You have a high priority project on your roadmap right now to add more AI features to your product. You are in good company. Across industries, teams are under the same pressure. Ship meaningful AI capabilities. Move fast. Show measurable impact. Do not fall behind.
The board is asking about it.
Customers are expecting it.
Competitors are announcing it.
The roadmap seems straightforward, but it’s not.
There’s no single right path from idea to value, you need the model to craft the experience, and you need the experience to shape the model.
Lenny Rachitsky
If you have actually built generative AI products, something quickly becomes clear.
Building AI products is not a funnel. It is a feedback loop.
1. From Sprint to Prototype: Start With the Question, Not the Answer
In traditional product design, a sprint yields a Figma prototype that looks finished. In AI, that prototype is only meaningful if you test it against a real model early.
For example, in Healthcare, we built what looked like a polished patient triage interface in Figma — but only once we connected it to a clinical LLM did we realize user workflow assumptions were wrong. The AI didn’t respond like we expected, and the experience collapsed. This matches the “counterintuitive” idea that:
First, understand what’s technologically possible — and then design the experience around it.
With generative AI, “build the solution” isn’t enough — you must explore tech limits first.
2. The Loop: Model Experience
This is where many teams stall. You might hear:
“We need the model first.”
“We need the UX defined before touching the model.”
Both are half‑truths.
In Fintech — where compliance, risk, and trust matter — we learned that we couldn’t finalize UX flows until the model’s risk profile was understood, and we couldn’t tune the model until users tried the UI. That tension drove us to a parallel path of engineering:
- Prototype UX assumptions,
- Evaluate model behavior against real sample data,
- Adjust UX based on where the model broke trust or legality,
- Iterate the model on where users hesitated.
This matches a core insight from Counterintuitive Advice :
Experiences inform the model, and the model constrains the experience — both matter early, not later.
It’s not sequential — it’s a loop.
3. The Self‑Driving Car Metaphor: You Need Both Car + Model to Iterate
A self‑driving car doesn’t emerge from training data alone — and it doesn’t emerge from simulated UX alone. You need:
- The vehicle (product context & interface)
- The driving model (AI core)
Only by putting them together can you measure:
“Can this interface let the model do its job?”
“Does the model enable safe, expected behavior in real interaction?”
In Logistics, we built routing tools that initially looked powerful, but when drivers used them, even strong models produced routes that ignored human patterns. Only when we observed real experience could we adjust model inputs, constraints, and outputs to fit the context.
This captures something deeply counterintuitive about AI product building:
Bringing model and experience into the same iterative cycle accelerates learning — and prevents wasted product effort.
4. The Real Value Isn’t the AI — It’s the Problem It Solves
Across all three industries, there was one consistent truth:
Cool AI doesn’t guarantee value — solving real, felt human problems does.
In Healthcare, we saw models spit out clinically plausible text — but it only mattered when the product reduced patient wait time or clinician burnout.
In Fintech, models that generated compliance summaries were impressive until they had to stand up to audit.
In Logistics, predictive arrival time forecasts were only useful once they cut customer support loads.
As builders featured in Lenny’s newsletter put it:
“Demo value isn’t user value.” — and that requires longitudinal user validation.
5. What This Means for Your Next AI Product
If you’re at the early stages of your AI journey — whether Martech, Healthcare, Fintech, Logistics, or beyond — here are the practical takeaways we’ve internalized:
Don’t push model work to the end.
Prototype with real models as early as possible.
Don’t freeze UX until the model is perfect.
Use design thinking to shape what you need the model to do.
Iterate the model and experience together.
Think in loops, not lanes.
Measure what matters.
User impact beats benchmark scores and impressive demos.
Above all, we cannot treat AI projects as just another feature. They demand a dual path mindset, where engineering and design are locked into an iterative feedback loop from the first week through launch.
After three years exploring this path, we are convinced: the quality of the product is determined not by which model we chose, but by how well we married that model to a meaningful human experience.
If you’re ready to move beyond AI as just another feature and redesign your product through a true AI-first design process, let’s build it together — Forwwward is ready to lead your GenAI transformation, today.
