How much does it cost to build an AI agent?

Most people who ask this question want a number they can put in a budget line. The honest answer is a range that depends on four things: how many actions the agent may take, how many of your systems it has to read and write, how much a wrong action costs you, and how many people have to check its work. Model tokens are the part everyone can see on a pricing page. They are rarely the part that decides the total.

This post walks through how we price those pieces, shows the token arithmetic with current published rates, and says plainly when an agent is the wrong purchase. We build agents and automations for clients, so we have a bias. The numbers below are arithmetic you can redo, not a quote for your project.

Start with what the agent is allowed to do

The cheapest agent is one that reads a request, picks from a short list of approved actions, and stops. The expensive one can plan across several tools, retry on its own, and change a record in three systems before a person sees it. Each extra permission adds build time, testing time, and something to monitor afterwards.

A useful first step is to write the action list on one page. For each action, name the system it touches, the data it reads, and whether it can be undone. A reply drafted in a shared inbox is cheap to get wrong. A refund issued through a payment API is not. The second kind needs a confirmation step, logging, and usually a rule about who can approve it, and those take real hours to build.

Anthropic's engineering team makes a related point in its guide on building effective agents. It recommends finding "the simplest solution possible, and only increasing complexity when needed," and notes that agentic systems "often trade latency and cost for better task performance" (Anthropic, Building effective agents). The trade is worth making only when a simpler design falls short. Price the simple version first, then price the agent as the upgrade.

The build lines that usually move the price

When we scope an agent, the estimate is built from a handful of lines. Each one can be large or close to zero, depending on the job.

• Discovery and the decision log. Someone has to decide which cases the agent handles, which it refuses, and who owns the exceptions. This is the cheapest line and the one most often skipped.
• Integrations. Reading from a CRM, a helpdesk, a spreadsheet, or a WhatsApp Business account is one job. Writing back safely, with idempotency and error handling, is a second job. Systems with poor APIs or no sandbox are where schedules slip.
• Evaluation. Before launch, the agent needs a fixed set of real examples, scored against the answers a person would accept. Without this, you cannot tell whether a prompt change helped or broke something else.
• Guardrails and the handoff. This covers the refuse list, the escalation path to a person, and the logs someone reads when a customer complains. For Gulf-market work it also covers language: a support thread that switches between Arabic and English needs an explicit rule about which language a draft may use.
• Monitoring after launch. Models and source data change. Someone has to check the failure rate every week for the first few months, and fix the cases that started failing.

Hosting and platform fees are usually the smallest of these, but they still need to be priced honestly. n8n, for example, bases its plans on monthly workflow executions rather than workflow complexity, and its pricing page lists plans from €20 a month on a billed-annually basis, with a self-hosted option for larger volumes (n8n pricing). Prices change, so check the page on the day you budget.

A worked token example with current rates

Model usage is the line you can calculate from the start. Take a support-triage agent that reads a ticket and the policy page it needs, about 4,000 input tokens, and writes a short routing note of about 500 output tokens. Anthropic's pricing page lists Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and Haiku 5.5 at $0.10 and $0.50 for prompts up to 100K tokens (Claude pricing).

For one model call:

• Sonnet 5.5: (4,000 × $2 + 500 × $10) ÷ 1,000,000 = $0.013
• Haiku 5.5: (4,000 × $0.10 + 500 × $0.50) ÷ 1,000,000 = $0.00065

An agent is rarely one call. If the task loops through five model calls with similar sizes, multiply by five. At 1,000 tasks a month, that is about $65 a month on Sonnet 5.5 or about $3.25 on Haiku 5.5. The pricing page also lists batch processing at a 50% discount, which suits overnight jobs that do not need an answer in seconds. Caching cuts the cost of repeated context further, but its write rate is higher, so it pays only when the same long context is reused.

Two things to keep in mind. First, these figures leave out retries, long conversations that grow the context on every turn, and the time spent on evaluation. Second, at this volume the token bill is the smallest line in the whole budget. The larger costs are the integration work, the review path, and the people who answer when the agent gets something wrong. If someone quotes you a figure that is mostly tokens, ask what the other lines are.

Where the total runs over

Gartner has forecast that more than 40% of agentic AI projects will be scrapped by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls (Outlook Business, reporting Gartner's forecast). Those three causes are a useful checklist for your own quote.

The overrun usually comes from one of three places. The first is scope that grew after the demo, when a team asked the agent to also handle the refund, the escalation, and the weekly report. The second is an integration that looked simple in a meeting and turned out to need a vendor's sandbox, a rate limit workaround, or a second copy of the data. The third is the review step that nobody priced, so the agent ends up with a queue of drafts that a person must read anyway.

What we have built that shows the shape of the work

Our own work follows the same pattern. For a media-buying team, we replaced a manual ROI report with a scheduled automation that pulls spend from several ad platforms and revenue from a call-tracking platform, then posts the calculated metrics to Slack. Most of the effort sat in the connections to each platform and in agreeing on the metric definitions, not in the language model. See the media ROI automation case study.

Our competition-research pipeline is another example. It collects ads from several sources, ranks them with a scoring method, and sends curated bundles to the team each day. The value depends on the sources and the ranking, which decide what reaches the team, so that is where the build time goes. See the competition research case study.

We have worked with 71 and counting clients, we are registered in the UAE, and we are expanding to Kuwait. Our service pages cover AI agents, AI automation, AI integration, and AI consultancy. Which one fits depends on whether the work is a new autonomous process, a repeatable workflow, a model call inside a system you already run, or a decision you have not made yet.

When a plain automation is the cheaper answer

If the steps are fixed, a scheduled workflow with one model call in the middle is often enough. Pull the rows, apply the rule, write the result, and alert a person when a field is missing. That is cheaper to build, cheaper to test, and easier to explain to an auditor. An agent earns its place when the next step truly depends on what the last step found, and when no fixed list of branches covers the cases you see.

A good test: write the five most common paths through the process on paper. If you can draw them as a flowchart without dead ends, start with the flowchart. If you keep adding boxes for exceptions you cannot predict, you probably need an agent, and you should budget for the evaluation and review work that comes with it.

Questions to ask any quote

• Which actions can the agent take without a person approving them, and which can it only suggest?
• Which systems does it read, and which does it write to? Who has the API keys, and where are they stored?
• What examples will you use to test it before launch, and who scores the answers?
• What happens when it is unsure? Who gets the handoff, and how fast?
• Which costs are in the quote: build, integration, evaluation, monitoring for the first months, and model usage at your expected volume?
• What does the team do each week after launch, and who owns that?

A quote that answers these clearly is usually a good one, even when it is higher than the one that only mentions tokens. For the vendor-side checks behind the same questions, see what to check before you buy agentic AI services in Kuwait.

If you want a number for your own process, bring the messy version of the workflow and three real cases where it went wrong. Book the free consultation and we will tell you which lines apply and which can be cut.