ByAhsan Nadeem@ashtech52Lead Full-Stack Engineer
The cost to build and run an AI agent for its first year comes to about $64,000 for a small internal agent, $232,000 for a mid-sized one and $755,000 for a high-volume one, on the assumptions shown below. These are our estimates, not quotes. The build is only 24% to 38% of each total. Most of the rest is people: checking the agent's work and doing the tasks it can't finish.
Prices were checked on 4 November 2026.
That's why a build quote is the wrong number to compare. Model your cost per completed task instead, then test it against the same work done with plain automation. Choosing between the two is the first decision in any workflow automation project, and you can make it with the sums here.
The cost to build and run an AI agent, in one worked example
Take a mid-sized agent handling 10,000 tasks a month, such as sorting and answering support requests. Done by hand, each task takes 8 minutes. At about $30 an hour, that's $4 a task, or $480,000 a year.
The $30 starts from the US median for customer service staff, $21.53 an hour in May 2025. We multiplied it by 1.4 for employment overheads. That multiplier is our assumption.
The agent's first year looks like this:
Item | Assumption | Year one |
|---|---|---|
Build | 140 person-days at $600 | $84,000 |
Tokens | $1,104 a month | $13,248 |
Hosting, logging, tools | $800 a month | $9,600 |
Upkeep | 20% of build a year | $16,800 |
Human review | 20% of tasks, 3 minutes each | $36,000 |
Handoffs to people | 15% of tasks, 8 minutes each | $72,000 |
Total | $231,648 |
That works out at $1.93 per task, against $4 by hand. The agent wins here. But look at where the money goes: tokens are under 6% of the year, while review and handoffs are 47%. Change the people assumptions and the answer changes fast.
Why published quotes range from $5,000 to over $500,000
Every figure in this table comes from a company selling agent builds or an agent product, so treat them as marketing with useful detail.
Vendor estimate | Scope | Price |
|---|---|---|
TierZero, Feb 2026 | Small pilot or simple bot | $5,000 to $30,000 |
RaftLabs, Dec 2025 | Single-workflow proof of concept | $15,000 to $40,000 |
TierZero | Mid-market agentic system | $20,000 to $100,000 |
RaftLabs | Production agent, 3 to 5 integrations | $80,000 to $200,000 |
RaftLabs | Multi-agent system | $150,000 to $400,000 |
TierZero | Enterprise production agent | $100,000 to $500,000+ |
These ranges don't really disagree. They price different things. A pilot with one integration and no monitoring isn't a cheaper version of a production agent. It's a different project.
Neither vendor prices the staff time spent checking individual tasks in its build or monthly figures. That's normal for a build quote, because you pay those staff yourself. It's still your cost.
For market context, Gartner's May 2026 release forecast $206.5 billion of AI agent software spending in 2026. Its September release puts "AI Agents and Assistants" at about $29.2 billion after changing its categories. The two aren't comparable. Neither tells you what your agent will cost.
Build costs: integrations, guardrails, evaluation
RaftLabs gives the most useful public breakdown, again as a vendor estimate:
- Discovery and architecture: $5,000 to $20,000.
- Each integration: $8,000 to $15,000 for a REST API, $20,000 to $30,000 for a legacy system.
- Agent logic and orchestration: $15,000 to $60,000.
- Evaluation and testing: $5,000 to $20,000, over 2 to 4 weeks.
- Deployment and monitoring: $5,000 to $15,000.
- Compliance (HIPAA, PCI, SOC 2), if needed: $20,000 to $60,000, plus $500 to $2,000 a month.
Integrations are the biggest item you can control. Five of them cost $40,000 to $150,000 on those figures, before any agent logic. Every system you take out of version one saves real money, so scope it carefully.
Don't trim evaluation to save money. It's how you find out your completion rate, and the sensitivity table below shows that's the number that matters most. If you're unsure how to contract work with this much unknown, compare fixed price and time and materials first. TierZero, also a vendor, says total cost usually lands 30% to 40% above projections, so hold a contingency.
