The uncomfortable part of AI, the bit that still divides the conversation, is the impact on jobs. The way the media reports it is simple: buy AI, shrink the workforce, watch productivity grow. We have already seen companies run at that story, and we have already seen them have to row back.
Klarna was the poster child. In 2024, it said its OpenAI assistant was doing the work of 700 customer-service agents. Then in May 2025, the CEO, Sebastian Siemiatkowski, told Bloomberg that quality had fallen, and that they were recruiting humans again. "Really investing in the quality of the human support is the way of the future for us." The chatbot did not get switched off; it still handles about two-thirds of chats, but what changed is the admission that you cannot expect AI to solve everything.
Meta showed what this looks like at scale
Then last week, Meta showed what this looks like at scale. Internally, Meta had a plan called "Project OT". The theory was simple: AI would take over much of the daily work, specialists would become general-purpose builders, and middle management would thin out. Some teams were looking at headcount cuts of as much as 60%. A second wave of layoffs was planned for November. But then it stopped.
But why? Did Meta suddenly have a change of heart? Did they suddenly stop believing in AI? The story is actually much deeper, much more interesting and is a story of input vs. output.
Input soared. Code changes to Meta's AI platforms were up 220% year on year, but features that reached users rose just 36%.
The more interesting number is what happened next: technical and security incidents rose 40%, and time spent trying to solve those problems rose 70%. That is a perfect demonstration of the "AI fire hose".
Shipping new code used to look like output
Shipping new code used to look like output. Now it is input. Creating a deck used to be the end goal. Now that is input too. These are the things you need in order to get to the output. Creating the code and shipping the PR is input. Getting features out that people actually love is output, and the output is the thing that matters. A deck about your Q4 strategy is input. Delivering the plan that moves the primary objectives of that strategy is output.
What we are seeing is organisational input exploding, and all of us drinking from the fire hose of AI. That work then gets pushed down an ever-slimming funnel. The deck a junior was asked to do for their manager gets turned around in an hour. The junior skims it; the analysis seems basically right, and it goes up the chain. The manager then has to sit down with that junior and unpack it. What do they actually mean by these sentences? Where do these numbers come from? Which bits of this are actually true?
The term for this, broadly, is work slop. What Meta is showing us is what happens when that reaches the scale of an organisation that size. These are the conversations we need to be having.
Put simply, the act of doing has got cheaper. The act of reviewing and judging has got more expensive.
A workforce problem, not a technology one
A similar story surfaced last week in banking. Dr Rita Fontinha, at Henley's World of Work Institute, told The Fintech Times that AI in banking is a "workforce problem, not a technology one". AI creates value only when employees have the confidence, skills and support to use it effectively and responsibly.
We are already seeing organisations take serious steps towards this. NatWest has combined tools with training for about 60,000 colleagues and a bank-wide AI and data ethics accreditation. HSBC has mandatory responsible-AI training and an AI Academy from beginner to advanced. These are a start. They do not, on their own, tell you what happens downstream.
One-off training is useful for getting people the confidence to begin. If you are responsible for operations and transformation, you still have to ask what happens at scale. How do we make sure the team has the judgement not to push the bottleneck to a different place in the organisation? The bottleneck used to be input. It no longer is. If all we have done is move the bottleneck somewhere else in the organisation, have we really solved anything with AI, other than bringing more cost?
That is the core of the narrative I am seeing everywhere right now. People have bought the tools, and they are not seeing the outcomes. That is where AI hits a human-driven plateau. Usage climbs. Outcomes do not, until people can specify the task clearly, apply the tool inside a real workflow, and get stronger at evaluating, critically, what comes back.
And that gap only compounds, because the models are still getting better. This week Claude Fable 5.1 landed: same headline price as Fable 5, much cheaper cache reads, stronger on the long writing and review work. The prose it brings back needs far less disentangling. It is a genuinely capable day-to-day model. But as the models get stronger and people get less practised at judgement, the opportunity cost for organisations gets worse.
That is the joy of the smaller, leaner, AI-first teams. They move quicker. They adapt to tool changes. They spend more time specifying a clear task, bringing the right context and the right success criteria, before they point AI at it, rather than getting the job done and throwing it over the wall for someone else to unpick. That is the challenge scaled organisations have to face.
Organisations will keep rolling out the tools. The missing layer is still workforce capability you can actually see. Assess the gap. Close it by role. Prove it moved.
Thanks, as always
Andy
Sources
- Will Shanklin, Engadget, 26 Aug 2026, reporting Reuters on Meta Project OT.
- Klarna hiring humans again: Fortune, 9 May 2025; CX Dive, 9 May 2025, citing Bloomberg interview with Sebastian Siemiatkowski.
- Commonwealth Bank reverse: Stephanie Chalmers, ABC News, 21 Aug 2025. Also Bloomberg, 21 Aug 2025.
- Uber 2026 AI budget: Janakiram MSV, Forbes, 17 May 2026; Fortune, 26 May 2026.
- Rita Fontinha, Henley, The Fintech Times, 28 Aug 2026.
- Claude Fable 5.1: Anthropic, 1 Sept 2026.
