Building an AI agent costs engineering time first and running costs every month after. Illustratively, at €45–65 per engineer-hour, a single-task assistant with tools comes to about €14,400–31,200, a multi-step workflow agent with approvals €32,400–62,400 and an agent working across several systems €54,000–93,600. The monthly bill (tokens, hosting, monitoring, human review, maintenance) decides whether it pays off.
Cost is a main reason agent projects stall: Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. This is the buyer's breakdown: what you pay for, how to estimate each line and where quotes hide cost. For use cases and running an agent safely in production, read our guide to AI agents for business; we do not repeat its checklist here.
The two cost buckets: build cost and run cost
Every agent has two bills. Most quotes show the first; most surprises come from the second.
Build cost (one-off)
Engineering hours to scope the task, connect your systems, prepare data, build prompts, tools and the evaluation set, add approvals and security, and go live. Priced in hours at a rate, or fixed for a defined scope.
Run cost (every month)
Model tokens, hosting, vector database, monitoring, the time your people spend reviewing the agent's work, and maintenance as models and systems change. It grows with volume.
The two are linked: build choices (model calls per task, context per call, which steps use a model at all) fix most of the run cost. A cheap build that ignores this can be the most expensive option over two years.
Build cost by agent complexity
"An AI agent" covers anything from an order lookup helper to a system working across your ERP, CRM and email. Three levels cover most business requests.
Single-task assistant with tools
One task, one or two systems, mostly read access. Example: looks up order status and drafts the reply for the support team to send.
Multi-step workflow agent with approvals
Several steps, some writes, a person approving the risky ones. Example: matches supplier invoices to orders, flags mismatches and prepares postings for approval.
Multi-system agent
Several integrations with write access, varied inputs, limits and monitoring. Example: order exceptions across ERP, CRM, email and the carrier portal.
Illustrative ranges at our published rates of €45–65 per engineer-hour and 40-hour weeks, covering the path to production for a well-scoped task, evaluation set included. Planning examples, not quotes; model usage and licences come on top.
| Scope (illustrative) | Team and duration | Engineer-weeks and hours | Illustrative cost |
|---|---|---|---|
| Fixed-price pilot: one task, read-only or sandbox, evaluation set, go/no-go | 2 engineers × 3–4 weeks | 6–8 weeks × 40 h = 240–320 h | 240 × €45 = €10,800 to 320 × €65 = €20,800 |
| Level 1: single-task assistant with tools | 2 engineers × 4–6 weeks | 8–12 weeks × 40 h = 320–480 h | 320 × €45 = €14,400 to 480 × €65 = €31,200 |
| Level 2: multi-step workflow agent with approvals | 3 engineers × 6–8 weeks | 18–24 weeks × 40 h = 720–960 h | 720 × €45 = €32,400 to 960 × €65 = €62,400 |
| Level 3: multi-system agent with several integrations | 3 engineers × 10–12 weeks | 30–36 weeks × 40 h = 1,200–1,440 h | 1,200 × €45 = €54,000 to 1,440 × €65 = €93,600 |
| Next agent on the same foundations: reuses integrations, logging, evaluation harness | 2 engineers × 4–6 weeks | 8–12 weeks × 40 h = 320–480 h | €14,400–31,200 |
The pilot is the first slice of the build, not an extra: it runs the agent on real cases and ends in a decision, so the larger budget is only spent if the pilot earns it (see our AI agent development page). These ranges are consistent with our AI development cost guide, where an agent connected to your internal systems (3 engineers × 11 weeks, 1,320 hours) comes to €59,400–85,800, inside the Level 3 band.
