Spotify, Adobe and LinkedIn Product Leaders on Designing for AI

Harry Stebbings convenes a product roundtable with Adobe's Scott Belsky, Spotify's Gustav Söderström and LinkedIn's Tomer Cohen. Belsky describes his move from overseeing Creative Cloud products to strategy and emerging products; Söderström is Spotify co-president, and Cohen leads product at LinkedIn.

Belsky expects AI to help teams develop, test and debug code and explore more interface alternatives. Instead of spending long cycles building each variation, designers could generate options and choose candidates for testing. His prediction is that teams could find better solutions in fewer cycles.

Söderström offers a broader shift: the interface used to be thought of as the product, with machine learning assisting it; increasingly, he says, AI is the product and the interface helps it gather useful signals. He points to TikTok's simple video interface as an example of design amplifying a recommendation system.

Belsky says an interface need not disappear just because AI is involved. It helps people understand what they can do, and it may take new forms such as tone, conversational inflection or persona. Spotify's AI DJ, for example, makes choices about how the product addresses listeners.

Cohen says AI strategy cannot simply be delegated to a specialist team; product leaders and executives must understand it. He compares the transition with moving from desktop to mobile, which required teams to unlearn old assumptions and make choices about relevance rather than fitting everything onto a smaller screen.

An AI-first product can produce probabilistic rather than fully deterministic experiences, Cohen says. Leaders accustomed to controlling every output must learn to set objectives and constraints while accepting variation. Stebbings asks whether unpredictable outputs and hallucinations are especially worrying for public companies.

Belsky distinguishes contexts: a surprising output may be useful in music discovery or Photoshop's generative fill, where a user can try again, but is much more consequential in a legal document. The acceptability of errors depends on the job and the stakes.

Söderström says designing Spotify's DJ involved deciding whether it should feel like a utility, a generic assistant or a digitized person with a real personality. Designers need to understand model capabilities as well as users. He cites Midjourney's presentation of several lower-resolution options as an interface adapted to a slow, imperfect generation process.

Stebbings asks how product people can understand multiple models at once. Cohen expects platforms to route tasks to appropriate models based on the objective, cost and efficiency, hiding some of that complexity from designers and developers. He sees room for startups in this layer.

Belsky predicts a long tail of models, including specialized, open and locally running systems, rather than a few cloud models doing everything. He expects routing based on both suitability and cost. Cohen also imagines coordination at the application level among specialized agents, alongside routing at the model level.

Söderström says cost already constrains product design at Spotify's scale: two minutes of generated voice for half a billion users would mean a billion minutes. Much of the company's internal work therefore focuses on delivering capabilities economically, not only on whether they are technically possible.

He also entertains a model that embeds a user's full history—interactions and content—as a common representation, while specialized models handle tasks such as voice and video. Belsky notes that a routing layer might instead use the user context to choose among models. Both are describing possible architectures, not a settled blueprint.

Asked whether AI costs will follow a Moore's-law-like curve, Söderström expects continued hardware and software efficiency gains, while noting that both larger frontier models and smaller models matching earlier capabilities are progressing. Cohen points to the potential of optimizing software and chips together for power, capacity and cost.

Cohen says model size and data volume cannot be ranked without knowing the task. A large undertrained model can perform worse than a smaller, well-trained one; a specialist job-seeker coach may be more appropriate than an all-purpose life coach. The relevant question is what problem the team is solving.

For Belsky, implementation includes substantial final-mile work that depends on knowing the customer and product. He warns against becoming so absorbed in technical breakthroughs that teams forget the experience users receive. With rapid change, he argues, empathy for customers helps leaders choose which new capabilities to pursue.

Söderström says Spotify's bottleneck has moved over time, from finding technical talent to preparing data and teaching the company to design around what models can actually do. Spotify maintained an internal priority called 'AI is the product' to build that shared understanding and counter both early underestimation and later overexcitement.

Cohen says product reviews now include examining prompts, much as his team once insisted on mobile designs. Prompting takes skill, but it also lets designers and product managers generate early ideas themselves instead of first persuading engineering to build a prototype. Belsky describes a design-led team working on Adobe Firefly's early interfaces.

Returning to models and data, Söderström expects model size to remain important for a time, but thinks strong understanding of individual users will matter over the longer term. He cites Medprompt as an example where a general model combined with careful prompting performed well on medical tasks; useful context still had to be supplied.

