How much does an AI agent cost to run?

It depends on the job, but the model bill is rarely the biggest line. In the worked example in this post, a research agent handling 2,000 tasks a month costs about $4,000 a month, or roughly $2 per task, and model tokens are under 20% of that. Human review, maintenance and tooling make up most of the rest. Budget per completed task, not per token.

What are the hidden costs of AI agents?

The hidden costs are retries, repeated context, tool and data fees, monitoring, human review of flagged outputs, and ongoing maintenance when prompts, models and upstream systems change. Agents also call the model several times per task and resend their history each step, so token use grows faster than the task count. Gartner expects at least half of GenAI projects to overrun budget through 2028.

How should I budget for my first AI agent?

Model one workflow, estimate volume, calls per task, retry rate, tool fees, review minutes and maintenance hours, then divide by completed tasks to get cost per task. Multiply by expected volume for twelve months, add your build cost, and hold back 20% contingency. Start with a small volume cap so a bad assumption costs hundreds, not thousands.

Is an AI agent cheaper than hiring an employee?

Often yes per task, but not always per outcome. A human SDR runs $95,000 to $115,000 fully loaded, while an AI SDR tool alone is $45,000 to $60,000 a year. Once you add review and maintenance, the gap narrows, and quality gaps such as lower show rates can erase it. Compare cost per outcome, not cost per seat.

AI Agent Cost: The 6 Layer Budget Model for Your First Agent

Search for AI agent cost and you will find the same page ten times: a table of vendor prices, a range like “$500 to $50,000,” and a sales pitch. Useful if you are choosing between two chatbots. Useless if you are about to put an agent into a real workflow and someone in finance asks what it will cost in year one.

Here is why that matters. Gartner expects at least half of GenAI projects to overrun their budgets through 2028, blaming poor architectural choices and thin operational know-how. Separately, it expects more than 40% of agentic AI projects to be cancelled by the end of 2027, with escalating costs and unclear value as the named causes. Cost is not a footnote in these failures. It is the reason.

The pattern I see is simple. Teams price the thing they can see, the subscription or the token rate, and miss everything that surrounds it. So this post does not list vendor prices. It gives you a six layer AI agent cost model, a formula for cost per completed task, and a worked first year budget you can copy into a spreadsheet. If you are still deciding which agent to buy, start with our field guide to the best AI agents for business, then come back here to budget it.

Why the sticker price misleads

A chat call is one model request. An agent is a loop: plan, call a tool, read the result, decide, call again. Each step resends the accumulated context. By step 20 you have paid for the same history twenty times.

One analysis from Atlan makes the gap concrete: one model call per chat versus six per agent workflow, which comes to roughly 19 times the model tokens before tools, retrieval, monitoring or human review enter the picture. Treat that as an illustration rather than a law, but the direction is right. Token spend scales with steps, not with tasks.

Vendor pricing hides the same thing from a different angle. Intercom’s Fin lists $0.99 per resolution. Salesforce launched Agentforce at $2.00 per conversation, charged whether or not the issue is resolved, which one comparison turns into about $3.33 per resolved conversation at a 60% resolution rate. HubSpot’s Breeze is listed near $0.50. Those numbers look comparable. They are not, because each vendor defines the unit differently. A resolution, a conversation and an action are three different denominators.

That is the first rule of agent budgeting: pick one unit, the completed task, and convert everything into it.

The 6 layer AI agent cost model

I think of agent cost as six layers. Most buyers price layer two and stop.

Layer	What it covers	Cost behavior	Typical owner
1. Build	Scoping, prompts, integrations, test set	One time, amortize over 12 months	Ops or engineering
2. Model usage	Input and output tokens, retries	Variable, scales with steps	Engineering
3. Tools and data	Enrichment, search, vector storage, third party APIs	Variable per task	Ops
4. Orchestration and hosting	Workflow runner, queues, servers, platform fees	Mostly fixed	Engineering
5. Monitoring and evals	Tracing, logs, regression tests, alerts	Fixed plus storage growth	Engineering
6. Human oversight and maintenance	Review of flagged output, prompt fixes, model upgrades	Fixed hours, grows with volume	Ops lead

Layers 1, 4 and 5 are close to fixed. Layers 2 and 3 are variable and driven by volume. Layer 6 is the one nobody budgets and the one that usually decides whether the agent is worth running.

Layer 1: Build

This is the time to scope the job, write prompts, wire integrations, and assemble a test set of real cases with known good answers. For a single, well scoped workflow, plan in days to a few weeks of work rather than months. The mistake is treating build as the whole project. Build is a deposit. The running costs are the mortgage.

