A prototype AI agent that proves the loop on your real data sits in the low five figures (USD). A production agent for one well-defined workflow — scoped tools, approval gates, evaluation, audit logging — typically runs into the five figures. Multi-system agents with write access across several systems of record start in the high five figures and up.
Those ranges are wider than a chatbot's for one reason: an agent takes actions. The cost is driven by integrations, evaluation, and the guardrails that make it safe to let software act on your systems — not by the model, which is very often the smallest line on the invoice.
If you are weighing an agent against a simpler assistant, our chatbot cost breakdown is the fair comparison. An agent is roughly double for the same subject matter, and the delta is almost entirely safety engineering.
The three tiers of AI agent
Quotes vary wildly because "agent" covers three very different things. Pinning your tier is the single biggest lever on cost:
| Tier | What it does | Typical build | Best for |
|---|---|---|---|
| Read-only agent | Gathers, enriches, and summarises across systems; proposes but never commits | Low five figures | Triage, research, enrichment, drafting |
| Production agent | Takes scoped actions in one workflow, with human approval on consequential steps | Five figures | One recurring multi-step process |
| Multi-system agent | Acts across several systems of record with role-based access and audit evidence | High five figures and up | Replacing a process that spans teams and tools |
What actually drives the cost
- Number of systems, not number of features. Each system the agent touches adds credentials, an access review, error handling, and a new set of states to evaluate. Two systems is not twice one — it is closer to three times.
- Write access versus read access. Reading is cheap. Writing attracts approval gates, reversibility design, audit logging, and usually a security review. The same agent costs substantially more the day it is allowed to change something.
- Evaluation. Building a task set with known-correct outcomes is real engineering, and it is the line teams try to cut first. It is also the only thing separating an agent that works in a demo from one that works in week six — see reducing LLM hallucinations for why unmeasured quality drifts.
- Exception handling. Deciding what the agent does when it is unsure is often more work than the happy path, and it is what makes the difference between escalation and a confidently wrong action.
- Compliance evidence. In regulated environments, the artefacts proving what the agent may do and did do can rival the build itself.
- Model choice barely matters. Swapping between OpenAI and Anthropic models is a configuration change in a well-built pipeline, so it is not worth optimising early.
Build cost vs running cost
Budget two numbers. Agents differ from chatbots here in a way that catches teams out: a chatbot spends tokens per conversation, but an agent spends tokens per *step*, and a single task can be dozens of steps.
| Cost | One-off (build) | Ongoing (monthly) |
|---|---|---|
| Engineering | The bulk of the build | Maintenance as your systems change |
| LLM usage | — | Per step, not per task — the multiplier that surprises people |
| Integrations | Setup and approval per system | Breakage when an upstream API changes |
| Evaluation | Building the task set | Re-running it after model or prompt changes |
| Oversight | — | Human review time, falling as autonomy widens |
The line item that runs away
An agent that retries on failure without a step cap can turn one task into hundreds of model calls. Cap steps and set a per-run token budget in code — not in the prompt, which is a suggestion the agent can route around. Our guide to cutting LLM API costs covers caching and model routing, both of which apply per step and so compound harder for agents than for chatbots.
How to keep an AI agent affordable
- Start read-only. You get most of the value before you pay for write-access approvals and reversibility design, and you learn whether the agent is any good at deciding before you let it act.
- One workflow, not a platform. Agents earn back on repetition of a specific task. Building a general-purpose agent is how budgets disappear without a single process being automated.
- Route by difficulty. Use a cheaper, faster model for routine steps and reserve the top-tier model for genuinely hard decisions. Because agents spend per step, this compounds far more than it does for a chatbot.
- Cache aggressively. Agents re-read the same records constantly across a run; caching retrieval and tool results is often the single largest saving available.
- Cap and observe. Step caps, per-run token budgets, and cost-per-run dashboards make a runaway loop impossible rather than merely unlikely.
When an AI agent is not worth the money
If the task runs a handful of times a month, the engineering will not pay back — agents earn on repetition. If the process is rigid with no exceptions, a plain script costs a fraction and is easier to debug. And if nobody can define what a correct outcome looks like, you cannot build the evaluation set, which means you cannot know whether the agent is working, which means you should not be paying for one yet.
We would rather tell you that on the first call than six weeks in. It is worth asking any agency the same question directly — our guide to vetting development agencies has the rest of the list.