How AI Actually Works, for Product Managers
You don''t need to build models to manage them — but you do need an accurate mental model of what AI is, because it behaves unlike any software you''ve shipped before. Get this right and every later module — strategy, metrics, ethics — makes sense. Get it wrong and you''ll promise things AI can''t deliver.
Traditional software vs machine learning
Traditional software follows rules a human wrote: if this, then that. The behaviour is fixed and, in principle, exactly specifiable.
Machine learning flips that. Instead of writing the rules, you show the machine many examples, and it learns the patterns itself, then applies them to new inputs it has never seen. You don''t tell it "an email with these words is spam"; you give it a million labelled emails and it learns what spam looks like.
That difference is the root of everything else: an ML system''s behaviour comes from data, not from a spec you author.
Training vs inference
Two phases every PM should distinguish:
- Training — the model learns from historical data. Done offline, it can be slow and expensive, and it produces the model.
- Inference — the trained model makes a prediction on a new, live input. This is what runs in your product, fast and cheap per call.
When someone says "the model," they usually mean the trained artefact being used at inference time.
The flavours you''ll meet
- Supervised learning — learns from labelled examples to predict (classify emails, forecast churn). The most common in products.
- Unsupervised learning — finds structure in unlabelled data (customer segments).
- Generative AI — produces new content (text, images); LLMs are the headline example.
The single most important fact: AI is probabilistic
Here is what makes AI products fundamentally different: the output is a prediction, not a certainty. A model returns an answer with a confidence, and it will be wrong some of the time. There is no "bug-free" version — there is an error rate you manage down, never to zero.
This one fact reshapes your whole job:
- No exact spec. You define a target metric and an acceptable error rate, not a precise input-to-output contract.
- Statistical QA. You test on data and measure accuracy; you can''t enumerate every case.
- UX for being wrong. The product must handle mistakes gracefully — show confidence, allow correction, route low-confidence cases to a human.
- It drifts. As the world changes, a model that was accurate degrades, so you monitor and retrain.
Garbage in, garbage out
A model can only be as good as the data it learned from. Biased, incomplete, or unrepresentative data produces a biased, unreliable model — no algorithm fixes bad data. As a PM, data quality is a product decision, not just an engineering detail.
Common pitfalls for PMs
- Expecting 100% accuracy. No model is perfect; plan for the error rate.
- Treating AI like deterministic software. The same input can behave differently; "it worked once" isn''t "it works".
- Believing the demo. A polished demo is the happy path; production data is messy and full of edge cases.
- Ignoring the failure UX. What the product does when the model is wrong matters as much as when it''s right.
Why this is your on-ramp
AI Product Foundations, Metrics & Analytics, and Ethics & Governance all assume you understand that AI learns from data and predicts probabilistically. Internalise "learned, not programmed" and "probabilistic, not certain," and you''ll manage AI products with realistic expectations from day one.