A focused, single-task AI agent built on a commercial model API typically takes 10 to 25 engineering days, which puts it at roughly €10,000 to €30,000. An agent that writes to your core systems, with approvals and proper testing, is more like €30,000 to €80,000. Orchestrated multi-agent systems in regulated settings start around €80,000 and go up from there.
Running costs come on top. Model usage is often smaller than people fear, while monitoring and maintenance are often larger. The rest of this article explains where those numbers come from, so you can adjust them to your own situation rather than taking anyone’s range on faith, including ours.
One assumption sits underneath every figure here: a blended rate of €800 to €1,200 per senior engineering day. That is a reasonable planning figure for AI engineering in Ireland and the UK. If your rate is different, substitute it. The effort in days is the part that carries over.
The short answer: three tiers of AI agent
“AI agent” covers everything from a chatbot with one tool to a fleet of cooperating processes running a back office. Price only makes sense once you know which tier you are in. The table below is the planning model we use when scoping.
| Tier | What it does | Typical build effort | Indicative build cost | Timeline | Main running cost drivers |
|---|---|---|---|---|---|
| 1. Single-task agent | One workflow, one or two integrations, mostly read-only or a single low-risk write action (e.g. triage inbound emails and draft replies) | 10–25 days | €10k–€30k | 2–5 weeks | Model usage (often tens to low hundreds of euro a month), light hosting |
| 2. Workflow agent | Several tools, writes to systems of record (ERP, CRM, ticketing), human approval steps, evaluation suite, audit logging | 30–70 days | €30k–€80k | 6–12 weeks | Model usage, observability, prompt and eval maintenance, integration upkeep |
| 3. Orchestrated or regulated system | Multiple agents or long-running processes, complex permissions, regulated data, formal risk assessment and documentation | 80–200+ days | €80k–€250k+ | 3–6 months | Everything above plus compliance reviews, incident handling, possibly self-hosted inference |
Most businesses asking this question for the first time belong in tier 1 or tier 2, even if the original idea sounded like tier 3. Getting one workflow agent into production and learning from it is nearly always cheaper than designing the grand system up front.
Where the money actually goes
The model call itself is the smallest engineering task in an agent project. A working prototype that calls an API and uses a couple of tools can be done in a day or two. The budget goes into everything needed to make that prototype safe to run unattended against real data.
Discovery and process mapping
Before writing code, someone has to pin down what the agent actually does. That means its inputs, the decisions it makes, which systems it touches, what a good output looks like and what happens when it is unsure. Expect 2 to 5 days for a tier 1 or tier 2 project. Skipping this step does not save money. It moves the cost into rework, because the first version will automate a process nobody fully described.
Integrations
Integrations are usually the biggest variable. A clean REST API with good documentation might take a day to wire up. A legacy ERP with no API, a SharePoint estate with inconsistent permissions, or a supplier portal that only offers CSV exports can each take a week or more. When a quote looks surprisingly high, the cause is usually here. Our guide to LLM integration in production systems covers the patterns in more detail.
Evaluation and guardrails
An agent that works on ten hand-picked examples will fail on the eleventh in ways you did not predict. Building an evaluation set from real historical cases is not optional work. You need a harness that scores outputs and a way to re-run it whenever a prompt or model changes. Budget roughly 15 to 25 per cent of the build for this. The same goes for guardrails: input validation, output schemas, permission checks on every tool call, and rules for when the agent must stop and ask a human.
Human-in-the-loop design
Any agent that writes to a system of record needs an approval path, and the approval path has to be quick to use or people will route around it. That means deciding which actions auto-execute and which need sign-off, and where the approval request appears: Slack, Teams, email, a phone. Plan for what happens when nobody responds. We wrote up our own approach in Human-in-the-Loop Action Tags: How to Make AI Agents Easy to Steer from Your Phone, which is how we run this ourselves.
