It depends, and no universal number holds. The deciding factor is how long a real check of one held item takes, multiplied against the focused time a reviewer actually has. Measure both on your own work, set a cap per reviewer from that arithmetic, and treat a reviewer who approves faster than the check can be done as a warning.
Why is there no standard figure?
Because approvals differ more than they look. Releasing a supplier payment that matches a purchase order, a receipt and a known bank account takes a minute of checking. Approving a hardship arrangement for a vulnerable customer can take twenty. A count of approvals per day says nothing until you know which kind they are.
The useful unit is minutes of scrutiny per item. If a careful reviewer needs four minutes per item and has three focused hours a day for approvals, the ceiling is about 45 items. Anything above that is not approval; it is clicking.
What drives how many a person can handle?
| Factor | Raises capacity | Lowers capacity |
|---|---|---|
| Evidence in the draft | Every figure points to its source | Reviewer must open three systems to check |
| Similarity of items | A queue of one item type | Mixed types that need context switching |
| Stakes per item | Low value, easy to reverse | High value or hard to undo |
| Exception rate | Most items are clean | Many items need a judgement call |
| Interruptions | Protected approval time | Approvals squeezed between other work |
| Familiarity | Reviewer knows the process and the customers | Cover reviewer from another team |
The first row is the one you control most directly. An agent that drafts a payment with the invoice, the purchase order, the receipt and the bank account check attached saves the reviewer far more time than any interface change.
What should the agent do, and what stays with the reviewer?
| The agent does | The reviewer does |
|---|---|
| Prepares the item and attaches every source | Checks the item against the sources |
| States which bound caused the hold | Decides whether the hold reason is satisfied |
| Groups similar items so they can be reviewed together | Approves or rejects each one individually |
| Flags anything unusual against past cases | Judges whether the unusual thing matters |
| Routes to named cover when the reviewer is away | Accepts or declines the cover |
Note what the agent does not do: approve. Another agent checking the first one's work is not a substitute for a person, for reasons set out in why one agent should not approve another agent's work.
How does this play out in practice?
Take a regional lender's operations team. An agent drafts loan disbursements and holds each one for approval. One team lead approves them alongside other duties. When volume doubles after a product launch, her approval time per item falls by more than half, and her rejection rate drops to almost nothing.
Nothing about the loans got simpler. She was rushing. The team responds in three ways. They split the queue: disbursements that match a fully verified application with no flags go to one reviewer, anything with a flag to a second. They cap each queue at what a careful check allows. And they look at whether the cleanest category has enough certified evidence to move to depth 3, with sampling instead of item-by-item approval. That last step is where most of the relief comes from, because the uneventful case is the common one.
Which controls keep approvals meaningful?
- A cap per reviewer, set from your measured minutes per item, with overflow routed to named cover.
- Time-on-item monitoring. Approvals faster than a plausible check are flagged for review.
- Rejection rate tracking. A reviewer who never rejects is either seeing perfect drafts or not looking.
- Approvals that expire. A held item that sits too long should return to the queue for a fresh look rather than wait to be cleared in bulk; see whether AI agent approvals should expire.
- Depth reviewed, not fixed. When a category is consistently clean and evidenced, consider raising its depth so people approve fewer, harder items.
Capacity planning gets harder as agents multiply, because each new agent adds its own held queue; see scaling from one AI agent to many. Approval design is one part of agent governance: bounds checked at the moment of action, with a named person wherever a bound holds the work.
Questions people ask
Should approval targets be part of a reviewer's performance goals?
Not as a volume target. Rewarding throughput encourages rubber-stamping. Measure accuracy of approvals found in later sampling instead.
Does batching approvals help?
Batching similar items helps the reviewer build context. Approving a batch with one click does not; each item still needs its own decision and record.
What if the cap means held items wait too long?
Then you have a staffing or depth question, not an approval question. Add named reviewers, or gather the evidence to move clean categories to a higher depth.