# Your AI Employee Has a Rate Card. You Set It.

Published: 2026-08-03T13:00:00.000Z · Updated: 2026-08-06T12:39:22.000Z · Author: Rod Amora · Canonical URL: https://rodamora.com/blog/your-ai-employee-has-a-rate-card-you-set-it

> A conversation with an agent costs me about eight cents at the median. Before that, the same work cost twenty times more, and the difference was a setting I picked. Agents give a service firm its first real cost per unit of work, and that cost is a dial, not a market price.

A conversation with an agent in my production system costs about eight cents. That is the middle cost across 2,000+ agent tasks in July 2026. The average is closer to seventeen cents, and the most expensive task cost $16.12.

Before that, the same work cost twenty times more. I was running a model that was overkill for the tasks, and I had days to fix it before the system stopped paying for itself.

The task did not set the bill. Settings did. Most firms running agents have never looked at them.

Putting an agent on the delivery team gives a service firm its first real cost for each piece of work. Unlike payroll, that cost is a dial you set: which model runs the task, how hard it thinks, how well you cache, and where it runs. Turn the dial and AI reaches your profit-and-loss statement. Leave it alone and you have bought a system that never pays for itself.

## Your firm now pays for each AI task

For service businesses, each piece of delivered work has always carried the cost of a person's time.

Payroll works differently. You pay the same salary in a slow month and a busy one, and the tenth report your team ships costs you what the first one did. That is how service firms have worked for as long as they have existed.

An agent changes that pattern. Every task carries a price that moves with volume. Run a new batch of tasks, and the bill grows with it. That price sits beside salaries in your cost to deliver.

Nobody mentions this when you [staff an AI employee](https://rodamora.com/blog/what-is-an-ai-employee). The demo shows the output. The bill arrives later, and your profit-and-loss statement has no place to put it.

## What does an AI agent actually cost per task?

Most tasks cost less than operators fear, but hard tasks cost much more.

Across those 2,000+ tasks, the average cost per conversation is $0.17 to $0.18 and the middle cost is $0.08 to $0.09. Over half of what real users ask lands between two and ten cents.

The expensive end of the list matters too. A daily collections nudge costs a fraction of a cent, while a customer service summary runs about eight cents. A full project analysis costs fourteen, and a monthly sales review costs $1.57. One project run hit $16.12, more than ten thousand times what the automated monitor loops cost as they run in the background at about a tenth of a cent each.

![Table of four agent tasks and their cost per run, from about $0.001 for a monitor loop to $16.12 for a project run.](https://rodamora-uploads-347628392068-sa-east-1-an.s3.sa-east-1.amazonaws.com/illustrations/2026-08-02-agents-as-billable-team-members/fig-diagram-1.jpg "The cost tail sizes your cap")

The answer changes with what you count. Counting only active user conversations, the middle cost is about eight cents. Counting everything the system does, including those cheap monitors, it drops to around three cents. The same month and platform produce numbers three times apart, and neither is wrong.

If you cannot tie a per-task cost to a clear list of tasks, you cannot defend the number. It is the same discipline problem as every other number in [the four numbers](https://rodamora.com/four-numbers): the metric is useless until you say exactly what it counts.

One caveat I'd rather give you than have you catch. These numbers are low partly because the system caches aggressively and I optimize for it, and cost isn't proportional to tokens. Across my own examples the effective rate swings from about three cents per million tokens to thirteen. Run agents without caching and your numbers will be higher than mine. The difference comes from how the system is built, not from the work itself.

## The cost is a setting, not a fact

You set the cost with four dials. Almost nobody turns them.

**Which model you use.** This is the dial behind my twenty-times drop. I started on a much more capable model than the tasks needed, watched the numbers, and moved off it within days. The work stayed the same, and so did output quality. The bill fell to one twentieth.

**How hard it thinks.** Most models now have an effort or reasoning setting. The same model and task can cost different amounts depending on how long you let it chew.

