Making AI a Product Differentiator, Not a Demo
We introduced proprietary AI capabilities that became the platform's core differentiator. Weekly active users more than quadrupled on the AI-powered workflow, over 90% of customers now save AI-generated outputs to their default dashboards, and the product strategy behind it contributed to EBITDA more than tripling year over year.
The setup
I joined as Head of Product with a CEO mandate to introduce AI capabilities "yesterday." The pressure was competitive optics; every vendor in the space was bolting on AI. That is precisely the trap. Technology-first AI produces features users don't trust and don't need, and in a data product, one hallucinated insight can poison trust permanently.
The company had a real asset most competitors lacked: proprietary data. So the question wasn't "how do we add AI?" It was "where does our data give us the right to win, and which customer problems does that intersect with?"
The approach: a five-step framework
Map the data moat
Inventory the proprietary data competitors can't replicate. That's where AI differentiates rather than decorates.
Find the friction
Deep discovery into workflow pain. The standout: users had data-rich visualizations but couldn't extract decisions from them.
Prioritize by business value and user pain
Using the same weighted scoring that governed the rest of the roadmap. AI ideas got no special treatment; they had to win prioritization like everything else.
Implement at the point of decision, constrained by data quality
AI is only as good as the data it sits on, so the capability was scoped tightly around data we could trust.
Measure outcomes, not vanity metrics
Did users save the output as a default view? How many reprompts before a useful result? Did the insight change a downstream decision?
What we built
V1 was deliberately simple: AI summarization and interpretation of the platform's most complex, predetermined visualizations. These were the graphs users already had but struggled to extract decisions from. The detail that made it powerful: the interpretations drew on proprietary datasets beyond the graph itself, connecting what users saw to context they would otherwise assemble by hand.
That simple version taught us where the real differentiation was. The magic wasn't summarizing a chart; it was combining our data assets into something a user could act on. That insight drove the expansion: natural-language querying across datasets and custom visualizations answering the user's actual question, saveable to their default dashboard and reused from then on. It's also why we measured saves rather than clicks. A saved output means the AI earned a permanent place in the user's workflow.
Key decisions and tradeoffs
Saying no to the demo. The fastest path to "we have AI" was a chatbot over everything. We declined, accepting slower time-to-announcement in exchange for a capability with a defensible data foundation. The tradeoff was real executive pressure in the gap.
Outcome metrics over adoption metrics. Measuring "did they save it as a default" sets a much higher bar than "did they click it," and early numbers look worse for it. It was the right bar. It's the difference between novelty and workflow.
Extending AI to self-service. Once trust was established, we extended the same capabilities to reduce support dependency: AI-powered data exploration and self-service audience building. This freed 4 to 6 hours per week per account manager and shifted the team to higher-value retention work.
Results
- Weekly active users on the AI workflow more than quadrupled
- Over 90% of customers save custom AI-generated outputs to their default dashboard, the strongest trust signal we could design for
- EBITDA more than tripled year over year, with the AI and self-service strategy as primary product drivers
- AI became the platform's stated differentiator in competitive deals
What I'd do differently
Our very first release of the AI assistant was premature, and I knew it at the time. We didn't yet have the systems to properly test AI outputs and translate what we learned into constraints. There was strong executive pressure to get it out the door, along with a belief that a "beta" label would make a lower quality bar acceptable. I wasn't comfortable, and I deferred anyway.
Users don't grade AI on a curve because you labeled it beta. We spent the following cycles constraining and correcting, at real cost to new development work. The framework above, especially the fourth step, was earned by that mistake rather than designed in advance. When trust is the product, and with AI it is, "beta" is not a shield. I should have pushed back, and now I do.
What I'd tell another product leader
Every AI feature request is a data-quality question in disguise. The framework is portable, but the sequencing matters most: moat first, then friction, then prioritization, then constrained implementation, then outcome measurement. Skip the first step and you build what everyone else builds.
← All case studies