AI agent development cost: what to budget to build and run an agent

The cost of building an AI agent depends on how many tools it can use, how much freedom it has, how well it must be tested, and how many model calls each task takes. A narrow agent that drafts replies for a person to approve costs far less than one that updates your ERP across several systems. Scenox publishes no prices, so this guide shows the cost drivers, three complexity tiers with hours-based ranges, and how to work out the per-run cost yourself.

#What makes an agent cost more than a workflow

A conventional automation follows fixed steps. An agent decides which step comes next, and that choice adds three costs.

1. Evaluation. You need a set of real cases with known right answers, and you re-run it every time the instructions or the model change.
2. Guardrails. Permission limits, validation rules and approval gates keep a wrong decision from becoming a wrong payment or an email to a customer.
3. Variable run cost. The agent might call the model five times on one ticket and twelve on the next, so you have to track cost per task.

If the task is predictable, a workflow tool or plain code is cheaper and easier to audit. See AI agents vs RPA vs workflow tools for how to decide by step.

#Build cost drivers

#Three complexity tiers (illustrative)

These tiers are our way of sizing work, not a price list. The hours are assumptions. The dollar figures are arithmetic: hours multiplied by an assumed blended engineering rate of $100 to $200 an hour. Your vendor's rates and estimates will differ.

The hours go into scoping, integrations, prompt and rule design, the evaluation set, the review interface, testing and rollout. A big share goes into testing and the controls around the model, and little into the model call itself. If a quote spends almost nothing on evaluation, ask why.

#Run cost: working out the cost per task

Running an agent has four lines: model usage, hosting, monitoring and human review. Model usage is the one people ask about, and you can calculate it from published prices.

We checked Anthropic's published API prices on 2026-10-15 at claude.com/pricing (opens in a new tab). Haiku 5.5 is listed at $0.10 per million input tokens and $0.50 per million output tokens (for prompts up to 100K tokens). Sonnet 5.5 is listed at $2 and $10. Other providers publish their own rates and change them often, so recheck before you budget.

#A worked example (illustrative)

Take a support-triage agent. Assume each ticket takes five model calls (classify, retrieve policy, look up the order, draft, self-check). Context grows with each call, so assume 20,000 input tokens and 2,000 output tokens per ticket in total, and 10,000 tickets a month.

Two things stand out. At this volume the model bill is small next to engineering and review time. And the choice of model moves it by a factor of 20, so send easy steps to a small model and save the larger one for hard calls. Measure accuracy on your own evaluation set before you pick. A cheaper model that sends many more cases to human review can cost more in staff time than it saves in tokens.

Now add the rest, with assumptions marked:

On your side, count review time. If people check 30% of tickets at 2 minutes each, that is 3,000 reviews, or 100 hours a month.

Key point: Agent cost per task is easy to calculate and easy to underestimate, because retries, longer context and tool calls all multiply tokens. Measure it on real cases in a pilot before you commit to volume.

#Costs that rarely appear in a first quote

Four items tend to surface after the contract is signed, so ask about them early.

• Your own team's time. Someone has to supply sample cases, label the right answers, approve rules and test the agent. Budget a few hours a week from a subject-matter expert during the build.
• Security and access review. Getting least-privilege credentials for each system can take weeks of calendar time and some IT hours.
• Change requests. The first version always teaches you something, and a new tool, a new document type or a tighter approval rule is a change that should carry a price.
• Model changes. Providers retire and update models, so re-running your evaluation set is a recurring task.

#How to stage the spend

You don't have to pay for Tier C on day one. To limit risk, start with a read-only version that drafts answers or proposed actions while people keep doing the work (a shadow run). Compare its proposals with what staff actually did. If agreement is high on a real sample, add write access behind approval, then widen the set of tasks. Each stage is a smaller purchase with a measurable result, and you can stop if the numbers don't hold.

#Build, buy or do it yourself

Do it yourself with a platform such as n8n or Zapier, which both offer agent features. That suits contained internal tasks and a small team with an engineer on it. Our n8n, Zapier and Make comparison shows how each prices usage.

Buy an agent product when your use case is standard, such as a help-desk assistant inside your existing support tool. Confirm what it can access and how it logs actions.

Build custom when the agent has to work inside your ERP, follow your approval rules and leave an audit trail you control. That is the work on our AI agents page.

We often steer teams away from a custom agent. If a rule can do the job, a rule is cheaper to run, easier to explain and never hallucinates.

#Questions to ask before you accept a quote

1. How many tools and systems does the agent touch, and which can it write to?
2. What does the agent do on its own, and where does a person approve?
3. How will accuracy be measured, and on whose data?
4. Who pays the model provider, and what is the estimated cost per task?
5. What is included after launch: monitoring, re-testing, fixes?
6. Who owns the prompts, code and evaluation set?

For the wider cost model, with build versus run and a manual-process comparison, read AI automation cost. To turn costs into a payback period, use the ROI method. Our sample audit shows how we size a workflow before anyone writes a quote.

If you have an agent idea and want a realistic scope, book a strategy call. We will tell you whether an agent is the right tool.

#Frequently asked questions

#How much does it cost to build an AI agent?

It depends on tools, autonomy and testing. Using an assumed $100 to $200 hourly rate, a narrow agent that drafts for human approval might take 80 to 160 hours, and a multi-system agent several hundred. Treat those as a sizing method, not a quote, and have a vendor scope your case.

#How much do custom AI agents cost to run?

Model usage is calculable: tokens per task times the provider's published price. Add hosting, monitoring and human review. At moderate volume, review time and maintenance usually exceed the model bill.

#Can I build an AI agent on my own?

Yes, for contained tasks. Platforms like n8n and Zapier include agent features, and model APIs are public. You take on the evaluation, guardrails and upkeep yourself, so be honest about who will own it.

#Why do agents cost more than simple automations?

Because they need evaluation sets, permission limits and approval gates, and their per-task cost varies. A fixed workflow is simpler to test and cheaper to run.

#Does a bigger model always give a better agent?

No. A smaller, cheaper model often handles classification and extraction well, and a larger one earns its price on harder reasoning. Test both on your own cases.