You must have sat through more AI demos this year than in the previous ten combined. Most would have been impressive. Fewer would have been entirely true. The gap between the two is rarely dishonesty — usually it is just a good team that has not yet been asked the right questions. So here are five questions you may want to ask before you believe any AI claim in logistics, including ours.
1. YOU MAY ASK WHAT PERCENTAGE RUNS WITHOUT A HUMAN
Not "does it use AI" — everything claims that now, even a Diwali card generator. You could ask what share of decisions ship with zero human review, today, on live freight. A good answer sounds specific and a little unglamorous — "92% on document checks, 60% on exceptions" — not a round 100%. Our POD Audit Agent clears the vast majority of proof-of-delivery documents on its own and hands over only the ones with a genuine dispute or an ODA (Out of Delivery Area) exception. That handover is not a weakness. It is the design.
2. YOU MAY ASK SINCE WHEN, AND AT WHICH CUSTOMER
A roadmap slide and a production deployment can look identical in a pitch deck. They are rarely identical in a pilot. It helps to ask for a customer name, a go-live date, and a volume — trip count, fleet size, shipment count. Our Voice Agents have been calling drivers in Hindi across live fleets like DTDC and Patanjali for months now, confirming pickups and pacifying an upset consignee at 11pm — a job that never existed done well, now done reliably. That is a claim you can verify. "Powered by AI" on a website usually is not.
3. YOU MAY ASK WHAT HAPPENS WHEN IT'S WRONG
Every model is wrong sometimes — the useful question is what the system does next. Does it fail silently, does it escalate to a human with the right context, or does it quietly generate a bad e-way bill and let someone discover it at the checkpost? We rebuilt two of our own agents last year, not because the underlying model was weak, but because our first version escalated the wrong 20% and let through the wrong 5%. That is the real work of shipping AI in logistics — tuning where the human sits, not removing the human.
4. YOU MAY ASK FOR THE BORING METRIC, NOT THE HERO METRIC
"Reduced costs by 30%" is a hero metric — impressive, hard to verify, easy to cherry-pick. "Average handling time per POD fell from 4 minutes to 40 seconds" is a boring metric — specific, checkable, and often more convincing precisely because it sounds unglamorous. As a tech company, we have found it useful to lead with boring metrics ourselves — they tend to hold up six months later, hero metrics rarely do.
5. YOU MAY ASK HOW LONG THEY'VE BEEN IN YOUR OPERATIONS, NOT JUST IN AI
A strong model wrapped around a shallow understanding of Indian logistics will usually get the demo right and the edge cases wrong — the address with no proper landmark, the GST mismatch that stalls a shipment at Bhiwandi, the driver who trusts a phone call over an app notification. We are still on a journey to learn even after twenty-plus years in this business, and that patience is what tells us where a model needs a human far more than any accuracy benchmark does.
AT WEBXPRESS
These are the same five questions we ask ourselves before letting an agent anywhere near a customer's freight — and the honest answer gets published on our own AI page rather than only the flattering one. Not every one of our fourteen agents sits at the same stage of maturity, and that is something you deserve to hear from us before your pilot, not during it.
None of this is a case for slowing down on AI. It is a case for asking better questions on the way in — the vendors worth working with will welcome them, and the answers will tell you more in ten minutes than any demo will in an hour.
The best AI claim you hear this year will likely be the one with a number attached, a customer attached, and an honest "but" attached.