Running costs: tokens, retries, tools, logging
Token cost per run is simple arithmetic. Our medium agent makes 6 model calls per task, each with 6,000 input tokens and 400 output tokens. That's 36,000 in and 2,400 out.
At Sonnet 5.5's list price of $2 per million input tokens and $10 per million output:
- Input: 36,000 x $2 / 1,000,000 = $0.072
- Output: 2,400 x $10 / 1,000,000 = $0.024
- Per run: $0.096
- With 15% retries: $0.11
- At 10,000 tasks a month: $1,104
Replace the calls and tokens with your own logs once you have a prototype. Until then, they are guesses, ours included. Four things move the figure:
- Model choice. The same run costs $0.0048 on Haiku 5.5, a twentieth of the price, if it does the job as well. Test that. A cheap model that hands off more tasks costs more overall.
- Caching. Sonnet 5.5 cache reads cost $0.10 per million tokens against $2 for fresh input, though cache writes cost $2.50. Long system prompts repeated on every call gain the most.
- Batching. Anthropic lists 50% off for batch processing, which suits work that doesn't need an instant answer.
- Price changes. Gemini 3.8 Flash costs $0.75 input and $3.75 output through 31 December 2026, then $1.50 and $7.50. Our run costs $0.036 now and $0.072 from January. Budget a full year at the later price.
Tools add their own meter. Google Search grounding is free for 5,000 requests a month, then $14 per 1,000. One search per task at 10,000 tasks a month comes to about $70, though the default cap of 1,500 requests a day would need checking at higher volumes.
For hosting and logging, RaftLabs estimates infrastructure of $100 to $300 a month at small scale, rising to $800 to $3,000 or more. Our $300, $800 and $2,500 allow for logging and tool fees on top of hosting.
The cost people leave out: human review
Review is a real payroll cost, and it doesn't shrink just because the model gets cheaper.
Three questions set it:
- What share of tasks does someone check? High-risk actions, such as refunds, may need all of them at first.
- How long does a check take? A good screen showing the agent's sources and proposed action can cut this. It's also part of the build.
- What happens when the agent stops? A handoff costs a full manual task, and the person may need to read what the agent did first.
Our model uses 3 minutes per review and 8 minutes per handoff. Here's what happens when you change one assumption at a time on the medium agent:
Change | Year-one cost per task |
|---|---|
Base case | $1.93 |
Token prices double | $2.04 |
Token prices halve | $1.88 |
Build costs 50% more, with upkeep staying 20% of build | $2.35 |
Review share rises from 20% to 40% | $2.23 |
Handoffs fall from 15% to 7.5% | $1.63 |
Handoffs double to 30% | $2.53 |
Doubling token prices adds 11 cents. Doubling handoffs adds 60 cents. On these numbers, time spent raising the completion rate pays back faster than time spent shaving tokens.
Year-one budgets from $64,000 to $755,000, with assumptions to replace
All three columns are our estimates, using Sonnet 5.5 for every run and $600 a person-day. That rate is illustrative, not a ZNova quote. It's about double what RaftLabs' published $6,000 to $6,500 per person-month works out at per working day, assuming 21 days a month, so replace it with your own quote.
Small | Medium | Large | |
|---|---|---|---|
Tasks a month | 2,000 | 10,000 | 50,000 |
Model calls per task | 4 | 6 | 10 |
Tokens per call, in / out | 4,000 / 300 | 6,000 / 400 | 10,000 / 600 |
Build (person-days) | $24,000 (40) | $84,000 (140) | $180,000 (300) |
Tokens a month, with 15% retries | $101 | $1,104 | $14,950 |
Hosting, logging, tools a month | $300 | $800 | $2,500 |
Upkeep a month (20% of build a year) | $400 | $1,400 | $3,000 |
Review a month (share checked) | $900 (30%) | $3,000 (20%) | $7,500 (10%) |
Handoffs a month (share) | $1,600 (20%) | $6,000 (15%) | $20,000 (10%) |
Running cost a month | $3,301 | $12,304 | $47,950 |
Year one, build plus 12 months | $63,614 | $231,648 | $755,400 |
Per task, year one | $2.65 | $1.93 | $1.26 |
Same work by hand at $4 | $96,000 | $480,000 | $2,400,000 |
The 20% upkeep sits just above TierZero's 15% to 20% a year. RaftLabs instead quotes $1,000 to $5,000 a month for a part-time engineer, which would raise the small agent's cost.
