What you are actually paying for
An agent is software that reads, decides and acts inside your systems. The cost sits in five places, and only one of them is the model.
| Cost area | What it covers | Scales with |
|---|---|---|
| Discovery and scoping | Mapping the process, the data and the risks | How complicated the process is |
| Build and integration | Connecting the systems and tools the agent may use | Number and quality of the systems and APIs |
| Guardrails and evaluation | Test sets, approvals, logging, failure behaviour | The cost of a wrong action |
| Running costs | Model usage, hosting and monitoring | Volume of work and model choice |
| Maintenance | Provider and model changes, process updates | How often your process or systems change |
What moves the price up or down
| Factor | Keeps cost down | Pushes cost up |
|---|---|---|
| Systems involved | One or two with good APIs | Many, or legacy systems with no API |
| Data quality | Structured and consistent | Messy, scattered, inconsistent documents |
| Cost of a mistake | Low impact and easy to reverse | Financial, legal or customer-facing impact |
| Autonomy | Drafts that a person approves | Acts without review |
| Volume | Modest and steady | High volume or strict response times |
| Sensitivity | Internal, low-sensitivity data | Personal or special-category data |
Start with a narrow first agent
The cheapest way to learn what an agent is worth is to build a small one against a real process.
- Pick a process with clear inputs, a clear output, an owner and something you can measure.
- Run it with a person approving every action, so you see real behaviour before it has any independence.
- Widen what it may do alone only where the evidence supports it.
Do not ignore running costs
Model usage, hosting and monitoring continue after launch, and usage grows with volume. Estimate them from real volumes during scoping, not after go-live, and ask the provider to show the assumptions.
How to judge the return
- Time spent on the task today, multiplied by how often it happens.
- Error and rework rates before and after.
- Cycle time from request to completion.
- Response time for customers or colleagues.
- Capacity released for work that needs judgement.
Measure these before the build. Without a baseline, “it feels faster” is the only evidence you will ever have.
Questions to ask any provider
Before you sign
- How will the agent be tested before it goes live?
- What can it do without approval, and what must wait for a person?
- What is logged, and for how long?
- Where is the data processed, and is it used to train models?
- What happens when the model or provider changes?
- Who owns the code, prompts and evaluation sets?
How we approach it
We begin with a free assessment that ranks opportunities by impact and effort, then agree a fixed scope and quote in writing. Read more about AI agent development in Ireland, or see why agents are not chatbots.