Hire or deploy? The 2026 US math

The short answer: on BLS figures, benefits add roughly 46% on top of wages for full-time private-sector workers, so a role advertised at $50,000 costs an employer closer to $73,000 before recruiting, equipment or management time. An agent covers a subset of that role — the recurring, text-shaped, reversible part — at a fraction of the cost, while judgement stays with people. The useful framing is not replacement. It is that the automatable subset stops consuming a person's day, which changes what you hire for, and when.

Third-party figures on this page were checked on 26 September 2026 against the publishing organisation's own release. Research in this field ages quickly and definitions vary between surveys — treat any undated statistic with suspicion, including on our own pages.

What a US role actually costs

From the Bureau of Labor Statistics' Employer Costs for Employee Compensation release for June 2026:

• Private industry, all workers: $46.89 per hour worked. Wages and salaries are $32.82 of that (70.0%); benefits are $14.07 (30.0%).
• Full-time private industry workers: $54.00 per hour. Wages $36.97 (68.5%), benefits $17.03 (31.5%).

Read the full-time line the way a budget holder has to: benefits are $17.03 on top of $36.97 of wages, which is about 46% added to the wage bill. The "30%" framing is the share of total compensation and is the number people quote by mistake when they mean the markup.

For the roles usually in scope for this decision, the BLS Occupational Outlook Handbook put the median hourly wage for customer service representatives at $21.53 in May 2025. Apply the same markup and the employer cost is roughly $31 an hour before anything else.

Then add what the ECEC does not cover: recruiting, onboarding, equipment and software seats, the manager's time, and the cost of the role being vacant while you fill it.

What the agent side actually is

The comparison people reach for — salary versus subscription — is wrong in both directions.

It overstates the agent, because a subscription is not the whole cost. There is model usage, the setup work of connecting it to your systems, and someone's ongoing attention. We wrote that out in full in what an AI agent actually costs to run.

It also understates the agent, because a deployment is not a seat. One deployment can run several agents covering several jobs, and the marginal cost of the second job is much lower than the first. A person cannot be allocated that way.

The useful unit is neither. It is: what fraction of this role's hours are recurring, text-shaped and reversible? That fraction is what is actually on the table.

Working one case through

A coordinator role: enquiries, scheduling, chasing, and a weekly report. Say the recurring text work is 60% of the week — high, but that is what these roles look like before anyone touches them.

• Hire: the full role, including the 40% that needs judgement, relationships and the ability to notice something is wrong. Cost: the loaded figure above, plus the overheads it excludes. Ramp: weeks. Capacity: one person's hours, in one time zone.
• Deploy: the 60%, at a cost that does not scale with hours, available overnight and at weekends, with every output reviewable. The 40% does not get done. Ramp: days to set up, then a couple of weeks of reviewing everything before you widen the boundary.
• Both, which is what usually happens: hire later, into a role where the recurring work already runs, so the person spends their week on the 40%.

Where a person stays essential

• Judgement about people. Noticing a customer is upset in a way the words do not say. Knowing which escalation is real.
• Accountability. Someone has to own the outcome. An agent produces work; it does not carry responsibility, and pretending otherwise is how governance problems start.
• Novel situations. The thing that has never happened before is exactly the thing an agent handles worst and a competent person handles best.
• The relationship. In most B2B sales and in any account that matters, a person is the product.

Anyone selling you the opposite is selling you a problem eighteen months out.

The four numbers to put in your own version

1. The loaded cost of the role — wage plus roughly 46%, plus recruiting and equipment.
2. The recurring-text fraction of that role's week. Measure it for one week rather than guessing; the guess is always wrong and usually low.
3. The all-in agent cost for that job, including setup and someone's weekly review time.
4. The cost of the work not happening, which is the one everybody omits. An enquiry answered in nine hours instead of nine minutes has a conversion cost, and it is often larger than either of the first two figures.

That fourth number is why this decision is frequently not about savings at all. The comparison that matters is not agent-versus-person; it is agent-versus-nothing, because the work in question is work nobody currently has time to do.

If you are hiring anyway

Then the question is what they walk into. A role where the recurring layer already runs as documented procedures is a better job and a shorter ramp than one where it does not, and the documentation is a by-product of setting the agents up. What to automate before you hire is the sequencing version of this.

This is a framework, not financial advice, and the BLS figures are national averages across all private industry — your region, sector and role will differ. Use your own numbers.