What 19 sessions at the world’s biggest AI summit actually said

Bosch rolled generative tooling out to 12,000 engineers, who self-reported feeling about 20% more productive. Then Bosch’s Head of AI Governance, Jochen Kokemueller, pointed to a controlled trial of experienced developers that found the opposite: they were 19% slower with AI assistance, and believed they were 20% faster. A 40-point gap between feeling and measuring. Code now ships in an hour and then waits three days for review, and Kokemueller’s read is that the constraint on AI adoption is no longer how fast machines generate, it’s how fast humans can verify. His fix borrows from algorithmic trading: machines checking machines, with a named human accountable at the end.

Meta’s Rohit Patel, Director at Meta Superintelligence Labs, described the engineering version: agent output is nondeterministic, so you need evals, a judge model scoring real outputs against curated golden answers, before you can tell whether your agent is getting better or worse.

That line came from Ferrovial’s Dimitris Bountolos and got the loudest reaction of the two days. His argument: models are commoditizing, so the thing that differentiates one company’s AI from another’s is the context around it — company history, domain knowledge, brand truth, kept current. An unconfigured agent is a new PhD on day one: technically brilliant, with zero knowledge of how your company works. Ferrovial’s answer was a federated context platform, deliberately not centralizing the data, but standardizing the meaning attached to it across departments.

And from a different angle, Profound’s Nick Lafferty shared a number with direct consequences for content teams: half of what AI answer engines cite is less than 13 weeks old. Freshness isn’t a nice-to-have; it’s a ranking signal.

Zalando’s Natalia Andrievskaya opened planning for a handful of automation workflows. Internal hackathons surfaced 102 candidate use cases in one month. Her answer was a feasibility/suitability scorecard and a triage system, with a named owner on everything that moved forward.

Zendesk’s Mirza Beširović showed the cost of skipping that discipline: one pilot ran $1,500/month in testing and then $1M+/month in production, because nobody modeled real costs before scaling. Melissa Reeve named the underlying structural issue: in most organizations, the same person who builds an AI tool is also the one who maintains it — the exact arrangement the software industry abandoned in 1968.

BOI’s Amir Ouki put the services-firm version of this on the table directly. A firm that delivers twice as fast can still end up smaller, if clients expect to pay half. His example: personal-injury demand letters, $2,000–$3,000 when a junior lawyer drafts them and about $500 when AI drafts and a lawyer reviews. His prescribed fix is pricing the outcome instead of the hours — Sierra AI now charges per resolved support issue rather than per agent seat.

BOI’s Stefano Orani set the market context: enterprise software spending reached $1.47 trillion, with AI capabilities accounting for roughly 85% of the year-over-year growth. McKinsey’s latest State of AI survey puts two numbers next to each other: 80% of people say AI makes them more productive, but only 6% of organizations tie AI activity to measurable profit. That gap is the story.

Two speakers, different industries, the same shape of change. Coca-Cola’s Dominik Heinrich said production time in creative roles has dropped from ~70% of the job to 30–40%, with craft and ideation time rising to fill the gap. Some companies are rehiring after discovering AI output still needs human refinement to meet the bar.

IKEA’s Explore team compressed a strategic exploration cycle from 18 months to 18 weeks to 18 minutes for a full simulated run — not by working faster, but by killing the decision paralysis that used to force them to choose one path before testing any. BOI’s Laura Stevens added the caveat: organizations default to familiar, low-uncertainty choices, and unless incentives and approval processes change, AI just speeds up the existing bottleneck.