**How well you cache.** Repeated context is the cheapest thing you can send. Structuring prompts to reuse it is a design choice, and it explains most of the difference between my numbers and someone else's.

**Where it runs.** Open-source models on your own hardware have a different cost structure from a cloud API. That dial matters at volume. For a firm spending a few hundred dollars a month, it is almost certainly not your first move.

You cannot pay a senior analyst twenty times less for the same output. People have no effort setting, and there is nowhere cheaper to run them. An agent gives you all four dials, and most firms leave them untouched because nobody told them they exist.

## Use the cheapest model that can do the job

Different tasks need different models. A cheap model can fetch data from a web page. A stronger model can proofread a client report and flag inconsistencies.

Use the cheapest model that can reliably clear the task.

Pulling structured data off a supplier's site is a job for the cheapest model that reliably does it. It either got the numbers or it didn't, and you can check that automatically. Reading a client-facing report to catch the one inconsistency that would embarrass you in a meeting earns the expensive model, because the cost of missing it is the client relationship.

When a cheap model gives bad results, many teams upgrade the whole pipeline. Put a stronger model in as a validator, checking and guiding the weaker one.

This gives better output than the cheap model alone and a smaller bill than running the expensive model for every step. The validator lets you step the primary model down one or two levels, and it runs only on the output rather than on every step of the work.

![Three blocks left to right: cheap model does every step, stronger model checks the output, better output and smaller bill.](https://rodamora-uploads-347628392068-sa-east-1-an.s3.sa-east-1.amazonaws.com/illustrations/2026-08-02-agents-as-billable-team-members/fig-diagram-2.jpg "The validator pattern")

The people who build these systems for a living already work this way. ICONIQ's survey of around 300 executives building AI products found companies [routing the majority of workloads to smaller or fine-tuned models, escalating only high-complexity tasks to frontier models](https://www.iconiq.com/growth/reports/2026-state-of-ai-bi-annual-snapshot), tied directly to margin outcomes.

The human version is the [named validation gate](https://rodamora.com/blog/delegate-the-inputs-own-the-outputs). The machine version is the fourth element of an AI employee: the validation method nobody sells you in the box. The same checking step can improve output and lower the bill when it lets you use a cheaper primary model.

## How should a service firm price against this?

Price carefully. The AI-services crowd's usual model leaves the risk with you.

Outcome pricing is having a moment. Intercom's Fin charges [$0.99 per outcome](https://fin.ai/pricing) and bills nothing when the agent hands off without resolving. Zendesk [prices its AI agents per automated resolution](https://www.zendesk.com/pricing/), charging only for requests settled without a human. Salesforce sells actions at [ten cents each](https://www.salesforce.com/news/press-releases/2025/05/15/agentforce-flexible-pricing-news/). Those companies promise a fixed price while their cost changes because millions of tickets spread the risk across the workload. A firm doing $3 million in revenue cannot spread that risk the same way. Sell a fixed price while your own cost can keep rising, and you are taking on a risk you have not priced.

Use the structure behind their slogan. Every serious vendor runs two layers: a stable price the client budgets against, sitting on a cost that grows with use but the client never sees. Anthropic's enterprise pricing is the cleanest public version, where a flat seat fee buys access and [every token bills on top](https://www.finout.io/blog/claude-pricing-in-2026-for-individuals-organizations-and-developers). Your retainer is already the top layer, and most firms are missing the bottom one.

Use the middle cost for your budget and the worst task for your cap. The middle cost tells you what a normal month costs. The worst task tells you what a bad month could cost, and your cap has to survive it.

This does not change my view on hourly billing. I've [argued that elsewhere](https://rodamora.com/blog/how-ai-is-changing-the-consulting-industry-and-why-the-retainer-wins) and the answer hasn't changed. This is the cost side underneath it.

What this changes is the top layer. Price is one of [the four numbers](https://rodamora.com/four-numbers), and AI never changes it by itself. Your cost to deliver just dropped whether you chose that or not. Price changes only when you choose it, and you have a window before clients notice. Move in that window and you keep the savings. Wait, and the market turns them into a fee cut you never approved.

