Every organisation has SOPs, but getting everyone to follow every step in full is actually very hard.
This is usually not a discipline problem but the predictable behaviour of people under time pressure: we tend to do what we feel like doing now, or whatever shows a result immediately, rather than what most needs doing.
When work piled up, getting a single task fully compliant with the business process was very hard: the order of execution between big and small tasks, the task’s own SOP, the records to be left afterwards, and so on. Once there is a lot going on, those steps get dropped one by one, and the person doing the task jumps straight to whatever produces a result right away, because that is the action most likely to quiet down whoever is chasing.
AI does not have this problem. It will execute in the compliant way, as far as it can, tirelessly, because it does not get anxious and does not cut corners just because someone is standing there waiting.
This is the use I think is most underrated: not using AI to do things faster, but to do them more completely.
1. Write-back: the part nobody ever has time for#
Closely related to normalisation is writing data back.
Say a ticket needs to be closed, and a lot of information has to be filled in for the record and for later tasks to refer to. Usually, just finishing the work the ticket is about has already used up the time, and there is nothing left for filling it in properly, never mind weekly reports, monthly reports, work logs and the like.
So it becomes the familiar picture: what was done leaves no trace, and the next person, or you three months later, steps into the same hole again.
For an agent, though, these work records are incidental. It has just finished the task, the whole context is in its hands, and the marginal cost of writing it down is close to zero.
A person has to recall in order to record; it does not.
This compounds a great deal. Once records start being complete, the next task of the same kind no longer starts from blank; you are effectively building a database for your future self, and building it takes almost none of your time.
2. Communication cost: the waiting itself is what is expensive#
Problems used to queue up and get handled one at a time, and during that queue the customer asks, complains to their manager, the manager emails to chase, and now you may have to reply explaining why it is not done yet. The time spent writing that reply could have gone into doing the thing.
The simple fact that work cannot be finished quickly generates a large amount of communication cost by itself.
It is a cost that rarely gets counted, because it does not appear on any ticket; it is scattered across email, meetings and people asking in person.
When the waiting disappears, the losses that the waiting produced disappear with it. What you save is not only the handling time but the whole layer of communication that existed to manage “not done yet”.
In my experience the amount saved at this layer is often larger than what is saved on the task itself.
These two are the same thing#
What they have in common is that both are things that should be done anyway, and that people cannot manage:
- SOPs should be followed, but people skip steps when in a hurry.
- Records should be filled in, but after finishing, people just want to rest.
- Progress should be reported proactively, but people only report when chased.
These are not really ability problems; they are closer to human-nature problems, and AI happens not to have that kind of problem. This is one of the places I find using AI most worthwhile.
If your organisation has a process nobody actually runs, that usually does not mean the process is written wrong; it means nobody is willing to pay the cost of executing it.
That cost can now be outsourced.