A familiar sourcing spreadsheet puts an employee wage in one column and a vendor's hourly rate in the next. The lower number wins. The problem is that those prices describe different things.
An employee wage excludes benefits and usually excludes recruiting, training, supervision, quality assurance, workforce management, technology, telecom, facilities, absenteeism, and attrition. A vendor rate might include some of those layers, exclude others, or apply minimum volumes and setup charges somewhere else in the proposal. Until the units match, the apparent difference is arithmetic, not analysis.
Define the unit before comparing the price.
Start by choosing the decision unit. It could be cost per paid hour, cost per productive hour, fully loaded cost per staffed seat-hour, or cost per resolved interaction. Each can be useful, but they cannot be mixed.
Paid hours include time that might not be available for customer work. Productive hours account for leave, breaks, training, meetings, coaching, and other shrinkage. A seat-hour adds the people, systems, and management required to keep that position operating. Cost per resolved interaction also incorporates workload and the contacts actually resolved.
The right unit depends on the decision. A workforce plan might need productive hours. A procurement comparison might need a loaded seat-hour. An automation proposal might be clearer at cost per resolved interaction because its escalation rate determines how much human work remains.
Build the in-house side from auditable layers.
Public wage data provides a starting point, not a complete contact center cost.
The U.S. Bureau of Labor Statistics reported 2,595,750 customer service representative jobs in May 2025, with a mean hourly wage of $22.40 and a median hourly wage of $21.53.
A separate BLS compensation table for March 2026 reported $36.42 in total compensation per hour worked for private-industry office and administrative support occupations: $25.00 in wages and $11.41 in benefits.
Those sources measure different but related populations, so they should not be presented as a single official benchmark. They can, however, illustrate a transparent bridge. Using the cited BLS inputs, the $11.41 benefit figure is an author-calculated 45.6 percent of the $25 wage figure. Applying that relationship to the $22.40 customer service mean produces an author-calculated $32.63 in wages plus benefits per hour worked.
That is still not a fully loaded seat cost. Buyers should add their own documented costs for recruiting, onboarding, training, supervision, quality assurance, workforce management, software, telephony, equipment, facilities, security, compliance, overtime, and replacement hiring. If a layer is unknown, label it unknown. An unsupported average creates the appearance of precision while hiding the real uncertainty.
Normalize vendor and automation proposals the same way.
Vendor proposals need the same treatment. Record what the hourly rate includes, then add setup fees, minimum commitments, management charges, platform licenses, telecom, overages, currency effects, and separately billed quality or reporting work. Convert every charge to the same period and workload assumption used for the in-house model.
Automation also needs a complete denominator. A per-interaction software price is not the final cost if a material share of contacts escalates to people. Include implementation, maintenance, monitoring, quality review, compliance oversight, escalation labor, and the cost of repeat or unresolved contacts. The goal is not to make any option look expensive. It is to stop one option from receiving a complete cost while another receives only its most attractive line item.
Put quality inside the financial model.
The least expensive operation is not economical if it creates avoidable repeat contacts, transfers, complaints, or rework. Quality belongs in the model as a constraint not as a paragraph after the price comparison.
Define first-contact resolution with an explicit repeat-contact window. Read average handle time alongside resolution, because shorter calls can simply move work into the next contact. Read average speed of answer, service level, and abandonment together, since any one of them can look healthy while customers still wait or leave.
Quality scores also need a denominator. State which interactions are eligible for review, how the sample is selected, what counts as a critical error, and how reviewers are calibrated. For automation, track escalation and containment with the same care. For digital channels, include response time and resolution rather than reporting message volume alone. Pair customer satisfaction with its response rate and cohort so a small voluntary sample is not treated as the whole customer base.
The comparison should use the organization's existing baseline and the same definitions for every option. If a proposal uses different windows or excludes difficult contact types, normalize it before accepting the claimed improvement.
Before approving a build, outsource, or automate recommendation, ask the following five questions that make the decision defensible:
- Are all options expressed in the same cost unit and time period?
- Which cost layers are included, excluded, estimated, or still unknown?
- Are workload, operating hours, seasonality, channels, and service scope comparable?
- Do all options use the same KPI definitions, denominators, and measurement windows?
- Who owns remediation, and its cost, when a service or quality guardrail misses?
This method does not predetermine whether an internal team, an outsourcing partner, or automation is the right answer. It makes the recommendation inspectable. When equivalent costs and equivalent outcomes sit in the same model, leaders can debate assumptions instead of defending incompatible prices.
Miki Furman is founder and CEO of Call Force Global.