Deployment, observability and handover
Getting the agent into production involves several pieces. Hosting, secrets management, logging every model call and tool call with enough context to debug it later, cost tracking per run, alerting, and documentation your team can actually maintain. This typically adds 3 to 10 days depending on how mature your existing infrastructure is.
Security and compliance
Personal data creates obligations. If the agent processes it, you need a lawful basis, a record of processing and probably a DPIA. If it touches anything the EU AI Act classes as high-risk, the documentation and oversight requirements go up sharply. Most tier 1 and tier 2 agents in operations, finance and customer service are not high-risk, but you need to check rather than assume. Our AI security and compliance guide for EU firms walks through the practical steps, and the Data Protection Commission publishes guidance relevant to Irish organisations.
Running costs: the bill after launch
Build cost is a one-off. Running cost is what decides whether the agent pays for itself over three years, and it has three parts.
Model usage
Model usage is easy to estimate once you know your volumes. Multiply runs per month by tokens per run, and then by the provider’s price per million tokens. Here is an illustrative example with round numbers. Check current pricing from your provider, because it changes often.
Say an agent handles 2,000 supplier emails a month. Each run involves several steps: classify, extract fields, look up the purchase order, draft a response. Together these consume about 15,000 input tokens and 1,500 output tokens. That totals 30 million input tokens and 3 million output tokens a month. At an illustrative €3 per million input tokens and €15 per million output tokens, the monthly model bill is about €135. Swapping in a cheaper model for the classification step could cut that further. Moving everything to a top-tier reasoning model could multiply it several times over.
The lesson is that token cost scales with design choices, not just volume. Several things push it up: long context stuffed into every call, unnecessary retries, and agents that loop through tool calls without a step limit. A well-designed agent routes simple steps to small models and caches stable context. It also caps the number of iterations per task.
Hosting and infrastructure
Most agents run comfortably on modest cloud infrastructure: a container or serverless function, a queue, a database for state and logs. The cost is usually tens to low hundreds of euro a month. Self-hosting open-weight models changes the picture. You trade per-token fees for hardware or GPU rental, plus the engineering time to operate it. That trade only makes sense at high volume or when data residency rules out external APIs.
Maintenance and drift
This is the cost people forget. Model providers deprecate versions, so prompts tuned for one model behave differently on its successor. Upstream systems change their APIs. The business process changes and nobody tells the agent. As a planning assumption, we budget 15 to 25 per cent of the original build cost per year for maintenance. That covers re-running evaluations on model upgrades, fixing broken integrations and refining prompts as edge cases surface. Agents with no maintenance budget tend to degrade quietly until someone switches them off.
What makes an agent cost more than it should
A few patterns reliably inflate budgets, and almost all of them are decisions rather than technical surprises.
The first is scope that keeps growing. The agent was meant to triage tickets, and halfway through it is also expected to issue refunds, update the CRM and write weekly reports. Each addition means another integration, another approval path and another set of evaluation cases. Fix the scope for version one and keep a written list for version two.
The second is unclear success criteria. If nobody agreed what “good enough” means — 90 per cent of tickets correctly routed, say, with zero refunds issued without approval — the project cannot finish, because there is always one more edge case to chase.
The third is poor data access. If the information the agent needs lives in people’s heads, in scanned PDFs or behind a system nobody has admin rights to, much of the budget goes on data plumbing. That work may be worth doing anyway, but it should be costed honestly as its own line item.
The fourth is treating autonomy as the goal. Fully autonomous agents need far more testing and guardrails than agents that propose actions for a human to approve. Start with approval on every write action, measure how often humans override the agent, and relax the controls only where the data supports it.
Build, buy or configure?
Not every agent needs to be built. In 2026 the major platforms ship agent features inside products you may already pay for, from productivity suites to CRMs to ad platforms. If your use case sits entirely inside one vendor’s ecosystem, configuring their agent may cost a fraction of a custom build. You pay for that in flexibility and in ownership of the logic.
