Approval is not authority: what an audit actually asks
An approval records that somebody clicked approve. An audit asks whether that person held the authority for that decision, and whether the approval still refers to the text that exists today. Those are different questions, and most tools answer only the first. Authority binding checks the decision against the roles your organisation actually holds, and version binding ties the approval to the exact content, so a later edit reopens review instead of inheriting a green tick.
The step everybody ships
Approval gates are now table stakes. Every workflow and agent product has one, and that is a good thing: it is the mechanism that keeps an automated step from doing something irreversible on its own.
It is also not the question a regulated organisation actually needs answered. Knowing that somebody approved a step tells you an event occurred. It does not tell you that the event counts.
The question an auditor asks second
Walk through a real review of an automated process and the questions arrive in a fixed order. What happened. Who approved it. Then, almost immediately: were they allowed to?
In most tools the answer to the third question is unsatisfying. Approval was granted by whoever the request was assigned to, or whoever was in the channel when it appeared. The system recorded a decision. It did not check whether that person held the authority to make it.
In a bank, a hospital or a public body that difference is the whole control. An approval from the wrong person is not a slightly weaker approval. It is no approval at all, carrying a timestamp that makes it look like one.
MindKeepr captures what your team knows and keeps it usable, even after people leave.
Binding authority to the decision
The fix is to make authority part of the decision rather than a property of the queue. Before a step runs, it is checked against the roles the organisation actually holds: who may decide this kind of thing, at this value, in this part of the business.
That sounds administrative until you try to write it down. Most organisations have a delegation of authority matrix somewhere in a policy document, and almost none of them have it connected to the systems where decisions are actually made. The gap between the two is where audit findings come from.
One rule follows directly and does not bend: the person who wrote an answer is not the person who approves it. Separation of duties is not a preference, it is the reason the control exists.
Binding the approval to a version
The second half of the problem is time. A procedure is signed off in March. Somebody edits it in June. Open it today and the sign-off is still sitting there, looking exactly as reassuring as it did on the day it was given.
Most systems treat approval as a property of the document. It is not. It is a property of a specific version: the exact text the approver read. The moment the text changes, the approval refers to something that no longer exists.
So in MindKeepr an approval binds to that version. Change the content and the approval does not quietly carry over: the review reopens. The green tick stops being decoration and becomes evidence.
Why this matters more with agents in the loop
When a person acts on a stale approval, there is usually a second person somewhere in the chain who notices that something looks wrong. An agent has no such instinct. It will act on the most recent thing it can retrieve, at the speed you allowed it, on every case that matches.
Salesforce published its reasoning for Agent Script in June 2026 and described the failure plainly: with too many rules in a prompt, the model skipped steps that were supposed to be mandatory and lost its place when a conversation wandered. Their answer was to take the business logic out of the prompt and run it deterministically.
We think the same argument applies one layer down. Rules and approvals should not live somewhere a model can forget them halfway through, and the knowledge an agent acts on should carry the evidence that a person with the right authority stood behind it.
What to ask a vendor
Five questions separate a real control from a convincing interface. Where does the system run, and under whose identity model? Does every answer respect the permissions a person already has in the source system? Can you see who declared an answer, when, and what it was based on? Is an approval bound to the exact version it approved, and does a change reopen review? And does the product score, rank or profile individual employees?
That last one decides whether your works council and your data protection officer will let the system near real work at all. Measuring the state of organisational knowledge is a different activity from measuring a person, and a product that blurs the two will not survive procurement in Europe.
- ✓An approval from the wrong person is not a weaker approval, it is no approval with a timestamp.
- ✓Authority is a property of the role, not of whoever happened to be in the thread.
- ✓An approval belongs to a version of the text, not to the document.
- ✓A stale sign-off certifies with full confidence something nobody reviewed.
- ✓The person who wrote an answer should never be the person who approves it.
FAQ
An approval that attaches to the exact text that was approved, rather than to the document as a whole. If the content changes, the approval no longer applies and the item returns for review.
Approval records that somebody confirmed something. Authority is whether that person was permitted to decide it, checked against the roles the organisation holds rather than who happened to be available.
No. The person who wrote an answer is never the person who approves it. That separation is the reason the control is worth anything in an audit.
The approval that relied on the previous version stops applying and the review reopens, rather than the old sign-off silently covering new text.
It moves the effort earlier. The alternative is discovering during an audit, or during an incident, that a decision nobody was authorised to make has been running in production for months.
Start free in minutes, or get a demo on your own tools and team.
Sarim Zafar is the co-founder and CEO of MindKeepr. He has spent twelve-plus years building and scaling cloud and AI platforms for large organisations, and still writes the code behind MindKeepr's governed memory.