Meta's AI transformation failure

Meta's top-down plan to replace staff with frontier AI models imploded, according to a Reuters investigation, with major security and technical incidents rising 40% and time spent on those incidents up 70% amid a 220% increase in changes to core platforms, while customer-facing changes rose only 36%. The commentary attributes the failure to imposing structural change on a culturally governed organization and to a category error by executives about what large language models can do, arguing LLMs create value where inputs are unstructured and output errors are tolerable. The author points to post-sales work, where AI bolted onto Customer Success, Support and Delivery tools improved speed, efficiency and customer satisfaction, as the more realistic path to new operating models.

Frontier models are not general digital employees. The real-world evidence of Meta's failed AI transformation in this Reuters article https://www.reuters.com/investigations/mark-zuckerberg-had-bold-plan-replace-meta-staff-with-ai-heres-how-it-imploded-2026-08-26/?utm source=AMP&utm medium=email&utm campaign=30-of-companies-diy-d-an-ai-tool-vs-buying-software-this-year& bhlid=16e6ec4e4c5fb2bfaea452f501bfe3ee14578faf makes it pretty clear that frontier models are not replacements for existing employees that can be rolled out top-down to reduce headcount with no operating consequences. Worse, the evidence in the article points to this style of deployment causing new problems. Major security and technical incidents went up 40% and time spent on those incidents went up 70%. Not too surprising given a 220% increase in changes to core platforms. Forgivable if the value is there, but customer-facing changes rose a more modest 36%. This is activity decoupling from value. The failure at Meta seems like it has two causes. First is trying to impose structural top-down change on a fundamentally culturally governed organisation. That's worth a post on its own some other day. But the second is a fundamental category error by the executives on what LLMs are actually capable of. And they are undoubtedly capable. Just as machine learning drove a boom in new technology wherever the problem space was a categorisation problem, so too will LLMs find their problem space and generate real value. What is that space? Well, that's the trillion-dollar existential question for the likes of Meta and other frontier model builders. To me, perhaps the most slept-on capability that has no equivalent in other tools is how well LLMs cope with unstructured input and turn that into pretty useful output most of the time. Any problem space that has messy inputs and a tolerance for error on the output is worth exploring. In the post-sales space I live in, I've improved speed, efficiency and customer satisfaction through early adoption of the AI capabilities now bolted onto Customer Success, Support and Delivery tools. Deflecting tickets and surfacing delivery risks improve KPIs and client outcomes, but they don't create opportunities for new operating models. Likewise, the acceleration in actual engineering and dev work means deliveries are cheaper and faster, but not fundamentally different - same model, better unit economics. Where I'm really intrigued is whether the capacity increase across the board by AI compressing work can open up new operating models. Are Agile and Scrum going to sunset, and what replaces them? Does this compression create spaces for new high-leverage individual contributors similar to what already exist in sales and engineering organisations? Do customer onboarding teams go the way of bank tellers in the face of ATMs? The focus on customer value will uncover the real winners. Tellers famously increased in number after the introduction of ATMs, but pivoted from account admins to value-added service providers. Companies that keep that search for value at the heart of their AI journey will stand a better chance than Meta in actually finding it.