The AI answer you can’t trace is the answer you can’t use

A crude tanker slows off a chokepoint and its AIS transponder, the automatic signal ships broadcast to identify themselves and their position at sea, goes dark for eleven hours. To a generic AI model, that’s a gap in a data stream. To anyone with money or compliance exposure on the line, it’s a question: whose ship, carrying what, under whose sanction’s regime and does the silence mean anything? The distance between those two readings has almost nothing to do with how advanced the AI model is. It has everything to do with whether the underlying data can be connected, and whether the answer that comes back can be trusted enough to act on.

I work in that world. I lead products for the AI capabilities customers use at Kpler, a maritime and commodity trade intelligence company. In plain terms, we track the movement of the world’s ships and the cargo they carry, and turn it into a picture of global trade that commodity traders, banks, compliance teams and governments rely on to make decisions. Most of that tracking begins with AIS, a system in which vessels continuously broadcast their identity and location. On its own, an AIS ping is just a dot crossing the ocean. All of the value comes from what you can reliably attach to it.

Those dots matter for reasons a technology leader in any sector will recognize as high stakes. Ship and cargo movements are leading indicators of commodity supply and demand, the kind of signal that moves energy prices. They expose geopolitical risk, from congested chokepoints to the growing “dark fleet” of tankers that switch off their transponders to disguise sanctioned oil. And they carry hard legal consequences: a bank or a trader that unknowingly finances a sanctioned vessel or cargo can face severe penalties, so knowing precisely which ship is which is not a nicety, it’s a compliance obligation. That is the moment the dark transponder stops being a data gap and becomes a question somebody has to answer.

So, I’ll say something that may sound odd coming from someone who ships AI features for a living: we never felt pressure to build “an AI product.” That was never the goal. We sat on one of the richest datasets in global trade, and most of the people paying for it could only reach a fraction of what it held. Nobody reads pages and pages of documentation. Plenty of users didn’t even know we could already answer the exact question keeping them up at night. AI, for us, was never a strategy box to tick. It was finally a good enough interface to close the gap between what the data could do and what people actually got out of it.

That reframing mattered, because it changed what we optimized for. We weren’t chasing a demo that looked intelligent. We were trying to make a genuinely hard dataset usable and, above all, trustworthy. And that pointed straight at two unglamorous problems most AI conversations skip past: whether your data can actually be connected, and whether every answer it produces can be traced back to where it came from.

Connection is harder than integration

Enterprise technology teams tend to talk about interoperability as if it were plumbing: wire system A to system B, pass the payload, done. But two systems can exchange data flawlessly and still mislead you. If a vessel is identified one way in your positional data and another way in your ownership data, joining them produces a confident, well-formatted, wrong answer. The real problem isn’t the pipe. It’s identity. Does “this vessel” mean the same entity everywhere it appears?

Reconciling that, which we call unification internally, is cumbersome work, and not for technical reasons. It’s cumbersome because it forces many different parts of a business to agree on a single definition of truth, and getting a commercial team, a data team and a compliance team to sign up to one canonical answer is a negotiation as much as an engineering task. We have a whole team dedicated to exactly that. We’ve done it for vessels, and that one win is instructive. Once a ship resolves to a single identity everywhere it appears, everything we know about it snaps together, and an AI sitting on top can reason about it without tripping over contradictions. This is the part of the work that never makes a keynote, and it’s the part that decides whether anything above it can be believed. Considering that, by some estimates, as much as 90% of operationally critical maritime data still arrives as unstructured text such as broker emails and port notices, the reconciliation problem only gets harder.

Traceability is the part I’d defend hardest

Once your data genuinely connects, you can let AI roam across it, and you immediately hit the trust wall. A user can ask, in plain language, “which sanctioned vessels discharged crude at this port last quarter,” and get a fluent paragraph back. But in a real workflow, a fluent paragraph is worthless unless the person can answer the next question: how do you know?

So, we made a rule that sounds obvious and is surprisingly rare in practice: no answer is delivered without its sources. Every entity in an AI response can be traced back to the exact signals that produced it, the position track, the cargo estimate and its confidence level, the ownership chain, the version of the sanctions list applied that day. We treat “show your work” as a first-class feature, not a footnote.

It does more than satisfy an auditor. It structurally addresses the hallucination problem, because an answer you can trace is an answer you can disprove, and one you can disprove is one you can finally rely on. This is the whole argument in a sentence: the answer you can’t trace is the answer you can’t use. In a regulated decision, where a wrong call can mean a sanctions breach rather than an awkward moment, an ungrounded output isn’t a smaller version of a good answer. It’s not an answer at all. The industry’s move from reactive to predictive operations only raises the stakes, because a prediction you can’t interrogate is a prediction no serious operator will bet on.

More trustworthy data means more bridges

Here’s the part that changed how I think about a roadmap. When your data is both reliable and connected, adding to it stops being additive and starts being multiplicative. Every new trustworthy, interoperable dataset you bring in isn’t just one more source. It’s a set of new bridges you can build between insights that used to live apart. Connect vessel movements to cargo, and you can see supply. Add ownership, and you can see risk. Add port and compliance data, and you can see intent.

Each reliable dataset you fold in doesn’t add one feature. It opens a combinatorial number of new questions the system can answer, because it can now be crossed with everything already there. That’s where intelligence actually comes from, not from a cleverer model but from more trustworthy things it’s allowed to connect. It’s also why so much of the value in AI-driven trade decision-making accrues to whoever has done the connecting work first.

This discipline cuts the other way too. A dataset that isn’t reliable, or that can’t be resolved cleanly to your model, doesn’t just fail to help. It poisons the bridges around it, quietly corrupting answers that used to be sound. So, the bar for what you let in has to be high, and holding that bar is one of the least glamorous and most important calls a product person makes.

What this means if you’re not in shipping

None of this is specific to trade. If you’re a technology leader being pushed to deploy AI this year, the sequence that actually works is the same in any domain. Start from a real user problem, not from the word “AI.” The best AI features are usually just old value finally made reachable. Make your data connect at the level of identity, not just format, and treat that reconciliation as an organizational agreement, not only a technical one. Make traceability a hard gate: if an answer can’t cite its sources, it doesn’t enter a decision. And judge every new dataset by how many trustworthy bridges it lets you build, not how many rows it adds.

The tanker is still off the coast, transponder dark. The organizations that will know what that silence means aren’t the ones with the flashiest model. They’re the ones whose data connects, whose answers can be traced to their sources and who kept adding reliable, interoperable pieces until the bridges between them started producing intelligence no single dataset ever could.