Amortize the build cost over twelve months so it shows up in cost per task. If you spend $6,000 on build, that is $500 a month in year one.

Layer 2: Model usage

Model cost per task is calls per task, times average input tokens plus output tokens, at your model’s rates, times one plus your retry rate. Three levers matter most:

Calls per task. Every extra planning or verification step multiplies spend. Fewer, better steps beat clever loops.
Context length. Long documents and conversation history get resent. Trim what goes in.
Model tier. Route simple steps to a smaller model and keep the expensive one for the decision that needs it.

Prices change often, so I will not hardcode a rate. Pull the current numbers from your provider and put them in an input cell.

Layer 3: Tools and data

Agents that act on the world pay for the world. Contact enrichment, web search, scraping, email verification, document parsing and vector storage all bill per call or per record. For research and outbound agents this layer can match or exceed the model bill. Count every external call per task, not just the model call.

Layer 4: Orchestration and hosting

Something has to run the loop: a workflow platform, a queue, a server. Some platforms charge per run, others per trace or seat. Self hosting moves the cost from a subscription to engineering time, which is cheaper only if someone already owns that time. For one agent this layer is usually small and flat. It grows when you add agents.

Layer 5: Monitoring and evals

You cannot manage what you cannot see. At minimum you need traces of each run, a log of failures, and a regression test set you re-run whenever you change a prompt or swap a model. Observability tools charge on volume, and logs accumulate. Budget a flat monthly amount and an hour or two a week of someone actually reading the traces.

Layer 6: Human oversight and maintenance

Two separate costs hide here.

Review. Some share of outputs needs a human. A flagged email draft, a low confidence classification, a deal summary before it goes to a customer. Estimate the flag rate and minutes per review. This is the layer that scales with volume.

Maintenance. Prompts drift, upstream APIs change, models get deprecated, and your own process evolves. Someone owns that. One vendor model I reviewed puts maintenance and tuning for a simple agent at roughly 1.5 times its annual model API bill. Vendor numbers deserve skepticism, but the shape matches what I have seen: maintenance is a recurring line, not a rounding error.

For the governance side of who owns what, see our AI agent deployment checklist.

The cost per task formula

Put the layers into one equation:

Cost per task = [(model cost + tool cost) x (1 + retry rate)] + (flag rate x review minutes x hourly cost / 60) + [(fixed monthly costs) / tasks per month]

Fixed monthly costs are hosting, monitoring, maintenance hours and the amortized build. Divide them by volume and you see the most important property of agents: cost per task falls as volume rises, until review hours start to scale with it.

Worked example: a research agent at 2,000 tasks a month

These are illustrative assumptions, not benchmarks. Replace every input with your own. The agent researches a target company and drafts an account brief for a sales rep.

Input	Assumption
Volume	2,000 tasks per month
Model calls per task	6
Average tokens per call	12,000 input, 1,000 output
Model rates (assumed)	$3 per million input, $15 per million output
Retry rate	20%
Tool and data cost	$0.15 per task
Flag rate	20% of tasks reviewed by a human
Review time	4 minutes per flagged task
Loaded cost of reviewer	$60 per hour
Maintenance	6 hours a month at $100 per hour
Hosting	$150 per month
Monitoring tooling	$100 per month
Build	$6,000, amortized over 12 months

Now the math. One call costs $0.036 in input and $0.015 in output, so $0.051. Six calls is about $0.31. With 20% retries that is about $0.37 per task, or roughly $740 a month.

Layer	Monthly cost	Share
1. Build (amortized)	$500	13%
2. Model usage	$740	19%
3. Tools and data	$300	8%
4. Hosting	$150	4%
5. Monitoring	$100	3%
6a. Human review	$1,600	40%
6b. Maintenance	$600	15%
Total	$3,990	100%

Review is 2,000 tasks x 20% flagged x 4 minutes, about 27 hours, at $60. That is the biggest line by a wide margin. The model bill is under a fifth of the total. If you had budgeted from the token rate alone, you would have been off by a factor of five.

The cost per task is about $2.00. Year one is roughly $47,900, and with 20% contingency, about $57,500. That is a number a CFO can sign.

Now compare it honestly. If a rep spends 20 minutes on the same brief at a $55 loaded hourly rate, that is about $18 a task. The agent is far cheaper per task. But only if the briefs are good enough that reps use them, which is a quality question the cost model cannot answer. That is why we pair every budget with the outcome metrics in our AI agent ROI measurement framework.

