As AI tools multiply across freight and logistics, the important question is how much control to hand over, and to what. On the latest Lean Quarterly Dive, host Malcolm Harris sat down with Lean Solutions Group CTO Alfonso Quijano to unpack that question. “experts in the loop,” a model that treats human judgment
Quijano didn’t waste time getting to the thesis. “Experts in the loop,” according to Quijano, is a model that treats human judgment not as a stopgap for AI’s limitations, but as the thing that makes AI worth using in the first place.
“Experts in the loop is the natural evolution of the logistics roles inside of every operation for every broker and every logistics player,” Quijano said. AI has unlocked real productivity gains across the industry, he said, but its full value only shows up once it’s paired with human judgment.
“You don’t really leverage its full capabilities until you pair it up with the right judgment and the right capabilities that come from a human expert,” Quijano explained, adding that Lean isn’t interested in building a world where AI makes decisions unsupervised. Instead, the model splits the work. AI handles the repetitive grunt work, while humans stay in the loop feeding it judgment so it keeps improving and can be trusted with the harder calls.
Standard operating procedures are still in a process of development as AI becomes integrated into more and more workflows. According to Quijano, most companies have never made SOPs usable by AI in a way that tightens the feedback loop. When an AI system hits a case the SOP doesn’t cover well, an expert needs to confirm there’s an operational gap before the document gets updated. Right now, most SOPs are static, reviewed once and then forgotten.
Quijano drew a comparison here that’s likely to resonate with anyone who’s spent time inside modern AI tooling: the CLAUDE.md or AGENTS.md files that let AI agents share context and coordinate with each other.
“What Anthropic and OpenAI and all these guys have figured out is that for the agents to work together and understand everything, they need to have a file where they write all the things that they’re doing in order to improve and get better,” Quijano said, noting that the concept is functionally identical to an SOP. It acts as a living source of truth for how a process works and how it evolves. Quijano believes that SOPs are headed in that direction industry-wide, and that they will be updated far more frequently by both AI and the people who oversee it in a feedback loop that will reshape how operational knowledge gets maintained.
The market is still determining where AI earns its keep today and where it falls short. Quijano’s list of clear use cases include document-heavy workflows such as invoice intake, discrepancy checks against bills of lading, routing emails to the right agent, and the kind of accounts-payable and accounts-receivable work.
These are tasks where fatigue and inattention cause real human error, and there’s a tradeoff.
“The AI is not going to miss,” Quijano said. “At least the accuracy for the AI is much higher in those cases.”
Quijano described the typical handoff point as built around dollar thresholds and client-specific rules. An invoice might suddenly include a batch of unfamiliar accessorial fees and get flagged for a human to verify, and that person’s decision then trains the system on whether to treat similar cases as exceptions going forward or adjust the threshold entirely.
Carrier sales tell a different story. “I’m bullish on the potential for the technology,” Quijano said, but he’s more cautious about full autonomy, largely because of what gets lost when a human isn’t involved in the early stages of a client relationship.
Trust is something AI still can’t deliver on its own. That’s generated through the human parts of sales and first contact that are harder to automate. He drew a distinction between inbound carrier sales, where a lot of spot-market inquiry volume is low-value noise AI can reasonably triage, and outbound sales, where the calculus is murkier and companies are still figuring out whether to go all-in or ease into automation gradually.
“Decision intelligence,” as Quijano described it, is Lean’s attempt to combine 14 years of operational knowledge with an AI layer that behaves like an embedded assistant, whether that’s in the format of a sidebar, desktop app, or always-on presence that watches for ambiguous data or risky decisions in real time and flags them before they become problems.
New hires at Lean, Quijano said, would typically route around half their tasks to a manager on day one. With decision intelligence pulling from SOPs, operational history and peer behavior to suggest next steps automatically, that onboarding drag shrinks dramatically.
Across document-based automations specifically, Quijano cited figures in the range of 65 to 70 hours of time given back to operators per month, or roughly a quarter of their working hours. But that gain isn’t pure profit. A share of that reclaimed time gets reinvested into training and giving feedback to the AI systems themselves, which is a necessary cost of the model, according to Quijano.
Asked what advice he’d give a logistics executive weighing how to bring AI into their operation without losing control, security or sound judgment, Quijano said that step one is mapping the business down to atomic-level workflows to identify where exceptions live and where automation makes sense.
Lean can benchmark that distinction against data from across its own client base rather than leaving companies to guess in isolation. Step two is understanding where a given AI tool sits on the autonomy spectrum, from simple rules-based (“deterministic”) processes up through fully agentic systems that can retain memory, deploy tools independently and act on what they observe. Not every workflow, he stressed, needs to be agentic. Plenty of companies are forcing agentic complexity onto processes that would run just fine on simpler, more predictable logic.
The last piece, according to Quijano, is planning for the share of the workload that stays human. In Lean’s experience, AI typically handles 70% to 80% of a given process, leaving the remaining 20% to 30% as exception management that requires trained people, many of whom are now operating in a different role. Rather than running the core process, those experts become coaches and ambassadors for the AI itself, reinforcing good outcomes and supplying evidence when something goes wrong.
“That’s the next challenge,” Quijano said. “That shift matters just as much as the technology decisions.”
Lean Solution Group will demo at F3, and Quijano says that much of this topic will be on display.
In addition, Leans Solutions Group recently announced the release of new Everest Group research showing that enterprises are adopting artificial intelligence much faster than they are redesigning operations to realize AI’s full business value. The report, “Designing the Human + AI Enterprise: A Transformation Agenda for Supply Chain-Intensive Industries,” is being introduced in conjunction with the FreightWaves F3 event, October 26–28, in Chattanooga, TN. A webinar will be hosted with an Everest Group analyst on December 8 to discuss the research and noteworthy insights for the industry.
For more information on Lean’s platforms, including decision intelligence and its Agent Edge product, visit leangroup.com.
The post Why Logistics AI Still Needs “Experts in the Loop” appeared first on FreightWaves.
