Most of what happens in a company has happened before. We built Boring Machine to run that work the same way every time, without paying a research-grade mind to work it out again.
AI labs are in a race to build models that can do anything, and frankly, they’re getting close. They’ll draft your board memo or argue philosophy with you at 1:19 a.m.
But every time one of them reads an invoice, you pay for all of that expertise. Paying a research-grade mind to do routine work makes very little sense, especially when it occasionally decides to get creative.
The first time you drove a car, you said each step out loud: mirror, signal, check, turn. Ten years later you drive home and remember nothing of the trip. Psychologists have mapped this for decades,1 and John Anderson named the last step knowledge compilation: rules you used to look up become a procedure you just run.2
Chat training teaches models to please the person in the conversation. Reasoning training teaches them to solve puzzles a test can grade. Nobody trains them to do the same procedure the same way ten thousand times.
So a reasoning model talks itself through each step of your invoice process in words you pay for, then forgets and starts over on the next invoice. Meta researchers showed that slow reasoning can be compiled into a model that answers directly, sometimes doing better than the slow one.3
Frontier models interpret your process on every run. Boring Machine runs it compiled.
Kahneman and Klein agreed that practised intuition can be trusted in a world regular enough to learn, with quick feedback.4 The back office is about as regular as work gets. Yet the people in it still disagree with each other far more than anyone thinks.
When an insurer gave its underwriters the same cases, executives expected their quotes to differ by about 10%. The median gap was 55%.5 Depending on who handles the file, the output varies. And nobody wants a creative accountant (unless you do, but that is a different topic).
The fix has been known for a long time: write the procedure down and follow it. A 19-item surgical checklist cut deaths after major surgery from 1.5% to 0.8%.6 For routine steps, we’d take a procedure over brilliance any day. Boring Machine is a procedure that runs itself.
A token is about three quarters of a word in English, more in French (we’re in Montreal, we know). Every lab counts them differently, and reasoning models bill for words they say to themselves. The same task can cost more on Tuesday than on Monday because the model thought longer.
Boring Machine has nothing left to think about, so a task takes roughly the same time on every run. We sell it the way businesses already buy work: by the hour. Book ahead, like a plumber on Tuesday instead of Saturday midnight, and it costs less.
The more boring the work, the less it costs.
Bury a good clerk under last month’s paperwork and even a good clerk makes mistakes. Models are no different: performance drops as the input grows, even on dead-simple tasks.7 Showing a model only the relevant tools took correct tool choice from 13.6% to 43.1%.8
Boring Machine sees only the tools a task needs, gets a short receipt for each result, and points at earlier results by reference, so it never retypes an amount or an ID.
Not everything is routine. Trust practised intuition when the world is regular, and slow down when it isn’t.
Work out new problems nobody has solved before.
Runs the procedures you already have, every day.
Makes one judgment about a document in milliseconds.
In CloudRaker Run, a large model helps you design the agent and Boring Machine runs it. When a run meets a case it wasn’t built for, it hands the case to a larger model or to a person.
We’ll let the frontier models be brilliant.
Ours is boring, and right.