Read time: 4 minutes

Welcome back, leader.

Big news week. Anthropic is now worth more than OpenAI, Claude Opus 4.8 dropped yesterday, and two of the loudest CEOs in AI walked back their jobs predictions. None of it matters as much as the one quiet admission from Uber's COO that I want to start with.

This week, Sam Altman (OpenAI) and Dario Amodei (formerly OpenAI, now CEO of Anthropic, makers of Claude) both admitted they were "pretty wrong" about AI. The white-collar jobs apocalypse they spent the last year predicting hasn't arrived.

Cynical read: both companies are heading for IPOs this year, and "we're going to end your job" is bad pre-listing copy. Worth keeping in mind.

But the honest read is more useful to you. The Uber admission and the walk-backs from Altman and Amodei are different angles on the same story. The jobs didn't disappear and your AI spend isn't moving the bottom line because we bolted AI onto an organisation built for a world without it.

Here is the analogy I keep coming back to.

When factories first got electricity, they used it for lighting.
Brighter floor, safer than gas lamps, but the work flowed exactly as before.
That is most people's first use of AI: faster emails, quicker first drafts.
The individual speeds up. The business doesn't.

Then factories got smarter. One electric motor driving a cluster of machines, replacing the old steam engine, saving on fuel. Cheaper, but the layout never changed.
That is where the keen companies sit now: agents handling whole workflows, screening more CVs, clearing more tickets, all bolted onto processes designed before anyone had heard of a language model. It saves money.
It doesn't transform anything.

The real shift took 30 years. Ford realised you could give every machine its own motor, which meant you could rip out the old layout and arrange the floor around how work actually flowed. That is the assembly line. Between 1919 and 1929, US manufacturing productivity grew more than 5% a year.

Almost nobody is here with AI yet. That is stage three, and it is the only stage where the gains actually compound.

That is where Uber sits. 70% of their code is AI-generated, and the COO can't draw a line from any of it to anything customers feel. Everyone is faster. The firm is barely quicker. The individual gains pile up and then queue behind approval steps, sign-offs and review cycles that never got any faster.

You cannot buy your way out of this. There is no license for it. The work is harder than swapping software.

It starts with decoding how your team actually operates.
Then splitting every workflow into what gets fully automated, what stays AI-assisted with a human checking, and what should not be touched at all.
Then wiring the new flow into the systems you already run on, NetSuite, Sage, Xero, whatever you've got, without forcing a migration.
Then running it in shadow mode beside the human first, so you can see where it gets things wrong and tune it.
Then going live, with the people who own the process bought in from week one, because the smartest agent in the world dies on contact with a team that wasn't consulted.

That is what redesigning around AI actually looks like. Most companies are still at stage one or two, which is exactly why the window to pull ahead is wide open.

Anthropic dropped Claude Opus 4.8 yesterday. It is genuinely impressive. It will not move your ROI an inch.

The model stopped being the bottleneck three versions ago. The bottleneck is everything around it: change management, implementation, data that isn't clean, security and access nobody scoped, and a finance team still running ChatGPT like it's a smarter Google. A better model doesn't fix any of that. It just widens the gap between what's possible and what you've actually deployed.

The teams getting real ROI didn't wait for 4.8. They picked a workflow, redesigned it around AI, wired the data, and built feedback loops so the system gets sharper every cycle. That was the hard work at 4.5. Still the hard work at 4.8. Still the hard work at 5.

The model isn't what you're behind on.

Most teams sprinkle AI across everything and wonder why nothing compounds. Try this mental model instead. Every workflow your team runs falls into one of three buckets.

The first is full automation. Pattern-based, low-stakes, high-volume work where the decision is the same every time and the cost of getting one wrong is small. Invoice matching against POs. GL coding for recurring vendors. Chasing approvals. Populating standard reports. Let the agent run.

The second is AI-assisted, human checks. Judgement matters but the pattern is there. First drafts of variance commentary. Board pack narratives. Forecast scenarios. AP exception triage. The AI does 80% of the work and you check the 20% that needs you. The team gets faster without losing the judgement layer.

The third is keep with humans. Pricing calls. Auditor conversations that matter. Renegotiating with a key vendor. Whether to back a new product line. High stakes, ambiguous, relationship-driven. AI can prep the inputs, but the decision stays with a person.

The most expensive mistake I see is putting bucket-three work into bucket-one. The second most expensive is treating everything as bucket two and never actually freeing anyone's calendar. Sort honestly, start with one bucket-one workflow, and the gains start to compound.

Pick one recurring task your team runs. Just one. Month-end commentary, the weekly forecast pack, AP exceptions, whatever's eating your Friday afternoons.

Write it down. Then decide which bucket it lives in. Full automation, AI-assisted with a human checking, or keep with humans. Be honest about why.

15 minutes, on a Monday morning. You'll have your first candidate, or your first clear "leave it alone." Either is a win. Most teams have never asked the question that explicitly, and that's exactly the gap.