The Math That Maths: Role-Based Agents and the New Cost of a Productive Hour

Most strategic conversations about AI still start with the model and provider. If you want to unlock capacity, start with the seat. A seat is a role your company already budgets for, with a job description, a reporting line, a manager and a line on the P&L. A role-based agent fills that seat the way a remote employee does: in email, in Slack, in meetings, and in the files and documents the team works from.

This is only the third time in modern business that every company can raise capacity and lower cost at the same time. The industrial age did it with machines on the floor. The software age did it with systems in the back office. The meaningful economics of the AI age will center on role-based AI agents in seats in your company.

The test is one number: cost per productive hour. Divide what a seat costs in a year by the hours of productive work it delivers. Do it for the person in the seat and for the agent in the same seat, and compare. It’s just math.

For the person, cost is salary times your loaded multiplier, plus the annual cost of turnover. Productive hours are the 2,000 hours of a working year, less the hours lost to vacancy and ramp.

For the agent, cost is its annual run cost. Hours can be counted three ways:

Here is the rubric on an illustrative $70,000 seat, with the measured view set at the average on-duty hours of our own three agent partners (4,832 a year):

The agent wins at the floor by about 6x. It wins by about 14x on measured hours and about 26x at the ceiling. The math holds in every view. Plug in your own salary, multiplier and run cost, and it holds for you. The rest of this piece shows where each number comes from.

The cost of a seat

Take the $70,000 seat. It's a round number so the arithmetic is easy to follow.

Leaders budget a seat as salary plus benefits plus taxes, usually salary times 1.3. That's conservative. The Bureau of Labor Statistics put benefits, including legally required ones, at 30.0% of private employer compensation in June 2026, which implies a multiplier of about 1.43 on wages. At 1.3, the seat costs $91,000 a year.

Then there's the cost of filling the seat again. Everyone in a role was recruited, onboarded and trained before reaching full speed. Then people get promoted, move or leave. That's healthy, but each move restarts the clock.

The public data on that clock is consistent. Median tenure with a current employer was 4.1 years in January 2026 (BLS). SHRM's 2026 benchmarking puts the median time to fill a nonexecutive role at 39 calendar days. A Mellon learning-curve study, cited in MIT Sloan Management Review, found professional hires took about 20 weeks to reach full productivity. The Work Institute puts the all-in cost of turnover at 33.3% of base salary.

Spread over tenure, turnover adds about $8,300 to $8,800 a year. The seat costs about $100,000 a year, or $54 per productive hour.

Now the agent. In our deployments, a role-based agent in a seat like this runs roughly 80% less than the loaded cost which puts the agent at about $16,800 to $18,200 a year. An agent that doesn't resign, it doesn't need a second 20-week ramp, and it doesn't take the role's knowledge with it.

The agent's hours

A salary buys a person on duty: present, reachable and able to act for a defined block of time. The fair comparison for an agent is the same thing. How many hours is it on duty, checking in and able to act on what comes in?

For our agents, that's a setting, not a guess. Each one runs on a schedule, and the scheduler records it.

Maker checks its inbox every hour, from 6:04 AM to 10:04 PM Eastern, every day of the week. That's 17 hours a day, or 6,205 on-duty hours a year.

Baker runs a Slack pulse every 30 minutes through the weekday, hourly on weekends, and at the edges of every day from 6 AM to 10 PM. That's 16 hours a day, or 5,840 on-duty hours. The scheduler shows the pulse running this week, every half hour.

Abbie runs working routines from a 3 AM memory pass to a 6 PM message check, plus a channel health monitor every 30 minutes around the clock. Counting only the clock hours that hold a working routine, she's on duty 2,451 hours a year.

Together the three partners are on duty 14,496 hours a year. That's 2.4 times the 6,000 hours three full-time people would give you.

