THE NOISE

Marketing Attribution Was Broken Before AI. AI Dashboards Just Make the Wrong Numbers Look More Confident

2026-09-20

Attribution has been a genuinely hard problem in marketing long before AI showed up, and anyone who's had to sit in a room where sales and marketing argue over who gets credit for a closed deal already knows this firsthand. Multi-touch journeys, last-click bias, and offline touches that never make it into any system at all are old problems, not new ones.

The newer pitch is an AI-powered attribution dashboard that finally solves it, with a model doing the work of untangling which channel actually drove the result. What the model can't do is invent data that was never captured in the first place. If the underlying tracking has gaps, a more sophisticated model built on top of those gaps just produces a more confident-looking number, not a more correct one.

This is worth illustrating with a specific, common scenario. A prospect sees a LinkedIn ad, forgets about it, gets referred by a colleague at a conference three months later, and finally converts after a sales call. Only the last two touches typically make it into any CRM. No AI model, however advanced, can recover the LinkedIn ad's contribution from data that was never recorded. It can only work with what was captured, and what gets captured is usually the easy-to-track digital touches, not the messier real-world ones that often matter more.

That's the actual risk worth naming: a dashboard that admits it's rough invites scrutiny. A dashboard that presents a single, clean, AI-generated number tends to shut scrutiny down, right at the moment a real budget decision is riding on it.

There's a specific organizational dynamic that makes this worse, one familiar from sitting on both the sales and marketing sides of budget conversations. Once a number comes from a dashboard labeled AI-powered, it tends to get treated as more objective than the same number would have been if a person had produced it, purely because of where it came from. That's backwards. A number's reliability depends on the quality of the underlying data, not on whether a model or a person did the arithmetic on top of it.

The useful check before trusting one of these tools with a real allocation decision: does it explain its own assumptions and methodology, or does it just output a number with no visibility into how it got there. If you can't ask it why and get a real answer, you're trusting a black box with money.

A practical version of that check: ask the tool, or the vendor, what percentage of your actual customer touches the model can see versus infer. A tool that gives you a specific, honest percentage is being straight with you about its own limits. A tool that dodges the question, or answers with something vague like comprehensive tracking across channels, is telling you it either doesn't know or doesn't want to say, and either answer should lower your confidence in the number it produces.

Our stance for now: use these tools for directional trends worth investigating, not for defining a budget outright, until you've stress-tested the model against something you already know to be true. A confident number is not the same thing as a correct one, and AI is very good at confident.

There's also a governance question worth raising internally before adopting one of these tools company-wide: who is accountable when the dashboard's number turns out to be wrong six months later, after budget has already shifted based on it. If the answer is nobody, because the AI made the call, that's itself a signal the tool has been given more authority than its underlying data quality actually supports.