Charging more for the same hour is a rate increase, and it runs out. The move is to change the unit. Sell a finished outcome, a capacity block, or a subscription to a result. Then your price stops tracking labor. An agent with a rate card only pays you back if the thing you sell on top of it has one too.

## Expect the cost to change

I don't expect eight cents to hold, and neither should you.

My costs are low partly because the system is young. They climb as it matures, as harder tasks get added, and as people get better at pulling real work out of agents. That $16.12 task was not a runaway loop.

It was someone who had learned to ask for something hard. People who use agents heavily pull the average up. The better your team gets, the more heavy users you will have.

The survey points the same way. ICONIQ's [2026 State of AI snapshot](https://www.iconiq.com/growth/reports/2026-state-of-ai-bi-annual-snapshot), built on survey data from around 300 executives building AI products, found that as adoption matures companies spend a higher percentage of revenue on AI, not a lower one.

This stat has a boundary. That survey covers companies where inference is the product. Bessemer's finding that [AI companies see 50-60% gross margins against 80-90% for SaaS](https://www.bvp.com/atlas/the-ai-pricing-and-monetization-playbook) describes those same companies. A service firm spending $175 a month against a $5,000 retainer is not in that story. Your AI delivery cost is around three and a half percent of revenue, and it is not coming for your margin.

![Three figures side by side: $175 a month in agent spend, a $5,000 monthly retainer, and 3.5% of revenue as AI COGS.](https://rodamora-uploads-347628392068-sa-east-1-an.s3.sa-east-1.amazonaws.com/illustrations/2026-08-02-agents-as-billable-team-members/fig-diagram-3.jpg "Agent spend against a service firm's revenue")

Cheaper and more capable models ship constantly. Over a longer horizon, I expect cost per unit of work to flatten or fall even as usage climbs. Usage can rise while the cost of each piece of work falls, which is why this needs an owner rather than a setup.

## Somebody has to own the dial

I expect some firms to create roles for this. Call it an AI system cost optimizer or call it nothing. Someone still has to own which model, effort setting, caching, and hosting the firm uses, just as someone already owns how much of the team's time gets used.

At your size, that is not a hire. Give it to someone already on staff, with a thirty-minute monthly review of what the agents cost and which tasks changed. The first time you run it, you will find something like my twenty-times problem, because almost everybody has one.

This is the [Stage 2 to Stage 3](https://rodamora.com/delivery-model-ladder) problem in a specific form. Revenue stops tracking headcount, which is the whole point of getting there. The automated work still has a bill of its own, and it is the first cost in your firm's history that responds to being tuned. It is also a role that appears rather than [disappears](https://rodamora.com/blog/fewer-managers-more-managing).

## Start here

Start with last month's agent spend. Split it by task type. Find the middle cost and the worst single task.

Most firms have never done this, so they are running whatever model they picked in week one, at whatever effort setting the tutorial suggested. That was my situation too, and it cost twenty times what it needed to.

For work agents suit, a well-tuned agent costs less than a labor hour, and I don't expect that to reverse. The tuning is the job.

## Questions operators ask

**What does it actually cost to run an AI agent per month?** It depends far more on how you configure it than on how much you use it. In my system, the typical conversation costs about eight cents and the average is closer to seventeen, because a few heavy tasks pull the average up. Budget for the average: a thousand tasks a month costs under $200. The same workload on a badly chosen model was twenty times that.

**How do I stop token costs from eating the retainer?** Set a cap, and pick the model per task rather than one model for everything. Set a hard spend ceiling per client engagement, budget the retainer from your typical task cost, and size the ceiling from your worst observed task. A cheap model with a stronger validator checking its output beats a single expensive model on both quality and cost for most work.

**Should the agent be a line item on the invoice?** For most firms, no. The client is buying a delivered outcome, and a line item invites a conversation about your cost structure instead of your value. Track it internally, watch it monthly, and price it into the fee.