Twenty steps.
Some AI tools now do a job as a chain of steps, one after another. Every step has to go right for the job to go right. Set the chain up, then run it a hundred times.
The chain
All 20 steps go right 35.8% of the time: about one time in three.
If each step goes right at the rate across the top, this is how often a chain of that many steps goes right all the way through.
| Steps | Each 90% | Each 95% | Each 99% |
|---|---|---|---|
| 5 | 59.0%about six times in ten | 77.4%about eight times in ten | 95.1%about 19 times in 20 |
| 10 | 34.9%about one time in three | 59.9%about six times in ten | 90.4%about nine times in ten |
| 20 | 12.2%about one time in eight | 35.8%about one time in three | 81.8%about eight times in ten |
| 40 | 1.5%about one time in 70 | 12.9%about one time in eight | 66.9%about seven times in ten |
What it means.
A step that goes right 95% of the time sounds safe. Chain twenty of them and the whole job goes right only about one time in three, because every one of the twenty has to come off.
So the longer the chain, the more often something somewhere goes wrong, even when every step is good. The fix is to keep chains short where you can, and to put a person between the steps that matter, so a wrong turn gets caught before the next step builds on it.
We picked 95% as an example. How often a real tool gets a step right depends on the tool and the job. The arithmetic is the same either way.
More like this on the Agents and automation shelf.
Want to see where your own chains break?
We map where the hours go first, and put a person at the steps that matter.