What a headcount costs

Thirty percent of what an employee costs is not wages

The Bureau of Labor Statistics measures employer cost directly, per hour worked, in the Employer Costs for Employee Compensation series. For March 2026, private industry employers spent $46.60 an hour on total compensation: $32.60 in wages and salaries and $14.01 in benefits. Wages were 69.9 percent of the total. Benefits were 30.1 percent.

Put the other way around, every dollar of wage carries about forty-three cents of non-wage employer cost. That is not an estimate of what a generous employer spends. It is what employers spent, measured.

Two of those rows are neither optional nor negotiable. Legally required benefits came to $3.38 an hour — Social Security and Medicare at $2.78 and workers' compensation at $0.42 sit inside that figure. The employer pays them because the employer is an employer. No arrangement of the org chart removes them.

Paid leave is the other one worth naming, at $3.54 an hour. That figure is the direct cost of hours paid and not worked. It is not the cost of covering those hours. Cover is a separate problem, and it is not in this table.

The wage is a band, and where you hire in it is a quality decision

For customer service representatives — occupation 43-4051 in the federal classification — the May 2025 national estimates put employment at 2,595,750, the mean hourly wage at $22.40, and the annual mean wage at $46,590. The median was $44,770. Those come from the wage survey, which counts and dates differently from the Occupational Outlook Handbook’s projection base — two measurements of one occupation, not a discrepancy.

The band around the middle matters more than the middle. The tenth percentile was $31,750 a year and the ninetieth was $63,590. One job title spans a factor of two. Where a brand hires inside that band is a decision about who answers its customers, and it is made once, at the offer.

Carry the benefit ratio across the mean and one seat costs an employer roughly $66,600 a year before recruiting, before supervision, before the empty weeks. The ratio is a private industry average rather than an occupation-specific figure, so substitute your own if payroll can produce one. The shape holds either way: the wage line is about seventy percent of the seat.

Supervision is a line of its own. First-line supervisors of office and administrative support workers — the occupation whose O*NET sample of reported job titles includes customer service manager and customer service supervisor — had an annual mean wage of $73,490 in May 2025. How many representatives one supervisor covers is your number, not a published one. Divide, add the benefit ratio, and the result belongs on the cost of every seat underneath.

Turnover is the line item that never gets a line

The Job Openings and Labor Turnover Survey counts separations — quits, layoffs, discharges, everything that ends an employment relationship. In retail trade across 2025, total separations averaged 3.8 percent of employment per month. The all-industry average was 3.3 percent.

The annual level is the number that lands. Retail trade recorded 6,987,000 separations in 2025 against average employment of 15,441,100. That is forty-five separations for every hundred jobs, in one year.

This is the point where most cost models reach for a figure saying that replacing an employee costs some percentage of their salary. That figure circulates without an author. It is repeated between articles that cite each other and it does not trace back to a published methodology, so it does not appear here.

The components can be built instead, and each one has a price the operation already knows:

• The vacancy. The queue does not pause for a notice period. The work is absorbed by the people who stayed, at their hourly cost, or it is not absorbed at all and the response time carries it.
• The search. Screening and interviewing consume supervisor hours, at the supervisor cost already computed above, plus whatever the job posting and the agency invoice.
• The ramp. A new representative is paid in full from week one and produces the output of a new representative. The gap is a real cost with a start date and no end date on the calendar.

None of that requires a study. Each is a number the operation has, and the sum is specific to the operation rather than borrowed from somebody else's.

One person does not cover a week

Coverage is where a headcount model stops being linear. A seat staffed nine to five, Monday through Friday, covers 40 of the 168 hours in a week. Every hour outside that window is either unanswered or a second seat.

Paid leave makes the same point inside the 40. The $3.54 an hour in the table is what employers pay for hours not worked; somebody still answers during those hours. In a team of three, one person on leave is a third of the coverage, and small teams take that arithmetic hardest — the smaller the team, the larger each absence is as a fraction of it.

Seasonality compounds it. Ticket volume moves with launches, promotions and the fourth quarter. Headcount moves with notice periods and start dates. The two curves do not have the same shape, and the gap between them is staffed either with overtime or with idle capacity, both of which are paid for.

Every hire restarts the learning curve

The last cost has no invoice. A representative gets better at a specific brand over months — which SKU fails in which way, which customer segment escalates, which sentence settles a complaint and which one enlarges it. That knowledge lives in the person, and separations remove it at the rate the survey measures.

The quality cost is the distance between the answer an experienced person gives and the answer their replacement gives while learning. It shows up as refunds granted that policy did not require, as escalations that did not need to escalate, and as customers who do not come back and do not say why. None of those three lands in a payroll system.

This is the part of the cost structure that architecture addresses rather than trims. Razviel is designed so the record does not sit inside a person: every ticket is built to carry an end-to-end record — context, the policy cell consulted, the decision, the execution, the outcome — and the design accumulates curated patterns per brand instead of in an individual's memory. Raw tickets, transcripts, attachments and customer identifiers stay siloed to the brand that originated them, permanently.

Whether that changes the number for a given brand is not something measured yet, and writing otherwise would be a claim with no measurement behind it.

The takeaway

The wage is about seventy percent of what a support seat costs an employer, before supervision. The rest is employer cost the federal series measure directly, the supervisor above the seat, the hours one seat does not cover, and the turnover that restarts the ramp at the rate the survey reports. Each has a published input or a number the operation already holds, so the full cost of in-house support is computable rather than estimated.

The method, line by line. Every published figure can be replaced with your own:

1. One seat: wage × 1.43. The wage you intend to pay, times the employer cost ratio from the BLS series, where benefits were 30.1 percent of compensation. At the occupation’s mean wage of $46,590, that is about $66,600.
2. Plus supervision ÷ span of control. The supervisor’s wage — a mean of $73,490 for the occupation — times the same ratio, divided by the number of representatives one supervisor covers. The divisor is yours.
3. Times the seats the week needs. One seat staffed nine to five, Monday through Friday, covers 40 of the 168 hours. Every hour you answer outside that window, and every hour of paid leave, is another seat or an unanswered hour.
4. Plus turnover × the cost of one departure. Seats times the share that separates in a year — forty-five per hundred jobs in retail trade in 2025 — times your own vacancy, search and ramp cost for one departure.
5. And the line with no invoice. The distance between an experienced answer and a learning one — refunds policy did not require, escalations that did not need to escalate — has no published coefficient. It stays a named line, unpriced, rather than a borrowed percentage.

The sum is the cost of in-house support: larger than the salary, and defensible, because every input is either published by a federal statistical agency or produced by your own payroll.