Custom builds earn their cost in two situations. The first is when the workflow spans several systems that do not talk to each other. The second is when the decision logic is specific to how your business works and represents a real advantage. The same trade-offs apply as with any software, which we cover in custom software vs off the shelf for Irish SMEs. If the aim is to absorb growth without adding headcount, our guide to scaling operations without hiring helps frame which processes are worth automating first.
A budgeting checklist before you ask for quotes
You will get more accurate quotes, and comparable ones, if you can answer these questions before talking to anyone:
- The process: which single workflow will the agent own first, and how many times a month does it run?
- The current cost: how many staff hours does that workflow take today, and what does an error cost?
- The systems: which tools must the agent read from and write to, and do they have documented APIs?
- The data: does it involve personal or regulated data, and where must it be processed?
- The risk: which actions can run automatically, and which need human approval?
- The definition of done: what measurable accuracy or throughput makes version one a success?
- The owner: who in your team will own the agent after launch, review its logs and decide on changes?
- The run budget: have you allowed for model usage, hosting and 15–25 per cent annual maintenance?
If you cannot answer the first and last points, that is the place to start. Commissioning a build comes later.
How D3X Solutions LABS helps
We start with a short scoping exercise. Together we map the target workflow, list every system it touches and agree measurable success criteria. That produces an estimate in engineering days, with integrations, evaluation and compliance as separate line items, so you can see exactly what drives the number and cut scope where it makes sense.
From there we build in stages. First comes a working prototype against your real data. Next is an evaluation suite built from historical cases, so quality is measured rather than asserted. Then comes a production deployment with approval paths, logging, cost tracking per run and documentation your team can maintain. We design for model portability from the outset, so a provider’s price change or deprecation is a configuration change rather than a rebuild. Where the agent handles personal data or edges towards regulated territory, we work through the GDPR and AI Act questions alongside the engineering, not after it.
If you have a workflow in mind and want a grounded estimate of what it would cost to build and run, book a free 30-minute call using the form below. We will tell you which tier it falls into, what the main cost drivers are, and whether an off-the-shelf option would serve you better.
Frequently Asked Questions
How much does a simple AI agent cost to build?
A single-task agent with one or two integrations typically takes 10 to 25 engineering days, roughly €10,000 to €30,000 at senior Irish or UK rates. That covers scoping, integration, an evaluation set and a basic production deployment. A no-code configuration of a vendor's built-in agent can cost far less but is limited to that vendor's ecosystem.
What are the monthly running costs of an AI agent?
Running costs have three parts: model usage, hosting and maintenance. For moderate volumes, model usage is often tens to a few hundred euro a month and hosting is similar. Maintenance is the bigger line: plan for 15 to 25 per cent of the build cost per year to handle model upgrades, integration changes and prompt refinement.
How long does it take to build a custom AI agent?
A focused single-task agent can be in production in 2 to 5 weeks. A workflow agent that writes to core systems with approval steps usually takes 6 to 12 weeks. Multi-agent or regulated systems commonly take 3 to 6 months, mostly because of integration, testing and documentation work rather than the AI itself.
Is it cheaper to self-host an open-source model for an AI agent?
Only at high volume or when data residency rules out external APIs. Self-hosting replaces per-token fees with hardware or GPU rental, and it adds the engineering time to run, update and monitor the model. For most tier 1 and tier 2 agents, a commercial API is cheaper overall and quicker to ship.
Why do AI agent quotes vary so much between providers?
Quotes vary mainly on scope assumptions: how many integrations are included and how much evaluation and guardrail work is costed. Compliance documentation and post-launch support also differ between quotes. Ask each provider to break the estimate into engineering days per component so you can compare like with like.
Want this working in your business?
D3X Solutions LABS builds AI agents and automation that pay back fast. Get a fixed-scope estimate for your AI agent build
Get in touch
