Scaling Spotify: product strategy, AI, and team autonomy with Gustav Söderström

Executive overview

Spotify built its reputation on music recommendation, but recommendation is now giving way to a third era: generation. The shift demands rethinking products from the ground up, not layering new tech onto old interfaces.

Gustav Söderström — co-president, CPO, and CTO — traces how Spotify navigated each transition: from user curation, to algorithmic recommendation, to generative AI. The same pattern recurs: new technology requires new UX, new business models, and disciplined restraint about what to build.

The hardest product decisions are redesigns, not new features — because participation is not voluntary.

The three eras of the internet product

• Curation era: users organised content themselves (Facebook, early Spotify)
• Recommendation era: algorithms replaced manual curation; required full UX rethinks
• Generation era: now beginning; no one yet knows what the right interface looks like
• Treat generative ML as a categorically different thing from recommendation ML — not an extension of it
• Two types of generative applications: iterative improvements to existing features, and genuinely new products that couldn't have existed before

The AI DJ: Spotify's first true generative product

• The "zero intent" use case — when users don't know what they want — was long unsolved
• Radio succeeded at this despite being non-personalised and non-on-demand; the knob-switching mechanic was the insight
• AI DJ combines a digitised real voice with dynamically generated, personalised commentary
• Key design principle: do as little as possible and get out of the way — it's there to serve the music
• Fault-tolerant UI: design must match algorithm performance; if hit rate is 1-in-4, show 4 options simultaneously (see also: Midjourney's early 2x2 grid)

AI and the future of music generation

• Generated music will likely be treated as an instrument, not a replacement for artistry — as the DAW was for EDM
• Truly original output remains hard; models excel at sounding like what already exists
• The distinction between "AI music" and "real music" is already dissolving; the question is degree, not kind
• Parallels to the piracy era: technology shifts first benefit consumers, then require new business models before creators can participate
• Rights holders need a new model before generation becomes a net positive — as streaming eventually outgrew piracy

Squads, tribes, and why Spotify moved on

• The squad model was right for its era: small, autonomous, full-stack teams of ~7
• Problems at scale: increments of 7 create enormous overhead; teams produced "heat" — 100 squads running in 100 directions
• Current teams are ~14 people, fewer overhead roles, more traditional structure
• Autonomy now sits at VP level, not the leaf nodes — large enough to generate diverse thinking, senior enough to have pattern recognition
• Extremes: Amazon (leaf autonomy, fast but ships org complexity to users) vs. Apple (centralised, coherent UX but bottlenecks speed)
• Spotify's strategy — one app, multiple content types — requires a centralised recommendation and UX layer

The homepage redesign: what went wrong and what was learned

• Problem being solved: users trapped in taste bubbles; low-intent discovery requires a fundamentally different UI from high-intent recall
• Built sub-feeds (podcast feed, music feed, genre browse) — these worked as intended
• Error: placed discovery-optimised UI on the homepage, flipping the ratio from 90% recall / 10% discovery to the inverse
• Users responded by fleeing to search and library; discovery tools were co-opted for recall they weren't designed for
• Key learning: Spotify's homepage handled recall — tracking multiple in-progress sessions simultaneously — better than most competitors; that was worth preserving
• Separating valid signal from change-aversion: look at new-user cohorts vs. existing ones; watch traffic shift to adjacent surfaces
• Updated hypothesis: make sub-feeds visible and voluntary, keep recall intact

Taking big bets and staying scientific

• Two types of product work: new features (voluntary) vs. redesigns (not voluntary — even users who dislike it are affected)
• Redesigns generate two kinds of negative feedback: "you changed things I liked" and "you made it worse" — hard to separate without quant data
• Commit 100% until data says otherwise, then commit 100% to the new direction — strong opinions, loosely held
• A/B testing is essential but insufficient for large rewrites: you need everything in place before you can distinguish a false negative from a real one
• The internal cost of a big bet (team time, opportunity cost) is harder than the external criticism

Organisation and planning

• 10% planning rule: spend no more than 10% of your cycle on planning; for a 6-month cycle, ~2 weeks
• If planning takes more, either the planning is excessive or the execution period is too short
• Spotify works in 6-month increments — similar to Airbnb under Brian Chesky
• Autonomy placed at VP level means many senior people can think independently without bottlenecking through one person

Explaining as a leadership tool

• The only way to truly understand something is to explain it to someone else
• Leaders owe their teams an explanation, not just a decision — "because I'm more senior" is not acceptable
• If you can't explain a product instinct, you probably don't understand it — or it isn't there
• "It's 0% art, 0% magic, 100% science" — a provocation to force rigour, not a literal claim
• Walk-and-talk one-on-ones (distributed, over AirPods) proved more generative than expected during the pandemic

Books and mental models

• Seven Powers — Hamilton Hemmer: strategy framework; having one is better than having none
• The Complete Investor — Charlie Munger: apply three different models to any problem; if all point the same way, confidence increases substantially
• The Beginning of Infinity, The Fabric of Reality — David Deutsch
• The Case Against Reality — Donald Hoffman: evolution optimises for fitness, not truth
• Gödel's Proof: in any axiomatic system, true statements exist that can never be proven
• The Demon in the Machine — Paul Davies: information as entropy; information engines

Rituals and frameworks that stuck

• Think it → Build it → Ship it → Tweak it: four-phase product model coined informally, now used inside and outside Spotify
• Think it: cheap, reduce risk; Build it: spend, but only after de-risking; Ship it; Tweak it
• The model stuck because it was concrete, memorable, and matched how product cycles actually feel