The Shiny Object Trap (Problem-First AI)
A regular PM ships the right product; an AI PM solves the right problem.
- Difficulty
- Easy
- Time to result
- ~weeks to results
- Steps
- 4
- Confidence
- 93%
Marily Nika's core guardrail against building AI because AI is exciting. The sequence is always pain point first, high-level solution second, model implementation last. She reframes the AI PM's job: not shipping features, but identifying which real user problem is best answered by a smart, data-driven solution, then handing that problem to a research scientist to model.
Origin
Marily Nika's own framing, developed while leading AI/ML products at Google (Glass, computer vision, speech recognition) and Meta, and taught in her Maven AI Product Management course.
Core principles
- 01Don't do AI for the sake of doing AI.
- 02There must be a problem, an audience, a user and a pain point before any model exists.
- 03The AI PM manages the problem as much as the product.
- 04Any behaviour you have data behind can be improved with AI — but only if the improvement matters to a user.
How to run it
- 1
Name the pain point, not the technology
Start from a user problem that is painful enough to matter. If you cannot state the audience, the user and the pain point in a sentence, you have a shiny object, not a product.
Pro tip Test it by removing the word 'AI' from your pitch. If nothing meaningful is left, stop.
- 2
Sketch the very high-level solution
Describe conceptually what a smart solution would do — personalise, recommend, detect fraud, speed something up, make it more accurate — without specifying architecture or model type.
- 3
Audit the data you already have
Inventory the data sitting unused in your product, or data from an adjacent product you can leverage. Nika finds many PMs have no dashboards and collect no data at all — fixing that is itself the first step towards AI.
Pro tip Hire a data science intern and see what they surface from the data you already own.
Watch out No data and no adjacent data source means you are not ready to model anything.
- 4
Only now reach out to implement
With the problem and the data understood, bring in the research scientist or data scientist and figure out how to actually build it. The model is the last decision, not the first.
In the wild
Rather than launching an 'AI product', Nika tells PMs on any existing product to look for one place where data can make the experience smarter: security, personalisation, fraud detection, ethics in healthcare, speed, accuracy, or better shopping recommendations.
→ A concrete, scoped AI feature attached to a real user pain point instead of a speculative model project.
Common mistakes
Starting with 'let's train a model'
Nika saw teams commit to building 'a really cool model' first; these consistently became low-ROI investments that ran six months to a year before anyone knew whether they worked.
Assuming you have no AI opportunity because you have no data
Having no data or dashboards is not a dead end — it is the first step. Start collecting behavioural data now so a model becomes possible later.
Is it for you?
Best for
Product managers at data-rich companies being pushed by leadership or hype to 'add AI' to the roadmap.
Not ideal for
Pure research organisations where exploratory model work without an identified product problem is the explicit mandate.
From the transcript
“there is something called the shiny object trap and I'm always telling people hey don't do AI for the sake of doing AI make sure…”
“generally PM helps their team and their company build and ship the right product but the aipm helps their team company solve the right problem”
“basically anything where you can get data behind the behavior users can be improved with AI”
From the episode
AI and product management
Marily Nika (Meta, Google)