When the agent loses
- Short tasks. The $4 is doing a lot of work. If the task takes a person 3 minutes, manual cost falls to $1.50, while the medium agent still costs $1.56, with handoffs down to 3 minutes and review unchanged.
- Savings you can't bank. The comparison assumes the staff time saved is really saved, through hiring you avoid or work you redeploy. Otherwise your payroll doesn't change, and the budget only got bigger.
Agent or plain workflow automation? A cost-per-task test
Anthropic, which sells models, advises teams building agents to start with the simplest solution. It suggests workflows for well-defined tasks and agents for open-ended ones, and warns that autonomy brings higher cost and compounding errors. Gartner also says many use cases sold as agentic don't need agents.
Here's how to test it on the medium case. Assume a rules-based workflow takes 30 person-days ($18,000), costs $600 a month to host and maintain, and has 5% of its output spot-checked. Whatever it can't handle goes to a person at $4.
Option | Year-one cost | Per task |
|---|---|---|
All by hand | $480,000 | $4.00 |
Workflow covering 40% of tasks | $316,800 | $2.64 |
Agent completing 85% of tasks | $231,648 | $1.93 |
Workflow covering 60% of tasks | $222,600 | $1.86 |
On these numbers the break-even is about 58%. If plain rules can handle more than that share, the workflow wins in year one, and it's easier to audit too. If your tasks vary too much for rules to reach it, the agent earns its cost.
You can run this test this week:
- Pull 200 recent tasks.
- Mark each one by whether a clear rule could have handled it.
- Work out that share, then replace our numbers with your own volumes and handling time.
- Prototype the agent only on the tasks rules can't reach.
Often the right design is both: rules first, with an agent handling what's left. Our guide to AI agents for small business covers where they tend to fit.
Should you buy per-agent software instead?
If a packaged product does the job, it can beat a build in year one. Prices below are from each vendor's page:
- [Intercom Fin](https://www.intercom.com/pricing): $0.99 per outcome, charged at most once per conversation. On Intercom's own helpdesk you also pay seat fees. With another helpdesk there's no seat cost, but a minimum monthly commitment applies.
- [Salesforce Agentforce](https://www.salesforce.com/agentforce/pricing/): $2 per conversation, or $500 per 100,000 Flex Credits at 20 credits an action, which is $0.10 an action. The Agentforce add-on for Sales, Service and Field Service is $125 a user a month.
At 8,500 completed tasks a month, Fin costs $8,415 before any seats. Intercom also bills $0.99 when Fin runs a Procedure that ends in a handoff, so treat that as a floor. Agentforce's conversation price on all 10,000 tasks comes to $20,000. At 3 actions a task on Flex Credits, it's $3,000.
Compare Fin with our medium build, leaving out human time, which you pay either way. Tokens, hosting and upkeep come to $3,304 a month. The saving of $5,111 a month pays back the $84,000 build in about 16 months. So buying often wins for year one, and building can win later if volumes hold and the product can't do what you need.
Read what you're billed for, too. Intercom counts a customer leaving without asking for more help as a resolution. That's a sensible rule for billing, but you'd still want to know whether those customers were actually helped. A packaged product also has to reach your systems, so an agent working inside your own back-office tools may have no packaged option.
What to ask for in a quote, and what Gartner's cancellation forecasts warn about
Gartner predicted in June 2025 that over 40% of agentic AI projects would be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. It's a prediction, not a measurement. In September 2026 it also predicted that 70% of enterprises would abandon agentic AI built by vendor forward-deployed engineers by 2028, citing rising costs and customers who can't change the systems themselves.
