it means shipping fewer flashy "the AI just does it for you" features than we technically could. Some AI demos look more impressive because they auto-execute actions that we deliberately keep behind a confirmation step.
We've made peace with that trade. A demo that auto-books a load looks great in a thirty-second clip. It looks a lot worse the first time it books the wrong one, and in freight, that first time is a phone call and a canceled contract, not a bug report.
The checklist, if you want to steal it
Before we automate a task fully - not just assist with it, actually remove the human step - we want honest answers to:
- If we sampled a representative set of real instances — including rare and high-impact edge cases — would they require essentially the same decision, or are meaningfully different situations hiding under the same label? Are we measuring variance at the level of the actual decision, rather than just the surface of the task?
- Does a wrong output stay a small, contained mistake, or can its financial, operational, or reputational cost compound as volume grows?
- Is the current “slowness” of this task mostly wasted motion, or is it where a human is actually doing the hard part?
- If the system gets this wrong, how quickly and reliably will we know — and who is likely to notice first, us or the customer?
That last one matters more than it sounds. Some bad automation decisions surface immediately, in your own metrics. Others surface three weeks later, in a broker's inbox, and by then it's not really a product decision anymore - it's damage control.
Closing thought
"What should we automate" gets asked constantly in AI products right now. "What should we deliberately not automate, and why" gets asked a lot less, and it's usually the more useful question - because it's the one that forces you to actually understand the task instead of just measuring how often it repeats.
What's one task your team has deliberately kept human-in-the-loop - and what made you decide not to automate it?
