Putting someone in front of the held queue is not the same as giving them power to decide it. Access lets a reviewer see an agent's held actions. Authority lets them approve, decline or escalate within defined limits, with backing when they say no. Without authority, reviewers defer or approve to avoid friction, and the control disappears.
What is the difference between access and authority?
Access is a permission in a system. The reviewer can open the queue, read the agent's draft, see the sources and press a button. It is granted by an administrator, often in bulk, and says nothing about what the person is expected to decide.
Authority is a decision right. It defines which held items this person may decide, up to what value or depth, with what options: approve, edit and approve, decline, or escalate. It is granted deliberately, usually tied to competence, and it carries an expectation that the person will use it.
A reviewer can have full access and almost no authority. They can see everything and are expected to approve most of it, with any decline needing someone else's sign-off. In that situation, they are not reviewing; they are relaying.
How does missing authority show up?
In patterns, usually visible in the record before anyone talks about it:
- Near-total approval rates. Reviewers approve almost everything, because declining needs a justification they are not sure they are allowed to make.
- Escalation of anything awkward. Items that a reviewer should decide go up to a manager, creating a bottleneck.
- Inconsistent outcomes. Similar held items are decided differently by different reviewers, because nobody defined who may decide what.
- Reviewers overruled informally. A reviewer declines a held refund and a manager approves it later without recording why.
Auditors will see these patterns. When they sample the reviewers, not just the agents, one of their first questions is whether each reviewer was authorised to make the decisions they made.
What does real authority need?
Four things.
A written matrix. Who may decide which held items, up to what limits. An approver authority matrix for a contact centre might give certified advisers authority over refunds up to one level, team leads above that, and two people for anything hard to undo. A professional services firm needs a different shape; see the approver authority matrix for a practice team.
Competence behind it. Authority should rest on evidence that the person can judge. A business can choose to tie each right in the matrix to a current certificate.
Backing. When a reviewer declines a held item within their authority, the organisation stands behind them. If a manager overrules, it is recorded with a reason.
Capacity. Authority is meaningless if the reviewer has no time to use it. A reviewer with a queue twice what they can handle will approve to keep up. Workload balancing for reviewers is part of giving them authority in practice.
Where is this hardest?
In areas where reviewers sit below people with commercial power. Wealth management is a clear example. An agent drafts a client communication or prepares a transaction, a reviewer holds it because it strays toward advice the agent may not give, and a senior relationship manager wants it sent. Unless the reviewer's authority is clear and backed, the relationship manager wins by default. The guide on which roles in wealth management need AI agent training shows how many people sit in that chain.
The same tension appears in sales-led businesses, where credit controllers face pressure from account managers, and in clinics, where administrative reviewers face pressure from clinicians.
What is the trade-off?
Giving reviewers real authority means they will decline things. Some of those declines will be wrong, and some will annoy customers or senior colleagues. Managers lose some control over individual outcomes.
Defining authority also takes work. Matrices need writing, agreeing, maintaining and checking. A matrix that is too rigid frustrates people; one that is too loose provides no structure.
These costs are the point. A reviewer whose declines are never wrong is probably not declining enough. The aim is calibrated authority: clear limits, competence behind them, backing when used, and a recorded route for disagreement. Mistakes within that structure are visible and correctable. Mistakes outside it, where reviewers approve because they feel they cannot refuse, stay hidden until they are expensive.
What should you do differently?
- Write an authority matrix for every team that reviews agent work, by type of held item and limit.
- Tie each right to competence, ideally a current certificate.
- Back reviewers' declines and record any overrule with a reason.
- Balance workload so reviewers have time to use their authority.
- Watch the patterns: approval rates, escalation rates and inconsistent outcomes tell you whether authority is real.
The guide to AI agent skills by role sets out what each role in the chain needs.
Where the Academy fits
Pryme Intelligence Academy's Operator track trains reviewers to use authority well: how a governed agent decides, human review, reading the workspace and when something goes wrong. Its practicum requires real decisions, including declines, with no missed red line. A business can then make approval rights depend on a current certificate, depth by depth.
Questions people ask
Should a reviewer ever be able to approve something above their limit?
No. Items above their limit should route to someone with authority for that level. If that happens often, the matrix or the agent's limits may need revisiting.
How do you handle a reviewer who declines too much?
Sample their declines and talk to them. Sometimes they are right and the agent's drafts need work; sometimes they need coaching on the criteria. Either way, the record tells you which.
Is access without authority ever appropriate?
Yes, for people learning the work or observing it, such as those in training or audit. Their access should be clearly read-only, so nobody mistakes it for a decision right.