The Permission Never Changed
Ask what your automations are allowed to do and somebody will answer. Ask who accepted that they should be, by name, and the question usually has no owner.
A step gets approved to draft. Months later somebody automates the step after it, for good reasons, and the draft now goes out.
Nothing about the first approval changed
March: a service delivery lead approves the assistant drafting renewal notices. It writes the letter, a person reads it, the person sends it. The approval is recorded and it is entirely sound.
September: a different team removes a bottleneck. Drafted notices were sitting in a queue for days, so they wire the send step to run automatically once the draft is produced. Reviewed, tested, shipped, celebrated.
A Tuesday in October, a renewal notice with a wrong figure reaches four hundred customers. Nobody edited the March approval. Its text is identical. What it authorises is now completely different.
The person who approved the drafting has no idea the sending changed, because nothing in the system connects the two facts and there was no reason for anybody to tell her.
Two kinds of authority
There is the authority an organisation has declared, and there is the authority actually in force, and the second one moves on its own.
It moves when a credential is reused for a second purpose. When a pilot quietly becomes a workflow. When the thing downstream of a permission is replaced by something faster. None of those touches the declaration, and all of them change what it permits in practice.
The declaration is a statement about what may happen. What is in force is a fact about the system as it stands this morning, and the two were last identical on the day the declaration was written.
Nobody is being careless
The declaration and the use are kept in different places, and only the declaration is ever read.
An owner reviewing an automation opens the console, reads the permission, confirms it is appropriate, and signs. Everything about that activity is diligent and it answers the wrong question. An owner can study what an agent was permitted to do and learn nothing whatever about what it has been doing.
The gap is not even hidden. It is just that one side of it has a page in the product and the other side does not exist as a record anywhere, so the review goes where the information is.
Why the quarterly review misses it
The standard control here is a periodic entitlement review. Worth tracing what happens when it meets this case.
The reviewer sees that the assistant may draft renewal notices. That is correct. It was correct in March and it is correct in October, and the reviewer approves it again in good conscience.
The thing that changed was not in scope. It was two steps away, in a different team's deployment, and it was not a permission change at all. It was a workflow change that altered the consequence of a permission nobody touched.
A control that re-reads the declaration can only confirm the declaration. To catch this it would have to notice that the behaviour downstream of a permission is not what it was when the permission was accepted, and no entitlement review is asked to look at that.
Put the acceptance on the act
Before a class of action can run, a named person accepts that class, and the acceptance is recorded against them with what they saw when they made it.
Not a setting in a console somewhere: a person, on the record, attached to the thing that is going to happen. The difference shows up in October. Sending a renewal notice without human review is a different class of action from drafting one, so it needs its own acceptance, and there is nobody it could have been recorded against because nobody was asked.
Then every exercise of it writes its own line. What ran, who had accepted it, and what was in front of them when they did. The record comes out of the work rather than being assembled afterwards, which is the only form of it anybody can rely on a year later.
What raises a hand is an event
Not a scheduled look. The first time a class of action is used. A permission that has gone quiet for a quarter. An agent doing work nobody accepted.
Those three catch different things, and the third is the one a quarterly review structurally cannot. It fires on the mismatch itself: an act running with no acceptance behind it, caught as it runs rather than at the next review date, by comparing what happened against what was accepted.
The first use matters for a quieter reason. A class accepted in March and first exercised in October has had seven months for its context to change, and for the person who accepted it to forget. An event on first use puts the acceptance back in front of somebody while the act is fresh.
That layer is what we build, and Task Accountability Infrastructure is the name for it. What may happen is a decision. What did happen has to be a record.
Questions
What is the difference between declared authority and authority in force?
Declared authority is what an organisation has written down that a system may do. Authority in force is what the system can actually cause to happen this morning. The second moves without the first being edited, when a credential is reused, a pilot becomes a workflow, or a downstream step is automated.
Why do access reviews miss changes in what an automation can do?
Because a review re-reads the declaration, and the declaration has not changed. In the common case the permission is still correct and what changed was a workflow two steps downstream, in a different team's deployment, which no entitlement review has been asked to look at.
How should an automation's authority be recorded?
As an acceptance by a named person of a specific class of action, with what they saw when they accepted it, bound to the act rather than held as a console setting. Every exercise of that authority should then write its own line recording what ran and who had accepted it.
When should an agent's permissions be re-examined?
On events rather than on a date: the first time a class of action is actually used, a permission that has gone quiet for a quarter, and any act running with no acceptance behind it. The last of those is the case a scheduled review structurally cannot catch.
Does an approval need re-confirming when a workflow changes?
Where the change alters what the approval causes, yes, and that is the hard part to detect. Approving an assistant to draft a letter is a different class of action from approving one to send it, so automating the send step requires an acceptance that nobody has been asked for.