I Told 200 PMs They Were Using AI Wrong. Nobody Argued.

My Women Who Code Summit keynote on AI in product management, and what separates sharp PMs from replaceable ones.

Earlier this month I had the honor of opening the Women Who Code Summit in NYC with a keynote called “Faith, Trust, and Pixie Dust: Integrating AI Effectively and Responsibly as a Product Manager.” The premise: Tinker Bell may only need faith, trust, and pixie dust, but building products at Disney requires rigor, depth, and deliberate design. AI can help PMs get there faster and think more clearly, but only if you use it with intention.

The energy in the room was incredible and the response was overwhelming, so I wanted to share the core of what I talked about for anyone who wasn’t in the room.

Right Now, There Are Two Types of PMs

The first group is using AI to make them sharper.

The second group is using AI, and it’s quietly making them replaceable.

You might not know which one you are yet, but by the end of this article, you will.

Most people think the biggest value of AI in product management is speed. Compress hours into minutes, produce more with less effort. That are legit use cases. But they aren’t the whole picture.

AI is making the gap between average PMs and excellent PMs wider, and more visible. When an average PM uses AI as a crutch instead of as an amplifier, the output looks polished on the surface. It checks the boxes. It sounds right. But it lacks the depth and rigor that separates strong product thinking from dressed-up filler.

Plausible outputs are the most expensive outputs you’ll ever generate, because they’re easy for no one to catch.

Beyond personal productivity, it’s also a leadership problem. At scale, plausible but wrong things will ship. Your team will ship what your bar allows.

The AI Maturity Model for PMs

I presented a four-stage maturity model to help PMs self-locate, and, more importantly, figure out what moving up looks like.

Stage 1: Autocomplete. This is table stakes. You’re asking AI to polish a paragraph, summarize a doc, or make something sound better. Low value capture. This is what we all learned in 2023.

Stage 2: Draft Generator. You’re asking AI to write a first version of a one-pager, draft an email, or produce a leadership update. Medium value. The speed improvement is real, but your voice and your point of view can get eroded if your inputs aren’t specific enough to inform what comes out.

Stage 3: Thinking Partner. This is where PMs should be, at minimum. You’re using AI to challenge your assumptions and consider edge cases you might not have thought of on your own. The goal is to sharpen your judgment, not replace it. This sounds like: “Argue against my roadmap. Challenge this timeline. What’s the strongest alternative I might have rejected? What assumptions am I making?”

Stage 4: Judgment Amplifier. This is the goal. This is when you’re building systems, not just writing prompts. You’re using AI as a competitive intelligence loop and creating synthesis pipelines that do real work alongside you. The inputs are completely amplified because it’s not about single prompts anymore. It’s about the systems you build to make sure your AI outputs are exceptional.

When I asked the room where they were, most people said Stage 2. In my experience, most people who think they’re at Stage 2 are actually at Stage 1+. The honest self-assessment is the first step.

The MAGIC Framework

This is the framework I use when I’m working with AI day to day. It’s what I teach my teams, because we know everyone is going to use AI. The question isn’t whether they do. It’s whether they use it in a way that amplifies their work instead of creating a crutch.

M: Mission. What outcome are you driving toward? Is it a decision, a draft, alternatives, a critique? Be specific about the task you want your tool to complete, and frame your ask clearly from the very start. If you don’t define the mission, the AI will define it for you. You won’t like its version as much as your own.

A: Audience. Who is this for? Who is going to read it? The way you present something to your executives is different from how you’d present it to your engineering team or your legal partners. Tell the AI who the audience is, what they expect, and what level of rigor they demand. You’d be surprised how much this one addition changes the output.

G: Guardrails. What do you not want the AI to do? What lines should it not cross? Don’t invent metrics. Don’t speak for users. Don’t make projections unless they’re grounded in data, and if you do, show your work. Guardrails are the fastest quality lift in prompting. Just telling it what not to do automatically makes your outputs sharper.

I: Inputs. This is where you share the real context: customer survey data, meeting notes, emails, design comps, legal guidance, experiment results. Ground your AI in real truth so it produces outputs based on actual information, not hypotheticals. The more specific your inputs, the less the AI has to guess. Guessing is where hallucinations live.

