Dead Reckoning: How Enterprise AI Drifts Off Course Without Continuous Validation

Before GPS, before satellite navigation, before any of the instruments that modern sailors take for granted, there was a technique called dead reckoning. It worked like this: you took your last known position, you tracked your speed, you held your heading, and you calculated where you must be now. No external reference. No confirmation. Just the logic of motion applied forward from the last point you were certain about.

It was a remarkable achievement of practical mathematics. It was also wrong in ways that compounded quietly, for days, before anyone knew how lost they were.

The error wasn't in the method. The method was sound. The error was in the inputs. A current you didn't account for. A wind estimate that was slightly off. A heading that drifted by two degrees and stayed there. Individually, none of these would have mattered. Accumulated over distance, over time, they could put a ship hundreds of miles from where its navigator believed it to be. Confidently, systematically, irreversibly off course.

This is exactly what is happening inside enterprise AI systems right now. And most organisations haven't looked up from the chart long enough to notice.

The last known position problem

Every AI system that operates on data is, in some sense, navigating from a fixed point. The training data has a cutoff. The records it references were accurate as of a certain date. The classifications it relies on reflect the world as it was when someone last checked. The system doesn't experience the passage of time. It doesn't know that the customer it's making a recommendation about has churned, that the regulation it's applying has been updated, or that the internal policy governing data use was revised six months ago and the change never propagated downstream.

It knows what it knew. It moves forward from there. It is, in the most literal sense, dead reckoning.

The challenge isn't that AI systems do this. Given how they work, some version of this is unavoidable. The challenge is that most organisations have no mechanism for understanding how far the drift has gone. They know the system is running. They can see it producing outputs. What they can't easily see is the gap between the position the system thinks it's in and where it actually is.

In navigation, that gap has a name: accumulated error. In enterprise AI, it doesn't have a name yet. It should.

Why the drift is hard to see

Dead reckoning errors are insidious for a specific reason: they don't look like errors. The ship is moving. The instruments are working. The navigator is doing exactly what they're supposed to do. The output looks like navigation. It produces coordinates, headings, and estimated arrival times. Everything has the form of correctness. The problem only becomes visible when you hit something that wasn't supposed to be there, or fail to find something that was.

AI outputs have the same property. A recommendation engine running on stale data still produces recommendations. A classification system working from outdated policy still classifies. A risk model trained on last year's conditions still generates scores. The outputs are real. They are acted on. Decisions are made. And somewhere, quietly, the gap between assumed position and actual position keeps growing.

The organisations most exposed to this aren't the ones that ignored data quality. They're often the ones that got data quality right at a point in time and then assumed the job was done. Freshness isn't a one-time audit. It's a continuous condition. And most data governance frameworks were built for the former, not the latter.

Celestial navigation and the value of an external reference

What eventually replaced dead reckoning wasn't better dead reckoning. It was a different kind of input entirely: a reference point outside the system itself. Celestial navigation introduced a check that didn't depend on accumulated calculations. You looked at something real, something external, something that existed independent of your assumptions, and you corrected against it.

The equivalent in enterprise AI is harder to build but just as necessary. It means introducing checkpoints that don't rely on the system's own internal logic to validate the system's outputs. It means auditing not just whether the AI did what it was configured to do, but whether what it was configured to do still reflects reality. It means treating data freshness, policy currency, and model calibration not as implementation details but as ongoing navigational inputs.

This is less glamorous than deploying a new model. It doesn't feature in product launches. But every serious navigator knows that the most dangerous moment isn't the storm. It's the clear sky that lets you believe you know exactly where you are, when the drift has been running for weeks.

The confidence problem

There's one more thing worth naming directly. Dead reckoning doesn't produce uncertain outputs. It produces confident ones. The navigator doesn't say "we might be somewhere around here." They say "we are here." The precision is part of what makes the error so costly, because confidence discourages correction.

AI systems have the same property at scale. They don't hedge. They classify. They score. They recommend. And the organisations relying on those outputs often have no easy way to distinguish the confident correct answer from the confident wrong one. Both look the same until something breaks.

The antidote isn't less confidence. Operational systems need to make calls. The antidote is a governance posture that treats every confident output as a hypothesis to be periodically checked against the world, not a conclusion to be filed and forgotten.

The sea doesn't care how good your last known position was. It only cares where you actually are.

‍