Sharing some notes from last week's festival - as usual there was an eclectic mix - and perhaps a greater than usual sense of ambiguity around the inevitable AI hot-topics.
AI business change sessions
Redesign before technology as an answer to the "90-95% of AI projects fail" stat.
▶ The Reinvention Window, Marc Warner (CEO & Global CTO, Faculty & Accenture). His argument: most projects fail, 90-95% by one count, because we bolt AI onto the existing process rather than redesigning it (his analogy was factories taking years to get value out of electricity while keeping the steam-era layout).
Key point: The same pattern came up in nearly every business session I sat in, the value isn't really in the model, it's in whether you're willing to redesign how the work gets done, and to rethink who the customer is.
▶ The Workforce Reimagined, Euro Beinat (Global Head of AI & Data Science, Prosus). One of the larger agent deployments I've heard described: 60,000 agents built in 12 months, graded intern to senior, with a few hundred driving most of the return.
Key point: What resonated was "jagged intelligence", the idea that AI is brilliant and brittle at the same time, which I think we'll all recognise from our own interactions with it.
▶ Why AI-Native Companies Are Playing a Different Game, Brian O'Reilly (COO, Writer). Only 10-15% of large enterprises have really re-architected their workflows and 80-85% of use cases don't need a frontier model.
Key point: the underserved opportunity is AI supervision (governance, auditing, observability) as businesses in their own right, because the capability is running well ahead of the controls. That's going to be the critical element as we re-architect and re-imagine businesses.
▶ Intercom On the Age of AI, Des Traynor (Co-Founder, Intercom). The story of putting 80% of engineering behind a $100k product (Fin) the day ChatGPT launched, against a $250m business.
Key point: Intercom's story is well known by now, but a good reminder that adoption will move fastest outside western markets, because there it's zero-to-something rather than an incremental upgrade.
▶ The Agent Has Entered the Store, with Daniele Bernardi (Toolhouse), Bouchra Kaabouz (Air France-KLM), Andy Fishburn MBE (Virgin StartUp) and Alessandra Bosco. Travel is the obvious early proof, given the friction (80% of carts abandoned, 23 tabs, three-hour bookings).
Key point: you need to serve two audiences at once, the model (which rewards consistency) and the human (who still wants relevance, or elasticity). Lose either and you optimise yourself into a sea of sameness.
The memory economy
The AI you talk to will probably end up knowing you better than you know yourself, while quietly working for someone else. It was one of the most thought provoking themes from last week.
A run of sessions kept circling the same question: who does AI actually work for, and who controls what we hand it?
▶ Meet Charlie, Sir Tim Berners-Lee & John Bruce (Inrupt), with Tania Bryer OBE (CNBC). Their argument is that the memory economy is more dangerous than the attention economy: assistants that know us intimately but answer to shareholders. Their proposed fix is a data vault you own, with agents working on top of it.
Key point: The data-vault model is one of the more important ideas for an agentic world. Whether it gets traction commercially is a separate question.
▶ The 6-Pack of Care, Audrey Tang & Caroline Emmer De Albuquerque Green (University of Oxford). Alignment reframed as a question of care rather than just code, with Taiwan's deliberative polling that cut deepfakes by 94% and became law in two months.
Key point: the bit that resonated was designing and making decisions to avoid lock-in. Alignment as an ongoing process, not a setting you configure once.
▶ Designing for Trust, Hovhannes Avoyan Hovhannes Avoyan (Picsart) & Nad Chishtie (Lovable), moderated by Hannah Parvaz (Aperture). Trust has shifted from "can it impress us?" to "can we rely on it?". The detail I liked: role-based agents flopped until they were given names and faces.
Key point: non-determinism (and the variability that comes with it) is a feature rather than a bug, but it does change how you have to design for trust (e.g. from empowerment, rather than consistency).
▶ From Policy to Practice: Building National AI Capability, George Osborne (OpenAI), moderated by Arjun Kharpal (CNBC). If the personal version of this is "who holds my data", the national version is "who holds the country's". Osborne on how governments pragmatically turn AI ambition into capability: trusted infrastructure, sovereignty, sharing cyber-defensive models with trusted partners, and regulation flexible enough not to be out of date in nine months.
Key point: it's the same data-vault instinct, one level up. Sovereignty isn't all-or-nothing; the harder question is which layer of the stack you trust others to run, and which you keep.
Fear, Trust and Hope
The corridor conversations at SXSW and the undercurrent in many sessions was how fearful people are of AI (and this from an AI-first and AI-literate crowd)... and yet I came away hopeful. A few sessions reminded me what we're capable of when we point the tech at the right problems and design it the right way.
The strongest sessions on staying ahead of AI's harms shared a spine: this is really about whether we can trust it, with our children, with the public realm, in our careers, and with each other. But we also need to remember that trust takes years to build and moments to break and so today is a formative moment.
▶ AI Is Already Shaping Childhood, Julia Gillard, Kanishka Narayan MP (UK Minister for AI & Online Safety) & Giovanni Salum (Child Mind Institute). The argument: act faster than we did with social media, and treat child safety as a moral non-negotiable rather than a cost-benefit calculation. A warning note is that AI is fundamentally different from social because it goes beyond social interaction to core cognitive functions.
Key point: the cleanest test I heard all week was whether a given use enhances or degrades someone's agency. It works well beyond children too.
▶ Reclaiming AI for the Public Good, Bo Young Lee 이보영 (AI4ALL), David Ryan Polgar (All Tech Is Human) & Elizabeth M. R. (Oxford). The harms aren't theoretical, bias enters at four different layers (algorithm, data, neural network, interface), and collective action does work. Bias and values aren't an accident that creeps in at one point, they are choices made at every layer. LLMs outputs are the direct result of design (and prompt) choices, not accidental
Key points: "inevitability" is itself a choice of narrative and collective action works (e.g. ozone layer). Existing law (IP, environmental, labour) is probably a faster lever than waiting for new AI-specific rules.
...and despite a shadow of fear in the rooms last week, there was still a sense of SXSW wonder - a fair bit of optimism about what creativity and technology can do together.
▶ Who Gets to Touch the Stars, Sheila Xu (AstroAccess) & Sandhya Sabapathy (Kaleidoscope). A deaf crew's light-based comms system ended up helping everyone once the cabin was too loud to hear.
Key point: design for the person your product wasn't built for. What breaks for them tends to break for everyone else under stress too.
▶ LEGO Futures Imaginarium, Joana Lenkova & David Pallash (LEGO). Probably my most joyful session of the week, partly because it was a room full of grown adults actually playing with LEGO. Play activates 24+ skills across creative, social, cognitive, physical and emotional development, and LEGO's Build the Change programme reaches 3 million children a year.
Key point: children from disadvantaged backgrounds "dream smaller", imagination can help change that. Also reminder that imagination and play are exactly the things we shouldn't be AI-ing away.
New writing by email
Occasional pieces on product, technology and AI — and how they actually play out in practice.
