Answers
What Does It Cost to Run an AI Agent?
An AI agent has three cost lines: the one-time build, the model API usage, and the cloud infrastructure it runs on. The build is a known number you agree up front. Infrastructure is usually the smallest of the three. Model usage is the line that varies and the one worth designing around, because the unit of cost is a token rather than a user or a task, which means two jobs that look identical on an org chart can differ by an order of magnitude in what they cost to run.
The short answer
Budget three lines. A one-time build, agreed in advance. Model API usage, billed by the token and variable with how much the agent reads and reasons. Cloud infrastructure, billed by your provider and usually the smallest line. Any ongoing engineering time is a fourth line you choose whether to buy.
Everything below explains what moves the variable line, what makes agents expensive by accident, and how the arithmetic compares with the per-seat pricing most teams are used to.
Line one: the build
This is the part you can fix in advance, and it should be a number rather than a range. On Sovereign, deployment starts at $10,000 for a single specialised agent with up to five connected systems, and $25,000 for the full tier with up to four agents, an orchestration layer between them and 60 hours of pre-deployment skill work included. Those figures and what each tier includes are on the pricing page.
What the build buys is not the model. The model is a commodity you rent by the call. The build is the harness around it: the skills that encode your procedures, the connections to your systems, the approval gates, the memory design and the logging. That is where the work is and where the difference between an agent that holds up and a demo that does not actually lives.
Line two: model API usage, the line that varies
Model usage is billed by the token, which is roughly a word fragment, counted on both the input and the output of every call. That single fact explains most of what surprises people about agent cost.
A task does not cost what the output is worth. It costs what the agent had to read in order to produce it. An agent that answers a one line question after reading a forty message email thread, three files and a procedure document has spent most of its tokens before it writes a word. Two tasks that both look like "reply to this message" can sit an order of magnitude apart for exactly that reason.
Three design choices move this line more than anything else:
- Context size per run. Give each run the material it needs and nothing else. Loading a whole knowledge base into every run because it is easier than deciding what matters is the most common cause of a bill nobody can explain.
- Model choice per job. Use the most capable model where judgment is the deliverable and a cheaper one for mechanical work. Running the top model on file moves and status checks is pure waste, and it is the most common form of it.
- Schedule frequency. An agent that wakes every two minutes to check for work that arrives twice a day is paying for nothing most of the time. Match the poll to the arrival rate of the work, or trigger on the event instead of polling at all.
The related question of whether the job needs an agent at all is worth settling before any of this, because a deterministic pipeline with one model step costs a fraction of an agent and is more reliable. The flowchart test settles it in about a minute.
Line three: infrastructure
The compute, storage and networking the agent runs on. In a Sovereign deployment this runs in your own cloud account and is billed to you directly by AWS, GCP or Azure, with no markup from us, because you own the account. For a single agent doing operational work this is typically the smallest of the three lines. It grows if you add heavy data processing, long retention or high availability requirements, which are scoping decisions you make deliberately rather than surprises.
How does this compare to a per-seat AI subscription?
The two price different things, which is why the comparison flips depending on your shape.
A per-seat subscription prices access. You pay per person whether or not that person uses it, and the bill scales with headcount. It is a good deal when many people each get modest value from a tool they use occasionally.
An owned deployment prices work done. You pay for the build once and then for the tokens the work consumes, and the bill scales with how much the agent actually does. It is a good deal when a small number of people need a large amount of automated work, which is the usual shape of an operations problem.
There is a second difference that does not show up in the monthly number. A subscription is rented capability; when it ends you keep nothing. An owned deployment stays in your account with its memory, its logs and its skills. What that is worth depends on how much procedure you have encoded, which is covered in migrating off a SaaS AI platform.
How to keep the running cost predictable
- Log cost per run. If you cannot see what each run costs, a behaviour change shows up as an invoice instead of as a signal. This is the single highest value thing to build first.
- Cap tokens per run. A hard ceiling turns a runaway loop into a failed job you get told about, which is the cheap failure mode.
- Route by job, not by habit. Decide per task class which model runs it, and make the cheap model the default rather than the exception.
- Keep deterministic work deterministic. If a step is a lookup, a format change or a file move, it should be code. A model call that does something a script could do is a recurring charge for nothing.
- Review monthly. Agent cost drifts as the work grows. A monthly read of cost per run by job class catches the drift while it is still small.
Those five are the same controls we run on our own deployments, which is where this page comes from. The full picture of what a deployment includes is on Sovereign Agent, and the security and isolation side is in the security baseline.
Frequently asked
Frequently asked questions
What does it cost to run an AI agent every month?
Two recurring lines you pay your providers directly, model API usage and cloud infrastructure, plus any retainer you buy for engineering time. The model usage line is the one that varies, because it moves with how many tasks the agent runs and how much reasoning each task needs.
Why is agent cost so hard to quote up front?
Because the unit of cost is a token, not a user or a task, and the number of tokens a job needs depends on how much context the agent has to read before it can act. Two jobs that look identical on an org chart can differ by an order of magnitude in tokens if one of them has to read a long thread first.
What makes an agent expensive to run?
Carrying more context than the job needs, running the most capable model on mechanical steps, and polling on a schedule that fires more often than the work arrives. Those three account for most runaway bills we have seen, and all three are configuration rather than anything inherent to the agent.
How do I keep agent cost predictable?
Route each job to the cheapest model that can do it, give each run only the context it needs, cap the tokens a single run may spend, and log cost per run so a change in behaviour shows up as a number rather than as a surprise invoice. Deterministic steps should be scripts, not model calls.
Is a per-seat AI subscription cheaper than running your own agent?
It depends on headcount and usage, and the comparison changes shape as you grow. A per-seat subscription scales with people whether or not they use it; an owned deployment scales with work done. Teams with many light users often find seats cheaper, and teams running heavy automation across a few people usually do not.
Who pays the cloud bill in a Sovereign deployment?
You do, directly to your cloud provider, because the infrastructure sits in your own account. That is a consequence of owning it rather than a markup, and infrastructure is typically the smaller line next to model inference.