Reliable Enough to Trust at 2 AM: Shipping Agentic AI Into Freight Ops
Reliable Enough to Trust at 2 AM: Shipping Agentic AI Into Freight Ops
Good AI in logistics isn't about being clever. It's about being reliable enough that an ops team trusts it at 2 AM when something goes sideways. A container gets rolled, a cut-off is missed, a shipment sits on a customs hold, and whoever is awake needs to know what the system did, why, and what's left for them.
Exceptions are the product
I've built agentic exception handling for exactly those cases: rolled containers, missed cut-offs, customs holds. It's tempting to treat exceptions as the messy tail after the happy path is done. In freight they are the job. The happy path is a booking that moves without anyone touching it, and nobody staffs a team for that.
Which means the design question isn't "can the agent handle the normal case." It's "what does the agent do when it is uncertain, and how quickly does a human find out." An agent that resolves eighty percent of exceptions silently and fumbles the rest invisibly is worse than one that resolves sixty and escalates the rest loudly.
The eval framework came before the features
At Winmore I established an AI evaluation framework across cost, latency, accuracy, and hallucination guardrails, and we adopted it across modules. At Freightender there's a dedicated AI evaluation layer attached to every feature, not a phase at the end.
Those four dimensions are deliberately unglamorous. Cost, because an agentic loop that calls a model eleven times per booking has an economics problem hiding inside a demo. Latency, because a rate quote that takes ninety seconds loses the business regardless of quality. Accuracy, measured against real documents rather than a curated sample. And hallucination guardrails, because a confidently invented surcharge in a contract summary is not a bug that gets caught downstream, it's a bug that gets invoiced.
Automation earns scope, it doesn't start with it
Booking Automation reduced manual touchpoints and ops cost by around twenty percent on key accounts. That number came from narrowing, not widening. We picked the flows where the inputs were reliable and the failure was recoverable, proved them, and expanded from there.
The same applied to Contract AI, which cut review and onboarding turnaround by roughly forty percent for enterprise forwarders. It didn't start by reading contracts end to end and pronouncing judgement. It started by extracting the things a human would otherwise hunt for, showing its source, and letting the reviewer stay the decision-maker until the accuracy record earned it more room.
What I care about as a PM
Solving real user problems over shipping shiny features. Honest build-vs-buy calls. Roadmaps that survive contact with reality. In AI that translates to a specific discipline: instrument the thing, publish the numbers internally, and let the automation's scope be a function of its measured record rather than the roadmap's ambition.
Ops teams are not anti-AI. They're anti-surprise. Give them a system whose behaviour they can predict at 2 AM, and they'll hand it more work than you'd have dared to ask for.
Lokesh skipped presentations and built real AI products.
Lokesh Radhakrishna was part of the June 2026 cohort at Curious PM, alongside 20 other talented participants.