Research-backed calculator guide

AI Agent Human-Review Staffing: Size the Queue, Shifts and Shrinkage

Calculate the reviewers, shift seats, paid FTE and monthly labor cost required for an AI-agent human-review queue, including peaks, coverage and shrinkage.

What the calculator needs to measure

1. Agent outputs and review demand

Use a period that your logs can support. Daily outputs are convenient for a continuously operating agent, while weekly or hourly data may be better for a batch pipeline. Separate mandatory escalations from quality sampling. If 10,000 outputs are created daily and 12% are sent to a person, the mandatory queue contains 1,200 items. If you audit 1% of the remaining 88%, add 88 sampled items. Total planned review demand is therefore 1,288 items per day.

Do not treat the escalation rate as a universal model-quality target. It is a measured property of your workflow, policy and threshold. It can change when you alter prompts, tools, confidence gates, risk categories or the share of actions that are reversible. Keep mandatory review and random audit in separate fields so a lower escalation rate does not accidentally remove your monitoring sample.

2. Seconds per review

Time a representative sample from the point at which a reviewer opens the case to the point at which they approve, edit, reject or escalate it. Include evidence reading, required policy checks, note-taking and reason codes. A screen that reduces clicks can reduce staffing; an approval policy that adds a second evidence source can increase it.

Use a distribution when the work is heterogeneous. For example, a short account-label review may take 30 seconds while a sensitive refund review takes five minutes. A single average is acceptable for a first pass only if you also run a sensitivity case for the slow category. For riskier work, staff by category or use a weighted average based on actual mix.

3. Arrival shape and coverage

Average arrivals hide queues. Let the peak multiplier describe peak hourly arrivals relative to normalized off-peak arrivals. If daily review demand is 1,288, coverage is 24 hours, and the peak multiplier is 2 for four hours, the normalized off-peak rate is:

1,288 ÷ (20 off-peak hours + 4 peak hours × 2) = 46 reviews/hour

The peak rate is 92 reviews/hour. This normalization keeps the total at 1,288 reviews instead of accidentally counting an average hour and a separate peak on top of the same volume.

Coverage is a product choice. A 24-hour queue may require overnight seats even when overnight volume is low. A business-hours queue can hold or reroute work outside its window, but that decision belongs in the service policy. If a customer is waiting, record the maximum acceptable wait. If the action is irreversible or high value, define what happens when the review window expires; do not silently turn a timeout into auto-approval.

4. Occupancy, buffer and shrinkage

Occupancy is the share of a reviewer's logged-in time spent actively processing work. At 80% occupancy and 90 seconds per item, theoretical throughput is 40 reviews per hour, but the usable planning rate is 32 reviews per hour. The remaining time absorbs bursts, evidence loading, handoffs and normal human variability.

A peak buffer is separate from occupancy. In the worked example, a 15% buffer turns 92 peak arrivals into 106 planned arrivals. At 32 usable reviews per hour, that rounds up to four peak seats. You can use a measured percentile instead of a percentage buffer when you have enough hourly observations.

Shrinkage converts staffed seat-hours into paid hours. It includes breaks, training, meetings, coaching, absence, leave and other paid time when a reviewer is not available to clear the queue. BLS compensation data is useful for setting a U.S. benchmark, but your own schedule and employment arrangement should drive the loaded rate and shrinkage assumption. BLS also publishes national occupational wage tables; the current page identifies May 2025 OEWS tables, which can help you choose an occupation-specific wage baseline rather than using a generic average. ^[2]

The staffing formulas

Use these formulas in a spreadsheet or script. Round seats up because a fraction of a reviewer cannot cover a live interval. Keep fractional FTE as an analytical output, then round hiring or vendor commitments according to your operating model.

Daily review demand

R = D × E + D × (1 − E) × A

Where D is daily agent outputs, E is the escalation rate, and A is the random-audit rate applied to non-escalated outputs.

Usable reviewer capacity

C = (3,600 ÷ S) × U

Where S is seconds per review and U is the target occupancy. The 3,600 converts seconds to one hour.

Normalized off-peak and peak arrivals

O = R ÷ ((H − P) + (P × M))

K = O × M

Here H is covered hours per day, P is peak hours per day, and M is the peak multiplier. O is the off-peak rate and K is the peak rate.

Seats and paid FTE

Off-peak seats = ceiling(O ÷ C)

Peak seats = ceiling((K × B) ÷ C)

Seat-hours/day = (off-peak seats × (H − P)) + (peak seats × P)

Paid FTE = seat-hours/day × 7 ÷ paid hours/week ÷ (1 − shrinkage)

Monthly labor cost = seat-hours/day × 30.44 ÷ (1 − shrinkage) × loaded hourly cost

Cost per processed output = monthly labor cost ÷ (D × 30.44)

B is the peak buffer, such as 1.15 for 15%. A strict SLA may need a queueing model instead of this deterministic screen. Do not call the resulting FTE a guaranteed service level without testing it against observed arrival and handling-time distributions.

Worked cost example

Assume a U.S.-oriented planning case, but treat every operating input below as an estimate until your logs replace it. The labor rate is the only dated external benchmark in this example. It uses the BLS June 2026 private-industry employer-cost average of $46.89/hour, not a vendor quote or a guaranteed reviewer wage. ^[1]

The staffing result is driven more by coverage and rounding than by the average number of review hours. The queue creates one seat for all 20 off-peak hours and four seats for the four peak hours. Shrinkage then turns 252 staffed seat-hours per week into 360 paid hours per week, which is nine 40-hour FTEs.