Both point at things a quote can settle. Ask each supplier for:
- A task definition and completion target. What counts as done, on which test set, at what rate?
- An evaluation set you own. Real cases you can rerun after every change of model or prompt.
- Projected cost per completed task. Include retries, tools and review time, with the assumptions written down.
- Per-integration pricing. So you can remove one and see the saving.
- Logs of every run. Tokens, tools called, outcome and who reviewed it.
- A spending cap and a stop rule. What halts the agent if costs or errors jump?
- The upkeep plan. Who updates it when a model is retired or an API changes, and can your own team do it?
- Handover. Code, prompts and evaluation sets are yours, in a form another team can run.
- A kill criterion. The cost per task at which you stop the project, agreed before the build starts.
The same discipline applies to any first version. See how we break down SaaS MVP development cost if the agent is part of a new product.
Frequently asked questions
How much does it cost to build an AI agent?
Vendor estimates run from $5,000 to $30,000 for a simple pilot, $80,000 to $200,000 for a production agent with 3 to 5 integrations, and $150,000 to $400,000 for multi-agent systems. Our own worked estimates are $24,000, $84,000 and $180,000 at an illustrative $600 a person-day. The number of integrations is usually the largest driver you control.
How much does an AI agent cost per month to run?
On our assumptions, about $3,300 a month at 2,000 tasks, $12,300 at 10,000 and $48,000 at 50,000, including review and handoffs. Without human time, those fall to about $800, $3,300 and $20,500. Tokens alone were $101, $1,104 and $14,950. Your own logs from a prototype will give a better figure.
Why do AI agent quotes vary so much?
They price different projects. A one-integration pilot without monitoring can cost $15,000, while a production agent with several integrations, handoffs and monitoring can cost $200,000. Ask each supplier to itemise integrations, evaluation and upkeep, so you can compare the same scope.
Is it cheaper to buy per-agent software than to build?
Often, in year one, if a product fits your task. Intercom Fin at $0.99 per outcome costs about $8,415 a month for 8,500 outcomes, plus seats if you use Intercom's helpdesk. Against our $84,000 build with $3,304 a month in non-staff running costs, the build takes about 16 months to pay back. Packaged products may not reach your internal systems.
What is the biggest hidden cost of an AI agent?
Human time. In our models, review and handoffs make up 57% to 76% of monthly running cost, while tokens are 3% to 31%. Doubling handoffs on the medium agent adds 60 cents per task. Doubling token prices adds 11 cents.
Should we use an agent or simple automation?
Test the share of tasks that plain rules could handle. In our medium example, a workflow beats the agent in year one once it covers about 58% of tasks. Above that, use rules. Below it, consider an agent, ideally only for the tasks rules can't reach.
Done the 200-task count? Book a free consultation and bring it, with your task volumes and one month of examples. We'll work through the cost per task with you, including whether plain automation would do the job.
Sources
- Gartner, autonomous business release, $206.5 billion agent software forecast (5 May 2026)
- Gartner, worldwide AI spending forecast (16 September 2026)
- Gartner, over 40% of agentic AI projects predicted to be cancelled (25 June 2025)
- Gartner, 70% predicted to abandon vendor forward-deployed agentic AI (29 September 2026)
- Anthropic, API pricing (read 4 November 2026)
- Google, Gemini API pricing (read 4 November 2026)
- Anthropic, Building effective agents (model vendor)
- Intercom, pricing (vendor, read 4 November 2026)
- Intercom, Fin outcomes help article (vendor)
- Salesforce, Agentforce pricing (vendor, read 4 November 2026)
- RaftLabs, AI agent development cost (vendor, 24 December 2025)
- TierZero, production AI agent cost (vendor, 6 February 2026)
- US Bureau of Labor Statistics, customer service representatives (May 2025 pay data)
