The Principle
Adopt the tool that compresses the boring half of the work, not the creative half.
Every AI tool pitches itself as a designer replacement. The good ones quietly do something different. They take the parts of the job that nobody wanted in the first place (writing tickets, summarizing research, generating boilerplate, drafting copy variants) and hand the time back to the team. That is the lens I use. If a tool tries to replace taste, I stop testing it. If it removes friction from the boring half, I keep it.
The other lens that has saved my team months of churn: a tool only stays in the stack if at least three designers are using it daily after the trial period ends. Anything below that bar gets cut. The cost of a tool is not the subscription. It is the cognitive load of remembering it exists, the integration debt, and the half-trained team that uses it badly because nobody owns it. Cut early, cut often.
Research and Discovery
Hey Marvin, Dscout, and Granola sit at the top of our research stack.
Research used to die in transcripts. Now it lives in synthesis. Hey Marvin transcribes and clusters interviews in minutes, surfacing patterns my team would have taken a week to write up by hand. Dscout has matured into a serious diary study and unmoderated testing platform that designers can run without a dedicated researcher. Granola sits in every internal meeting and turns hallway conversations into searchable artifacts. The combined effect is that insight cycle time drops from weeks to days, which changes the kind of decisions a team can make.
The unlock is not the transcription. It is what becomes possible when synthesis is cheap. A designer can run six customer conversations on a Monday, have clustered themes by Wednesday, and walk into Thursday's roadmap meeting with evidence. The team that used to do one round of research per quarter now does two per month. The decisions get sharper because they are anchored in user voice that is hours old, not weeks old.
- Hey Marvin for interview transcription and pattern clustering.
- Dscout for unmoderated tests and longitudinal diary studies.
- Granola for ambient meeting capture and shareable summaries.
- UserTesting for fast unmoderated tasks when you need quantitative signal in 24 hours.
Design and Prototyping
Figma for source of truth. Lovable, Cursor, and Claude for everything around it.
Figma is still our canvas, but it is no longer where ideas are born. My team starts in Lovable when we want to see a working prototype with real data behind it, often within an hour. Cursor and Claude live next to Figma for the in-between work, generating component variants, writing helper text, refactoring tokens, and turning rough flows into shippable code. Designers who learned to prompt well are operating at roughly twice the speed they were six months ago. Designers who did not are slowly being out-produced by their peers.
What changed in 2026 is that prototyping moved from a designer activity to a designer plus engineer activity that lives in code. The fidelity gap between a Figma prototype and a working app collapsed. The teams that adapted fastest stopped treating Figma as the final artifact. It became the place to think, while the prototype lived in code from the second week of the project. That shift alone has been worth more to my team than any single tool.
The new design seniority is not visual taste. It is judgment about which AI output to keep, which to throw out, and which to ship.
Writing and Decisions
Claude Cowork for management work. ChatGPT for everyday drafting.
The single highest leverage tool for me personally as a director has been Claude Cowork. I use it for weekly digests, leadership update drafts, performance review prep, and structured deep research. It does not replace judgment. It compresses the eight hours of reading, formatting, and chasing that used to sit in front of every leadership decision. ChatGPT stays in the stack for everyday drafting (emails, copy variants, quick comparisons) because the muscle memory is hard to break and the latency is unbeatable.
The pattern I encourage on my team is one heavy tool and one light tool per workflow. Heavy tool for the work that compounds (Cowork for management, Cursor for code, Hey Marvin for research). Light tool for the throwaway tasks (ChatGPT, Granola). Trying to make one tool do everything is how teams end up with the worst version of three different workflows.
Knowledge and Decisions
Obsidian as a second brain. ChatGPT and Claude as thinking partners.
Every designer on my team keeps an Obsidian vault. Notes from research, decisions from reviews, links to artifacts. It compounds. Pair it with Claude or ChatGPT and a leader can interrogate six months of context in seconds, which makes coaching conversations and quarterly reviews substantially sharper.
The tools matter less than the habit. If your team is not writing things down, no model in the world will save you. If they are, AI turns that written history into the most valuable asset on the team.
What We Dropped
The tools that did not survive the trial.
Naming what we dropped matters as much as naming what we kept. A leader who only talks about the wins creates a culture where everyone is afraid to admit a tool did not work. We post the kill list in the team channel every quarter. It makes it normal to try things and walk away when they do not earn their place.
- Standalone AI design generators that produce "final" UI. Output looked finished, but never integrated with our system or our brand voice.
- Heavy AI plugin suites inside Figma. The cognitive overhead and slow updates made them a net negative against lightweight Claude prompts in a side window.
- Auto-summarizers attached to Slack and email. They saved minutes and lost nuance. The team stopped trusting the summaries within a month.
- AI "copilots" for design reviews that promised to score work. Scoring design without context is worse than not scoring it at all.
How To Roll It Out
A 90 day plan for taking your team from curious to fluent.
Month one: pick one workflow per designer to AI-assist. Not a whole job, one workflow. Research synthesis, copy variants, ticket writing. Set a clear before and after on time saved. Month two: pair people up and run a 48 hour hackathon on a real product problem. The hackathon is where fluency actually lands. Month three: name a tool owner per category (research, prototyping, writing, knowledge) and let them publish a one page "how we use this" doc for the rest of the team. The team that has owners ships faster than the team that has tools.
Skip the all-hands training. People do not learn AI tools by being lectured at. They learn by trying to ship something with a peer who is six weeks ahead of them. Design that into the rollout from day one and you will not need a single training session.