AMA: Head of Product, Atlassian (Trello), Victor Dronov on Creating AI-enriched experiences

It comes down to a question: Is AI solving an existing user friction point, or is it a solution looking for a problem? Feature theater you mention usually looks like a generic chatbot slapped onto a sidebar because leadership panicked about an AI strategy. It creates high cognitive load for low utility. To ensure genuine enrichment, I filter ideas through three criteria: High Frequency, Low Joy: Does it automate a tedious, repetitive task? Many people use Trello's Slack and MS Teams apps - to ca ...Read More

We are moving from "permission-based" design to "reversion-based" design. Instead of forcing users to approve an AI action upfront - which just adds the friction right back - let the AI execute or draft the action, but make it incredibly easy to review, tweak, or undo. The best products in the space are using these principles: "Draft" pattern: AI creates the output (a roadmap, a response, a ticket), but the user owns the final publication. The AI acts as the intern; the user is the editor-in-chi ...Read More

Great question. As PMs, we are increasingly managing variable unit economics, not just shipping features. It is not unreasonable to anticipate significantly reduced subsidizing by the LLM providers once the market matures (remember cheap Uber rides, anyone?). Here is how product teams can protect margins at scale: Model cascading: Never use a sledgehammer to crack a nut. High-volume, low-complexity tasks (like text formatting) can be routed to smaller, highly optimized, or open-source models. Fr ...Read More

Trust isn’t broken when AI makes a mistake; it’s broken when AI acts completely confident while being dead wrong. Design language of predictable imperfection is what can help here. Here is what I love in the emerging AI experiences, which are designed for those inevitable misses: Tone down the certainty: Shift the UI copy from definitive to suggestive. Use terms like "Suggested tags" or "Drafted summary" rather than framing outputs as absolute facts. Visual hedging: Use distinct visual wrappers ...Read More

When foundation models become a commodity, the model itself is just an API call away for your competitors. The defensible moat shifts entirely to context, experience, and user trust. The PM craft now dominates in three specific areas: Building proprietary context: It’s not about the LLM; it’s about the unique data graph you feed it. Defensibility lies in capturing user interactions, metadata, and feedback loops that competitors simply cannot access. UX as the moat: Models are smart, but AI agent ...Read More