Our fleet costs $37,135 a year: the whole meter, at list price. Here are the three views at that cost:

How much of an on-duty hour is the agent computing? Very little, because the work takes minutes. Abbie's logged runs over the past seven days total 2.9 hours of compute. Baker's pulse logged 95 runs in the same week at about a minute and a half each, roughly 2.4 hours. Nobody pays a person for the minutes their hands are on the keyboard either. You pay for the hours they're on duty and the work that comes out.

This morning I asked for a revision to a full article, a LinkedIn post, an X post and its reply at 8:54 AM. The revised set was checked and logged at 8:55 AM.

The sub-agent multiplier

Each first-class partner leads its own fleet of narrow sub-agents: seven in Abbie's A-Fleet, four in Baker's B-Fleet and three in Maker's M-Fleet. That's 14 sub-agents and 17 agents in all. The sub-agents carry the narrow, repeatable work, such as a weekly CRM sync, deployment health, model checks, claim audits and agreement reviews. The partner deploys them, briefs them and manages them, the way a team lead runs a pod.

At the ceiling, 17 agents is 148,920 hours a year, or about 25 cents an hour at $37,135. At the floor, it's 34,000 hours at $1.09 an hour.

Getting role-based agents ready

None of this works out of the box. It needs readiness, and most companies overestimate what that takes.

Nobody needs an agent that knows everything. You need an agent that knows everything relevant to your company, your products, your services, your clients and your goals, locked into the role it holds. Most of that already exists in your back office: training materials, role descriptions, employee handbooks, operating procedures, reports. The cost of readiness is assembly, not invention.

Here's the leverage. Anthropic, OpenAI, xAI and Google have together put hundreds of billions of dollars into frontier models and the compute behind them. You don't need to match that. With a modest investment in assembling your company's context and standing up a place to manage the agents, any company can co-opt that investment and build what I call a company language model of its own.

Two pieces of work make it real. The first is the context layer. Your existing documents, assembled and staged, become a persistent, versioned layer every agent reads from and writes to, including what the agents learn when a person overrides them.

The second is the management space. Think of it as Workday for role-based agents. Ours is the Agent Management Center. It gives a department manager one place to see, coach and manage the agents that report to them. Every agent is visible (its brief, permissions and guardrails), auditable (its history of changes and actions) and stoppable by a named owner.

Then the agent gets the access points a new remote employee would get. At a digital marketing agency, that meant email, Slack, Zoom and Drive, with agents mapped to four existing seats under human leads. We run the same pattern for a senior living advisory startup and a newly launched restaurant concept, where the owner works with the agent in iMessage.

Trust builds one rung at a time. We use a Trust Ladder: draft and wait, then act and report, then act within set boundaries, then full delegation with periodic audits. At the agency, no agent wrote to the system of record on its own. Writes staged in a review queue, and an operator approved them.

Readiness is assembly, staging and a few weeks of a manager's attention. Price it from your own documents.

Funded from a line you already have

None of this needs new money.

For almost every business personnel and staff costs are the largest line on the P&L. Whatever the share is for you, the line already exists, it's already approved, and it recurs every year.

So the question isn't where you'll find new dollars for AI. It's what portion of the dollars you already spend on seats you'll convert to a new employee type. When a seat opens, you can determine if that role can be refilled with a person or fill it with a role-based agent and redeploy most of the budget. Additionally, when a team is stretched, you can add capacity without adding the historical fully loaded burden of an FTE.

That changes the internal conversation. An AI budget request competes with every other new project. A seat conversion is a decision about money that's already in the plan, and the math is the rubric above, at your own salary and your own multiplier.

The math only gets better

The math maths at today's prices, and that alone should be compelling enough to at least go interrogate it and test it in your own business. It also only gets better, for three reasons.

First, the market cost of tokens keeps falling. The venture firm a16z found that for a model of equivalent performance, cost has been falling about 10x every year. The cheapest model at GPT-3's level fell from $60 to $0.06 per million tokens in three years. Epoch AI found the price to reach a fixed level of performance falling between 9x and 900x a year, depending on the task.

