Stop starting with AI

A one-word strategy

A client came to me with a one-word strategy: "AI."

No details. Just AI.

They were a health and wellness company. Their customers loved the product, but growth had stalled, and they wanted AI to bring in more customers. They weren't unusual. On the Product Insider podcast, Dr. Nancy Li and I talked about how often companies treat "AI" as a plan in itself. (I shared three takeaways from that conversation separately.) It gets used as a verb, a noun, a strategy and a budget line, all at once.

The problem is that "add AI" is a solution looking for a problem. And when you start with the solution, you build whatever the solution is good at, not whatever the business needs.

Finding the real problem

So we didn't start with AI. We started with a three-hour deep dive into their mission, their sales funnel and how the business actually ran day to day.

That's where the real problem showed up.

They didn't have an acquisition problem. They had a churn problem. Customers were leaving, and when we looked at why, it came down to one thing: they didn't feel heard when something went wrong.

That changes everything about what to build. If you want more customers, you build marketing. If you're losing the ones you have, every new customer is filling a leaky bucket. It was more valuable to stop customers leaving than to find new ones.

And the first fix didn't need AI at all. If you want to find your own version of this, start with the three questions to ask before any AI build.

The first fix was deterministic

We built an intake and ticketing flow, so every customer problem got logged in one place. Every customer got an automatic email telling them someone was looking at their issue.

That's it. No model, no prompts. A deterministic process: the same input produces the same result every time.

It sounds too simple to matter. But a customer who hears "we're on it" within minutes feels very differently from one who hears nothing for three days. The process worked before AI touched it, and that's exactly why AI could improve it later.

Then AI, with rules from day one

Once the process was running well, we had something AI never has on its own: real examples of what "working well" looked like.

That's when AI came in. Its job was triage, sorting every ticket into high, medium and low severity so the urgent ones never waited behind the routine ones.

The part nobody had thought about was governance. We set the rules before the system scaled, not after:

• Keep customers' personal information out of the model.
• Never let the system give medical advice. For a health and wellness company, that one isn't optional.
• Make every AI suggestion clearly a suggestion, with legal language to match.

Over six months, churn dropped by about 60%. The full case study walks through each stage.

The order I use now

Here's the sequence I follow before putting AI on any problem:

1. Name the real problem. It's rarely the one you walked in with. (Here's a two-minute test for it.)
2. Build a process that works without AI. Manual, deterministic, boring.
3. Run it until you've collected examples of it working well. These become your quality bar.
4. Teach AI to copy that quality, using those examples.
5. Add governance before scale: personal data, legal limits, human review.

Most teams start at step 4.

That's why so many AI projects demo well and then fail in real use. The demo shows what the model can do. Real use shows whether anyone defined what it should do.

Why this matters more for a PMgineer

I call myself a PMgineer: one person doing the product manager, AI engineer, software engineer, DevOps and forward-deployed engineer work, directing a fleet of AI agents. That's how I built Varsity Page, a school spirit app that puts every program in the spotlight, in nine months. I explain why we build our own products first in its own post.

When agents do the building, this order matters even more. AI is a capable but naive assistant. It builds exactly what you describe and misses everything you didn't. If step 1 is wrong, the agents build the wrong thing quickly and well.

The same goes for governance. Varsity Page handles information about students, so the privacy rules aren't a feature request. They're written down before any agent touches the code, and every agent building the product has to follow them.

Governance is a product decision. Make it before something forces you to.

Start with the problem

AI is a remarkable tool. It's the e-bike of thought: every push takes you further. But it doesn't know where you're going. You still have to pick the destination.

So the next time someone in a meeting says "we need AI," ask the question that comes first: what problem are we actually trying to solve?

If you're not sure what that problem is in your business, or whether AI belongs anywhere near it, that's exactly what my free consultation is for. In 30 minutes we'll find your real bottleneck and what to fix first, with or without AI. Book a free 30-minute call with me.

This article expands on my conversation with Dr. Nancy Li on the Product Insider podcast, "End-to-End Roadmap to Become a PMgineer."