Anne Griffin shipped ten agents in under two and a half months by treating each agent as assembled from smaller tools and skills. That let her team reuse pieces when an agent failed or when a new one needed similar context and tools, instead of starting from scratch.
Anne Griffin describes designing for the threshold where a skeptical user will trust the output enough to use it. In cyber risk, one misplaced data point could make users abandon AI and return to manual work for a week, so overbuilding before feedback is risky.
Anne Griffin found users hitting a chat feature five times a day to draft vendor emails, a task outside the original agent plan. Talking to those users confirmed the workaround, so the team added email drafting to the workflow instead of assuming only in-product steps mattered.
Anne Griffin built a cyber-risk agent that called a data science model to estimate vendor risk from similar companies. Generative AI was too probabilistic for a use case where accuracy mattered, so the agent combined that model with LLM output rather than relying on the LLM alone.
Anne Griffin onboarded a hotel owner and social media manager to Claude and saw they did not know the project feature existed. For people outside tech, even a tool that feels simple can create too much friction, so they never reach the value product leaders assume is obvious.
Anne Griffin says product leaders should experiment enough to understand how AI works, why it fails, and whether it will work for a customer. She compares AI to the internet: you can build internet products without explaining every technical layer, and companies differ on how granular leaders must be.
Anne Griffin recommends product people take a prototype to an engineer and ask what would be hard to productionize. She sees first-time AI builders assume they can call an API directly, missing how probabilistic outputs, feedback loops, evals, and model behavior make shipping much harder.
Anne Griffin says the right UI for the customer gets left behind, and building it is a bigger challenge than people give credit. She points to the wave of AI companies hiring designers because a good model does not automatically make a workflow usable for non-technical users.
Anne Griffin warns against adding AI just to say a product is AI, especially when the team has existing customers. She argues that unless you are a frontier lab, copying features without knowing which user and value you are targeting makes the product hard to market and risks leaving customers behind.
Anne Griffin repeats that whether a product uses AI or not, the work starts with who the customer is, what they are trying to do, and their level of understanding. In cyber risk, that meant turning a wall of vendor data into plain-language highlights and a ready-to-send email.
Polly Allen points out that OpenAI and Anthropic hire consultants to help non-tech businesses adopt AI because the products still require too much setup and context. Anne Griffin agrees this is a product problem temporarily solved by human hand-holding rather than a lasting adoption strategy.
Anne Griffin notes that once users write off an AI feature, winning them back is hard; a little note in the app probably will not do it. In high-distrust groups, a single misplaced data point can send them back to manual work for a week, so teams must build to that trust threshold.