Spotify builds some models and works with partners on others, Söderström says. It is not trying to build a hidden GPT-4 competitor: Spotify's goals, such as producing audio economically for many listeners, differ from a general AI lab's. He says Spotify works substantially with Google Cloud.

Belsky says Adobe should build models when its expertise and data give it an advantage, as with imaging, and partner for general-purpose language models. Cohen makes a similar distinction: LinkedIn's opportunity is integrating models with its own understanding of professional relationships and user needs, not competing to build every underlying technology.

Cohen calls data collection and quality a central product responsibility, not something to hand off entirely to data scientists. In his view, the first job is to define the objective of the algorithm; the next is to make sure the system receives the relevant, high-quality data needed to pursue it.

Belsky says Adobe pays particular attention to the provenance and commercial safety of training data. He states that Adobe does not train its models on customer creations unless customers submit work to Adobe Stock for that purpose and are compensated. He presents this as the company's policy at the time of the conversation.

Belsky also sees potential for companies to use their own customer data to personalize digital experiences. Söderström says Spotify learned the value of usage data through playlists and early representations of user behavior. He adds that licensed data can solve a business and intellectual-property problem as well as a technical one.

Söderström thinks the first wave of AI is impressive but has largely reinforced existing business models. He expects more disruptive change when a new technology enables a business model incumbents find difficult to follow, while acknowledging he does not yet know what that model will be.

Belsky argues that each platform shift is different. Moving from on-premises software to cloud or from desktop to mobile posed major transitions, whereas Adobe can add an AI interface within an existing product such as Photoshop and also offer Firefly separately. He believes established customer relationships can favor incumbents in this particular shift.

Söderström notes that Adobe's subscription transition made the new capabilities less disruptive to its own model. He compares this with Spotify's streaming model, which he says competitors took years to match. He allows that AI might simply make existing players more productive instead of overturning them.

Cohen raises a possible tension in seat-based software pricing: if AI lets an organization do the same work with fewer people, charging by user can undermine the provider's own revenue model. The panel considers paying for completed work or value rather than seats, as an emerging possibility rather than an established replacement.

Belsky extends the question to hourly professional fees. A lawyer's judgment may remain valuable even if AI shortens the work from hours to minutes; the question is how to price that value. He expects software purchasing to evolve too, but does not claim a specific timetable.

On enterprise adoption, Belsky says change can happen in jumps when a company sees a peer achieve useful results. He favors working with small teams as design partners before trying to sell broadly, and describes experimenting with customer-specific Firefly models using a brand's own intellectual property.

Söderström agrees adoption is early but expects companies to experiment with AI in parallel with existing systems more easily than they could move on-premises infrastructure to the cloud. He predicts a sharper adoption curve once use becomes widespread; this is his forecast at the time.

Cohen tells early-career product people that the tasks within jobs will change, making continuous learning essential. He cites LinkedIn's estimates of changing skill requirements and recommends combining broad skills with a specific area of expertise, plus interpersonal abilities and imagination. His numerical forecast is an estimate, not a guaranteed outcome.

Belsky advises people to experiment with new tools in their own work, then run small pilots with their teams to discover real utility. Junior staff can introduce useful practices even when colleagues have more experience. Söderström adds that newcomers can benefit from fresh knowledge and that learning prompting does not require asking permission.

Belsky draws a leadership lesson from running: rather than stopping immediately to record an idea, he keeps thinking about it, often improving it in the process. At work, he applies this by waiting a little longer before giving feedback or deciding how to solve a product problem.

Cohen invokes the conservation of complexity: product teams can absorb complexity in design or leave users to deal with it. LinkedIn serves people with many different goals, and he hopes AI can simplify the experience by adapting to the user's need instead of adding more visible features.

Belsky says he has changed his mind repeatedly about centralizing design. At one point distributing designers helped business teams resource their work; later he brought design together to serve a different strategy. A leader should be willing to change the playbook when circumstances change.

Cohen says he would not hire a product leader unwilling or unable to learn AI deeply. They need not be a model engineer, but should understand the objective, probabilistic behavior, responsible data collection and how to learn quickly. Söderström closes by saying technological disruption is an exciting time for designers and product leaders, even though its eventual outcome is uncertain.