Faizaan Charania treats generative AI as a tool for solving user problems, rather than a reason to add AI to a product.
2
Teams should start with a simple MVP, test their assumptions, then replace an expensive general model with a smaller model suited to the use case.
3
AI products need defined evaluation criteria and feedback systems because model errors can damage user trust and the product's reputation.
Summary
Faizaan Charania explains how his technical background shapes his work as an AI product manager at LinkedIn. He starts with the user problem, checks what data is available, and works with tech leads to turn a broad product idea into an implementation. For new machine learning products, he prefers a simple MVP that tests assumptions before the team invests in a more complex model. He applies the same thinking to generative AI. Teams can use hosted large language models to test demand quickly, then move to a smaller, more focused model if the product proves useful. Faizaan prefers AI experiences that meet users inside an existing workflow, such as helping write an email, rather than requiring users to open a separate assistant and construct a detailed prompt. He also discusses prompt iteration, output constraints, model temperature, evaluation criteria, feedback pipelines, bias, and trust. His advice is practical and candid about the costs and uncertainty of AI features.
A technical product manager starts broad and hands implementation to tech leads
Faizaan says a machine learning product manager needs to understand how a system works under the hood. He begins with the user problem, then asks what data the team has, what data it needs, and what experience the product should provide. His working style starts at a high level and adds detail over time. He and the tech leads define the solution together until the remaining decisions concern implementation. At that point, he hands the work to the engineers and stays available for product questions. He does not spend his time choosing which Kafka stream supplies the data. He focuses on the features the team can use and the experience they are trying to create.
A simple MVP gives the team evidence before it invests in a complex model
Faizaan argues that an AI product cannot begin with the best possible model because too many assumptions are built into a new idea. A team can spend a large amount of effort confirming or rejecting the wrong hypothesis. He therefore prefers a simple MVP that tests whether the team is seeing the right signal. Depending on the problem, that may mean a heuristic or a basic clustering approach using existing user attributes. If earlier work or strong data already proves that a more complex approach works, the team can increase the complexity sooner. The model choice follows evidence about the product, rather than enthusiasm for a more advanced technique.
Generative AI should be chosen to serve a user problem
Faizaan's main product principle is that generative AI is a tool, not the product idea itself. The team still needs to build the best experience it can for users. He suggests starting with relatively forgiving use cases, such as text creation, where an imperfect answer may be acceptable. As the product takes on factual tasks, the team has to address hallucinations and may add retrieval augmented generation. He does not think assistants are the right starting point for every company. A product should first ask which user problem has the highest value and whether an AI feature can solve it inside the existing product.
AI experiences should do more work for users instead of demanding a separate conversation
Faizaan contrasts pull experiences with push experiences. An assistant is a pull mechanism because the user has to start a conversation, supply context, and ask for value. He prefers products that do more of this work in the user's current workflow. He uses Google Gmail's Help me write feature as an example. The user can provide a simple description, while the system can use the email thread and contact context behind the scenes. The complexity is hidden from the user. Faizaan also points to Notion, where users can select text already in front of them and ask for it to be cleaned up or improved, without moving to another window and bringing the result back.
Hosted models are useful for testing demand before teams build specialized infrastructure
Faizaan places himself in the group that values go-to-market speed. Large language model APIs can make it fast to test whether an AI feature is useful, even though the same approach may become expensive at large scale. Most products have a narrow use case, so they do not need a general model that can answer questions about every subject. After a feature shows promise, the team can replace the hosted model with a smaller model trained or tuned for the specific task. Faizaan expects that focused model to perform better in many cases while costing less. Open source models can also help when a company needs to host the model itself or control the data and sources used by the system.
Prompt quality has to be tested across many users, not one impressive example
Faizaan warns that the first successful prompt can create false confidence. A response may look brilliant when a team tries one example, but the same experience can produce anomalies across 10,000 users. Prompt engineering and repeated prompt iteration are needed to remove bad cases while keeping the useful behavior. He recommends adding as much relevant context as possible and constraining the output. Temperature can help with consistency, although it has tradeoffs. An external evaluator can check one model response against another, but making two model calls also doubles the cost. His preference is to make the prompt as effective as possible before adding more machinery.
AI evaluation needs explicit criteria and a feedback loop
Faizaan says teams need to define what success means before they build the experience. They should identify the problem, confirm that users care about it, set a time period, decide what the MVP should do, and compare the metric they want to move with the cost of building the feature. Generative AI adds another layer because it creates a new interaction pattern. A team needs clear criteria for deciding whether an output is good or bad, especially when many people will score results differently. Online feedback can expose corner cases that offline testing misses. Teams should track failures, improve the system, and watch for user behaviors that differ from their original assumptions.
Trust becomes part of the product when AI speaks for a company
Faizaan treats trust as a major concern for generative AI experiences. Language models can sound confident while producing something that does not make sense. A feedback mechanism tells users that the team is willing to learn and fix problems. It also gives the team evidence about where the system fails. Faizaan says the AI output can become a voice for the brand because it adds a human element to the product. A foolish answer can therefore reflect badly on the company. Product and machine learning teams need to consider bias, unexpected interaction patterns, and corner cases as part of the experience, rather than treating them as issues to address only after launch.
"The first time you hit a prompt about a product that you want to build, you'll see the output and you'll be like, 'Oh my God, this is brilliant. I've hit jackpot.'"Faizaan Charania29:54
Who should watch
You are a product manager working with machine learning engineers and need a practical boundary between product decisions and implementation decisions.
Your team wants to test a generative AI feature quickly, but you need a way to think about model cost, prompt quality, and when to build a narrower model.
You are shipping AI outputs to users and need to define evaluation criteria, collect feedback, and protect user trust.
