When GPT-4 launched, putting "AI-powered" on your product was enough. Investors leaned in. Users signed up out of curiosity. Competitors held emergency meetings.
That era is over.
Today, a mid-size SaaS company has access to the same Claude API, the same GPT-4o, the same Gemini as everyone else. Foundation models have become infrastructure — like electricity, like cloud hosting. Who uses it doesn't matter anymore. What you do with it does.
And that shift creates a more uncomfortable question for product managers: if AI is no longer a differentiator, what exactly are you building on?
I noticed something last month while working on a feature in DentalBulut. A user wanted automated SMS suggestions inside the appointment module. "Use AI to do it," they said. Technically, we could ship it in two days. But four other apps already had this. They all used similar models. They all produced similar outputs.
So where was the difference?
In our three years of patient-appointment-treatment data. Nobody else had that. And a recommendation model built on top of that data says something completely different from a generic SMS assistant.
That's where I stopped and thought: the product isn't the AI feature anymore. The product is the data that makes the feature mean something.
A few months ago, a16z published a note on AI apps that stuck with me. Most AI product failures in 2024 weren't about model quality. They were about three things:
Wrong workflow integration. Insufficient context data. Lack of user trust.
The models were good enough. Everything around them wasn't.
That tells me the PM's focus needs to shift away from model selection. "Which AI should we use?" is too shallow a question now. The real questions are:
What data do we have that nobody else does?
Why would a user trust this output over a generic one?
How deep are we embedded in their workflow — and how painful is it to leave?
When you start asking these, some things get clearer.
Data accumulation beats feature velocity. A competitor can copy a feature you built in six months within three weeks. They can't copy three years of clinical data.
Workflow depth creates stickiness. In DentalBulut, a dentist logs in, checks appointments, records treatment, invoices the patient, and fires off an SMS — all inside one system. Migrating away from that is a real, painful process. Creating that pain (in a good way) is the PM's job.
Trust is still the hardest thing to measure. Even a perfectly functioning AI feature fails if users don't trust its outputs. In healthcare, building that trust takes months. And it's not about the AI — it's the sum of every interaction the user has had with your product.
I've also noticed something this year. "Let's use AI for this" no longer functions as a reason. Almost every feature request comes with "can AI handle this?" attached to it. Not a problem in itself — but worth being careful about.
Because AI is a very convenient way to obscure the real problem.
"Let's do AI sentiment analysis on patient reviews" — but the actual problem is that your review collection flow is broken. "Let's generate treatment recommendations with AI" — but the actual problem is that doctors aren't entering data consistently. No matter how good the model, garbage in still means garbage out.
I see people who call themselves "AI PMs" miss this more often than I'd like.
Short version: competitive advantage is still possible. It just no longer comes from the model — it comes from what makes the model meaningful.
Which is what it's always been: data, integration, trust.
AI just surfaces those three faster now.