Second, the models keep getting better. The same dollar buys an agent that can hold a more complex seat each year, and the share of roles an agent can hold grows with it.

Third, more of the agent work will be able to run on local models, on hardware you own, inside your own walls. That's where I believe this is heading for a material portion of the tasks a role base agent will be expected to complete. That helps on cost, because routine work isn't billed by the token. It helps with security, because sensitive data doesn't have to leave the building.

This is a blinding glimpse at the potential, with math that maths.

Snap it to any business

The math is the same at 50 people and at 50,000. Only the inputs change.

For loaded cost, I use the BLS's June 2026 data on employer compensation by establishment size, annualized at 2,000 hours. Total compensation averaged $41.03 an hour at establishments with 50 to 99 workers, $47.96 at 100 to 499, and $67.68 at 500 or more. That's about $82,000, $96,000 and $135,000 a year per worker. The implied benefits multiplier rises with size, from about 1.38 to 1.55. The tiers, the 10% to 20% share of roles and the 80% are illustrative assumptions.

Change any input. Use 50% savings instead of 80% and 5% of roles instead of 10%, and the small business still frees about $80,000 a year, and the enterprise more than $13 million. The mechanics don't care about your size.

"Freed" doesn't have to mean cut. Take it as savings, in-market pricing leverage, or take it as expanded capacity.

A PE lens

I’ve spent a good deal of my career working with PE firms as an operator inside their portfolio. Private equity partners think in portfolios, so think about this one.

Take an illustrative portfolio of 10 companies with 500 employees each, at the mid-market loaded cost of $95,920 per seat. Convert 10% of roles. That's 500 seats, $48.0 million of payroll converted, and about $38.4 million a year freed at 80%.

Take it as cost and it's a run-rate operating-cost reduction that flows to EBITDA. Multiply by the exit multiple you underwrite and you have the value effect. Take it as capacity and it's 500 seats of new output across the portfolio without 500 requisitions, each on duty well beyond 2,000 hours.

Most of the work repeats across the portfolio, which is what makes it a portfolio play. Assess the roles at diligence, so the seat map is part of the investment thesis. Run readiness in the first 100 days. Stand up the Agent Management Center once as a shared service. Carry the Trust Ladder and the review discipline from company to company, so each deployment is faster than the last.

Change management with familiar constructs

Treat this as a technology project and it stalls. Treat it as hiring and it lands.

Familiar structures shorten the change curve: onboarding, org design and reporting lines. An agent gets a job description that doubles as its brief. It reports to a human lead, the way a new hire does. It earns trust one rung at a time. When people can see where the agent sits on the org chart, the dip before the climb gets shorter.

It needs owners at the top.

And people do move up, from execution to oversight, judgment and relationships. The work that was always too expensive to staff becomes possible: the new segment, the new offer, the faster response, more coverage. The point was never a smaller company. It's a company with more capacity than its competitors and the cash to use it.

Is this role ready for a role-based agent? A checklist

The Obvious

Every CEO, CFO and COO should be assessing open roles now to find which can be role-based agents, then adding capacity to existing teams the same way. Run the rubric on one real role. Pull your own loaded costs and your own share of roles into the template. Then test it with one seat and one manager.

The firms that get this going will change the force they bring to market, and they'll be able to pull levers their competitors don't have, to take share. The math works at the floor. It works better on the hours the agents are actually on duty. And it gets better every year.

So think about what that capacity is worth to you. What would you do with tens of thousands of hours of new capacity, at $2.56 an on-duty hour against the $75 to $100 an hour it would take to staff it with people? Clear out the backlog? Put it behind your pricing to take share? In my career, I have never found a company that complains about too much capacity. It's always "my team" or "my department" that needs more resources, and it's always an exercise in priority management. Role-based agents change that exercise.

You don't want to be the firm that slow-rolls adding role based agents to your team.