The monthly cost is not the total cost of the AI agent. Add model usage, tool calls, storage, monitoring, evaluation, management and any vendor or BPO fee separately. If the reviewers are contractors, use the actual invoice rate and model paid non-working time differently. If the team is internal, use a loaded compensation rate that includes the employer costs relevant to your location and role. Do not mix an hourly wage with an FTE count that already includes benefits.

Sensitivity: what changes the answer?

A staffing model should expose the levers rather than present one precise-looking answer. Holding the worked example constant, the following directional cases show why a small improvement may not immediately remove a hire:

The last case is a governance trade-off, not free efficiency. NIST's Generative AI Profile describes risk management as spanning the AI lifecycle and identifies human-AI configuration as a risk area. That supports keeping a deliberate control plan alongside the staffing calculation, especially where automation bias, over-reliance or consequential decisions are possible. ^[4] The framework does not prescribe a universal audit percentage or reviewer-to-agent ratio, so choose those values from your use case, policy and observed error data.

How to operate the queue after launch

First, instrument every review request. Record creation time, queue entry time, reviewer start, decision time, outcome, reason code, category, model version and whether the item was mandatory or sampled. These fields let you distinguish a staffing problem from a routing problem. A long wait with low reviewer utilization can indicate a schedule gap or time-zone mismatch. A full queue with high utilization can indicate insufficient seats, too much review time or a threshold that escalates too broadly.

Second, report the distribution, not only the average. Track p50 and p95 wait, p50 and p95 handling time, abandonment or timeout rate, approval-as-is rate, edit rate, rejection rate and audit-discovered error rate. If sampled auto-approvals find errors at a materially different rate from mandatory reviews, the threshold may be hiding risk rather than reducing it.

Third, review category-level economics. The marginal labor cost of an extra approved output is not simply the average monthly labor cost divided by total output. If the extra volume lands in an already full peak interval, it can create a whole additional seat. If it lands overnight outside coverage, it may increase backlog or require a new shift. Use the calculator at the hour or category level when the decision is operationally important.

Fourth, define timeouts. For reversible, low-value work, you may hold the item for later review or allow a constrained fallback. For irreversible, regulated or high-value actions, require a human decision before execution. NIST's AI RMF is a risk-management framework, not a staffing rule, so document how your own approval gate maps to the action's impact and reversibility. ^[3]

Finally, rerun the model whenever a material input changes: new workflow, new model, new action permissions, new coverage promise, new review interface or sustained change in escalation rate. The calculator is most useful as a living operating baseline. It should reconcile to schedules and invoices rather than serve as a one-time launch estimate.

Frequently asked questions

How many human reviewers does an AI agent need?

There is no universal reviewer-to-agent ratio. Calculate review items from output volume, escalation rate and any audit sample, then divide hourly arrivals by measured reviewer capacity. Peaks, coverage hours, shrinkage and rounding can make the paid FTE count much higher than review-hours-per-day divided by an eight-hour day.

What is shrinkage in human-in-the-loop staffing?

Shrinkage is paid time when a reviewer is not available to process queue items. Include breaks, training, meetings, coaching, leave and absence. If shrinkage is 30%, divide required staffed hours by 0.70. Lowering the denominator increases paid hours without changing the number of live seats.

Should I staff for average or peak AI review volume?

Use the average to reconcile total demand and the peak to set live seats. Normalize the daily volume across the covered hours, apply the observed peak multiplier, then round peak seats up. If you have enough history, replace a simple multiplier with a percentile or simulation and check the resulting wait-time distribution.

How do I calculate reviewer FTE for AI escalation?

Use staffed seat-hours/day × 7 ÷ paid hours/week ÷ (1 − shrinkage). For example, 36 staffed seat-hours per day, 40 paid hours per week and 30% shrinkage produce 36 × 7 ÷ 40 ÷ 0.70 = 9.0 FTE. Round the staffing commitment to the schedule or contract you can actually operate.

What hourly cost should I use for reviewers?

Use your loaded hourly cost, not salary alone. Include employer-paid benefits and other recurring employment costs that are actually attributable to the role. As a dated U.S. benchmark, BLS reported average private-industry employer compensation of $46.89 per hour worked in June 2026, with wages and salaries at $32.82 and benefits at $14.07. ^[1] That average is an orientation point, not a quote for your role, geography or vendor.

Does a lower escalation rate always reduce FTE?

No. Seats are discrete. A lower escalation rate can reduce review hours while leaving the same one or four seats in place until demand crosses a rounding boundary. It may still reduce backlog risk, overtime or the need for future seats. Measure the effect on peak seats, paid hours and quality rather than assuming a linear saving.

Is this a replacement for Erlang C or a workforce-management system?

No. This page provides an auditable capacity screen for budgeting and early planning. It does not model an exact waiting-time probability, abandonment, multiple priority classes, skill-based routing or time-varying arrivals. Use a queueing model or simulation when a service-level commitment, safety requirement or large staffing decision depends on the result.

Limitations and assumptions

This model assumes that reviewers have comparable skills, that work can be represented by an average seconds-per-review input, and that the peak window can be described by a normalized multiplier. It does not include paid breaks as live seat capacity; shrinkage handles them as paid but unavailable time. It also excludes model/API spend, implementation, supervision, QA management, facilities, taxes beyond the chosen loaded rate, overtime premiums, vendor minimums and the cost of incidents or rework.

All dollar outputs in the worked example are estimates derived from stated inputs. The only external dollar benchmark is the BLS June 2026 average cited above. BLS data is not a staffing quote, and NIST guidance does not establish a mandatory staffing formula. Replace every assumption with your own logs, schedule, employment terms and risk policy before approving a hiring or outsourcing decision.