C: Criteria. How do you want to evaluate the output? This comes in two parts. First, how should the AI evaluate its own output before presenting it to you? Second, once you get the output, how will you evaluate it against your own standard? This creates a double-check system where you set the bar the AI needs to meet, and then you hold it to your own bar on top of that.

Crafting the MAGIC: Workshop

During the workshop portion of the talk, we ran a live exercise using this scenario: an app called Little Chef Cooking Companion, built for people who don’t think of themselves as cooks, inspired by Ratatouille. The feature is simple. You snap a photo of what’s in your fridge or pantry, and Little Chef suggests things to cook.

Your task is to draft a one-pager as the PM for this product using AI.

This is where we can see the magic prompting framework in action. Let’s try two different ways to complete this task.

Round 1: The Wish Prompt. Open your AI tool of choice. Write a quick prompt for the scenario. One line, fast. Run it and evaluate the output.

Round 2: The MAGIC Prompt. Open a new chat. Rewrite the prompt using all five elements of MAGIC. End the prompt by telling your tool to ask you clarifying questions about your task. Run it. Compare the outputs from the first chat and the second.

Try both and compare the outputs. The difference isn't subtle. The gap between them is the gap between Stage 2 and Stage 4.

It’s staggering. One attendee shared that with the wish prompt, the output was instant and skeletal. With the MAGIC prompt, her tool thought for many minutes and asked her clarifying questions before producing a complete PRD with suggested North Star metrics, a prioritization framework, and a slew of open questions and considerations to challenge her thinking.

The scenario was intentionally simple. But think about what that gap looks like when you’re not drafting a one-pager for a workshop. Think about what it looks like when you’re building the actual product.

Three Habits That Separate Sharp PMs from Replaceable Ones

The MAGIC framework is the system. These are the habits that make it stick.

1. Prompt for disagreement, not validation. Confirmation bias is baked into LLMs. You could tell it you’re going to skip rent this month to buy a Birkin and it would call it an investment in yourself. Instead of asking AI to confirm your thinking, tell it to argue against you. Generate the strongest cases for opposition. Ask it what assumptions you’re making. This is how AI sharpens your judgment instead of softening it.

2. Ground in real artifacts, not hypotheticals. Instead of telling AI to “imagine a user who…,” give it real user data and customer feedback. If you’re asking it to help you understand your users, give it a clear criteria to work from so it’s not pulling random information out of thin air. If you don’t have actual data, ask it to help you research.

3. Rewrite before you ship. AI drafts often sound like AI. No matter how many parameters and skills you layer in, the output has quirks that are not how you write. Use the AI output as a draft, then rewrite in your own voice. The goal is not to replace you. It’s to help you recognize patterns you might have missed on your own. Then you bring the voice and the judgment.

Set the Bar. Don’t Gatekeep

I have an intern joining my team this summer, and this is her first PM internship. I was going back and forth about whether to give her access to AI tools.

Part of me was like: when I started as a PM, I had to write 20-page documents by hand. Uphill. In the snow. By myself. She should have to go through that too.

But after thinking about it more, I realized that this is where we are. This is the landscape. These tools exist, and they should be helping her. So instead of gatekeeping, I’m going to spend this summer teaching her how to use them effectively to boost her learning, without using them as a crutch to write her complete output for her.

That’s the job now. Not gatekeeping. Setting the bar.

The PMs you manage and the ones coming up behind you are going to use AI no matter what. Your job is to make sure they use it well.

---

Three things I’d challenge you to do this week:

Honestly self-locate on the AI maturity model and plan one move from your current stage to the next. Just one step. It doesn’t have to be a revolution. It can be as simple as applying the MAGIC framework to a single email you’re drafting this week.

Audit the last three AI-assisted outputs you or your team produced. Check them for ghost writing, over-validation, and outsourcing patterns. Name what you find. Figure out where better prompting could improve those outputs.

Identify a recurring PM task, write a MAGIC prompt for it, and turn it into a reusable skill. Stop writing the same prompt from scratch every time. Build systems.

---

Here’s to bigger and better.

Abundantly yours,

Alexandra 🤖