Anthropic released Fable this week, and the discourse followed the usual script: benchmark charts, superintelligence timelines, someone declaring the end of white-collar work by 2030. Most of that conversation is aimed at other AI labs. Very little of it is aimed at the people who actually run businesses.
So here is the operator's version. Fable is genuinely remarkable at some kinds of work and genuinely unremarkable at others, and the line between the two is not how hard the work is. Once you see where the line actually falls, you can predict, with useful accuracy, the order in which work gets automated — including yours.
Why models got good at what they got good at
For years, model capability tracked a simple recipe: more compute, more data, better results. That recipe still works, but the step change of the last two years came from somewhere else. Instead of training models on examples of correct answers, the labs started training them on tasks whose answers can be checked automatically.
Give a model a hard coding problem. Let it try a thousand times. Run every attempt through the test suite. Keep the attempts that pass, and train on those. No human had to write the ideal solution; a machine just had to recognise one. One well-built task can generate thousands of training signals, so the labs stopped buying answers and started buying checkers.
The consequence is written all over Fable's benchmark results. The model is astonishing at software engineering, mathematics, and structured analysis, work where an attempt can be scored in seconds by a compiler, a test, or a proof checker. It is far less impressive at long-horizon, open-ended business tasks. The now-standard benchmarks that ask a model to run a small business over weeks show the same thing every generation: the models that demolish coding evaluations still stumble over inventory, suppliers, and customers.
That is not a temporary embarrassment the next release will fix. It is the shape of the training pipeline showing through. Capability grows fastest wherever an answer can be verified cheaply. Everywhere else, it grows slowly.
The frontier does not advance evenly
This is the part that matters if you run a company. The frontier of AI capability is not a rising tide that lifts every job function at the same rate. It advances along a gradient of checkability.
Software engineering went first, not because code is the most valuable work in the economy, but because it is the most verifiable: bounded inputs, tests that pass or fail, versioned files, reviewable diffs. Point the training recipe at that and you get the extraordinary coding agents everyone now takes for granted.
Now look at a finance close, a sales pipeline, or a maintenance schedule. The work spans half a dozen systems that do not share state. The real process differs from the documented one. The output is "the numbers were right" or "the machine did not break," which takes a person days to confirm. None of it is checkable by default, so the training recipe has far less to grip, and generic AI dropped onto it underperforms — not because the model is stupid, but because nothing in its training ever looked like your Tuesday afternoon.
The question executives ask is the wrong one
Every leadership team we talk to eventually asks some version of "when will the models be able to do X?" The honest answer is that for most operational work, the model has been able to do X for a while. The real question is: what would it take to check whether X was done correctly?
Because the moment that check exists, everything changes at once:
- Today's models clear the bar. A workflow decomposed into steps with defined inputs and a scoring rule at the end is exactly the shape of task current models already excel at. You do not need to wait for the next release.
- Accuracy compounds instead of decaying. When an agent's output can be verified, every correction becomes a training signal. This is why agents we run in production climb from high-80s accuracy in their first week to high-90s within a couple of months, and why unverifiable deployments plateau and get quietly switched off.
- Trust becomes a policy, not a feeling. "Should the agent be allowed to do this autonomously?" stops being a debate and becomes a threshold you can read off a dashboard.
Making work checkable is not a research problem. It is engineering and process work: defining what the inputs are, writing down the rules everyone carries in their heads, logging outcomes, and decomposing the workflow until each step can be verified on its own. It is unglamorous, and it is the entire ballgame. When we audit a business, most of what we produce is precisely this: a map of which work is already verifiable, which work could be made verifiable cheaply, and which work genuinely needs a human.
What about superintelligence?
The timelines debate will run for years, and you do not need to pick a side. Here is the arithmetic that holds under every scenario.
If a workflow in your business is verifiable, current models can already run it, and the return on deploying an agent today is measured in months, not decades. If a workflow is not verifiable, a dramatically smarter model mostly fails at it more fluently, because the bottleneck was never intelligence — it was the absence of a check. Either way, the expensive strategy is waiting.
And if the optimists are right, the ordering matters even more. Whatever the labs release next will arrive first at the businesses whose operations are already instrumented, decomposed, and checkable, because that is the substrate the technology runs on. The companies that did the boring work early will absorb each new model generation in weeks. The ones that waited will still be arguing about pilots.
Fable is a milestone, and there will be another one within the year. Our advice does not change with the release cycle: stop tracking the models and start instrumenting the work. If you want to know which of your workflows are already on the right side of the line, that is a two-week question, and we answer it for a fixed price.