At Viacom I built the company's first social media data team. We had something genuinely new: every public reaction to every show, at a scale nobody in the building had seen before, running through models we ended up patenting.
The first thing everyone asked for was a dashboard. So we built a good one. Filters, cohorts, trend lines, a share button. It was opened a handful of times in the first month and then, mostly, it wasn't.
That's the most useful failure of my career, and it generalizes: nobody buys insight. People buy a decision they can defend, a workflow that got shorter, or a line item they can charge to somebody else. Insight is the raw material. The distance between the raw material and any of those three is where most data programs quietly die, and where a lot of AI programs are dying right now.
What people actually pay for
Three things, in my experience, and a data product that doesn't deliver at least one of them isn't a product yet.
- A decision someone can defend. Somebody has to walk into a room and say "we're doing X." Hand them the sentence and the backup for it.
- A workflow that got shorter. The thing that took four hours takes forty minutes. This is the most reliable of the three and the most consistently undervalued.
- A line item. Something a third party will pay for, packaged so that someone in sales can actually sell it.
Our dashboard offered none of these. What it offered was the opportunity to go find an insight, which is work, and which we had just handed back to the person we were supposed to be helping.
Why teams build the dashboard anyway
The reason isn't technical, and it isn't stupidity. A dashboard is the artifact that satisfies the largest number of stakeholders simultaneously, which makes it the easiest thing to get approved and the hardest thing to be blamed for. You ship the surface and let other people supply the judgment. If it doesn't get used, that's an adoption problem, and adoption problems belong to somebody else.
There's a structural version of this too. Data teams are usually funded as a service function, and service functions produce artifacts on request. Nobody in that arrangement is accountable for whether the artifact changed anything. If you want to know quickly whether a data team is building products or taking orders, ask whether they've ever told a senior stakeholder no.
What actually worked
The fix wasn't a better dashboard. It was taking the same models, with mostly the same output, and pushing that output into the places where decisions were already being made: into the commercial conversation as something a buyer would pay a premium for, into programming and marketing as a specific recommendation with a confidence attached, and out to the brand teams as something that arrived on a schedule rather than something they had to remember to go get.
Same data. Same models. Different packaging and different distribution. Across more than 500 brands it became over $50 million in new revenue, and the thing that changed was never the modeling.
I want to be careful about the lesson here, because it's easy to over-learn. The models mattered. Without something defensible underneath, better packaging just gets you to a credible-sounding wrong answer faster. The point is that the modeling was necessary and nowhere near sufficient, and we spent our first year acting as though the ratio ran the other way.
Two questions before you fund it
These are the ones I ask now, and they've saved me more money than any framework I've used.
Who is the named person whose job changes the day this ships? Not the department. Not the persona. The person, with a name, whose Tuesday afternoon is different.
What were they doing instead, and how long did it take? If nobody was doing anything, you haven't found a workflow, you've found a hypothesis. That's fine, but fund it like a hypothesis.
If you can't answer both, you don't have a product. You have a capability and a hope, and those get funded very differently.
This got easier to get wrong
AI made this failure mode cheaper to commit, which means it's about to get much more common. A model will produce a fluent, confident, well-formatted insight about nearly anything you point it at, in seconds, at almost no cost. The raw material is now effectively free and infinite. Which means packaging and distribution aren't part of the problem anymore. They're the whole problem.
A lot of what I see in enterprise AI right now is structurally the dashboard. A surface, a box to type into, an invitation to go find value. Then everyone is surprised when usage decays after week three. It's the same mistake with better typography.
The teams I've watched get this right have one thing in common. They can tell you exactly whose Tuesday afternoon gets better, and they can describe it without using the word insight.