Three levers that cut cost without cutting quality

Once you have the model in a spreadsheet, sensitivity analysis is easy. Change one input at a time. In the example above:

Lever	Change	Effect on monthly cost
Cut flag rate	20% to 10%	Saves $800 (review halves)
Cut calls per task	6 to 4	Saves about $250 plus fewer retries
Cap tool spend	$0.15 to $0.10 per task	Saves $100

The flag rate dominates. That tells you where to invest: better evals and tighter prompts that raise first pass quality are worth more than a cheaper model. Teams that obsess over token price and ignore flag rate are optimizing the 19% line.

How to read vendor pricing models

Most agent products fall into four pricing models, and each shifts risk differently. Knowing which one you are signing changes how you fill the spreadsheet.

Pricing model	Example from published rates	Who carries the risk	Watch for
Per resolution	Intercom Fin at $0.99 per resolution	Vendor, if the agent fails	How “resolved” is defined; assumed resolutions can include customers who gave up
Per conversation	Agentforce at launch, $2.00 per conversation	You, because failures still bill	Escalated chats cost the same as resolved ones
Per action or credit	Salesforce Flex Credits	Shared	Actions per task are hard to predict before shadow mode
Seat or platform fee	Typical AI SDR tools	You, fixed regardless of output	Paying for idle capacity at low volume

Per resolution looks safest, and for support it often is. But the definition does the heavy lifting. If a vendor counts silence after a reply as a resolution, your real cost per solved problem is higher than the sticker suggests. Ask for the definition in writing and sample 50 resolved tickets yourself.

Per conversation pricing is the opposite. You pay for attempts, so your effective cost is price divided by resolution rate. At $2.00 and a 60% resolution rate, that is about $3.33 per solved conversation, which is why one comparison lands there. Run the same division on any vendor before comparing.

Seat pricing is the easiest to budget and the easiest to waste. If an AI SDR tool costs $4,000 a month and you only feed it 300 prospects, your cost per prospect is thirteen dollars before review. The same tool at 3,000 prospects is a tenth of that. Volume is a cost lever, so test at the volume you actually plan to run.

Finally, notice what none of these models include: your review time, your maintenance hours and your monitoring. Those sit on your side of the line in every contract. Add layers 5 and 6 to any vendor quote before you compare it to building your own.

The first year rollout: a budget you can defend

Do not commit twelve months of spend on day one. Use stages with caps.

Weeks 1 to 2: Scope and test set. Pick one workflow with a clear unit. Collect 50 to 100 real cases with known good outputs. Spend: build time only.
Weeks 3 to 4: Shadow mode. Run the agent on live inputs, but humans do the work and compare. Measure flag rate, retries and calls per task. This replaces my assumptions above with yours.
Month 2: Capped pilot. Limit volume to 10 to 20% of target. Set a hard monthly spend cap in the provider dashboard and an alert at 70%.
Month 3 onward: Scale in steps. Raise volume only when cost per task and flag rate are stable for two weeks.
Quarterly: Re-price. Model rates, tool fees and your own process all move. Re-run the spreadsheet.

If your question is whether to build the agent at all or hire for the role, run the decision through our AI agent vs hiring framework first. In that comparison a human SDR runs $95,000 to $115,000 fully loaded against $45,000 to $60,000 for an AI SDR tool alone. The gap narrows once you add the layers above, and quality differences can erase it entirely.

Common budgeting mistakes

Budgeting from the token rate. In the worked example, tokens were 19% of the cost.
Ignoring retries and repeated context. A 20% retry rate is a 20% tax on layers 2 and 3.
Treating review as free. If a person looks at it, it costs money. Put minutes and a rate on it.
No maintenance owner. Unowned agents decay silently, then fail loudly.
Comparing vendor units. A resolution, a conversation and an action are not the same denominator. Convert to cost per completed task.
No spend cap. A looping agent with a bad stop condition can burn a month of budget in a day. Caps and alerts are cheap insurance.
Skipping contingency. I hold back 20% in year one because the first assumptions are always wrong in some direction.

What to do this week

Open a spreadsheet. Name one workflow. List the six layers as rows and fill every input from the worked example with your own estimate, flagging the ones you are guessing. Then run shadow mode for two weeks and replace the guesses with measurements.

If you want a second pair of eyes on the model, we build these agents for B2B teams and can walk through your workflow. Book a free growth audit and we will map your specific cost structure. You can also try the free AI growth tools at app.momentumnexus.com.