What drives the build cost
The model call is a few lines of code; the hours go into everything around it. Seven drivers explain most of the gap between two quotes for "the same" agent.
| Cost driver | Cheaper when | More expensive when |
|---|---|---|
| Integrations and API quality | One or two modern SaaS tools with documented APIs and test environments | Several systems, an older ERP without an API, no sandbox, screen-scraping |
| Data preparation | Clean, current records and documents in one place | Duplicates, outdated policies, PDFs and spreadsheets, knowledge in people's heads |
| Permissions and approvals | Read and draft only; a person sends or posts | Writes, payments or access changes; per-user permissions; approval steps in several tools |
| Evaluation set | Historical cases with known correct outcomes are easy to export | Outcomes must be labelled by experts; quality is a matter of judgement |
| Security and prompt-injection defences | Internal, trusted inputs; no irreversible actions | Reads external emails, web pages or uploaded files and can act on them |
| User interface | Runs inside an existing tool (helpdesk sidebar, chat, email) | A new web app with review queues, dashboards and role-based screens |
| Compliance | No personal data, low-risk decisions | Personal or financial data, audit duties, a possible high-risk area under the EU AI Act |
Integrations are usually the largest line: authentication for the right user, scoped permissions, safe retries, useful errors. The Model Context Protocol (MCP), an open standard introduced by Anthropic in November 2024 and since widely adopted, lets you wrap a system once and reuse it across agents and models, which mostly lowers the cost of the second agent. Systems with no usable API need an integration layer first, which is legacy modernisation work and should be quoted as such.
Security is not optional once the agent reads untrusted text. The OWASP Top 10 for LLM Applications lists prompt injection as the top risk (LLM01): instructions hidden in an email or document can redirect an agent. Defences belong in code (least-privilege tools, approvals before irreversible actions, step and spend limits) and take engineering time. On compliance, the EU AI Act places extra obligations on AI used in high-risk areas such as credit, employment and education; its timelines have been changing, so check current dates and your classification with counsel before scoping.
Run cost: what an AI agent costs every month
Five lines. Tokens get the attention; human review and maintenance are often larger.
1. Model tokens per task
Providers charge per token (units of text sent to and returned by the model), with different prices for input and output. An agent makes several calls per task, and each resends instructions, tool definitions, retrieved context and the steps so far. Estimate it like this:
Model cost per task = sum over all model calls of (input tokens × input price per token + output tokens × output price per token)
Monthly model cost = tasks per month × model cost per task
- Measure calls and tokens per task from pilot logs (APIs report token counts per call), on a realistic mix of easy and hard cases.
- Read current prices on your provider's price page, usually per million tokens: input, output and cached input, any batch discount, larger and smaller models.
- Multiply and add a margin for retries and long cases; recheck prices each budget cycle.
- Set spend alerts and a per-task limit from day one.
An illustrative example with hypothetical prices (placeholders for the arithmetic, not any provider's actual prices):
| Item (illustrative) | Figure |
|---|---|
| Model calls per task | 5, each with about 4,000 input and 400 output tokens |
| Tokens per task | 20,000 input, 2,000 output |
| Hypothetical prices | €2.50 per million input tokens, €10 per million output tokens |
| Input cost per task | 20,000 × €2.50 ÷ 1,000,000 = €0.05 |
| Output cost per task | 2,000 × €10 ÷ 1,000,000 = €0.02 |
| Model cost at 10,000 tasks a month | 10,000 × €0.07 = €700 |
2. Hosting, vector database and monitoring
The agent runs in your cloud account; a vector database (managed, such as Pinecone, or self-hosted, such as Weaviate) holds what it retrieves; logging and evaluation tooling record every step. Price these from your cloud and vendor price lists. They are partly fixed and do not fall with volume.
3. Human review time
Approvals and spot checks are work. Illustratively, if 10% of 10,000 monthly tasks need a 3-minute review: 1,000 × 3 = 3,000 minutes, or 50 hours. At an assumed loaded internal cost of €50 an hour, that is €2,500 a month, more than three times the token bill above. It should fall as evaluation scores justify fewer checks; budget for it anyway.
4. Maintenance as a share of build effort
Agents need upkeep: evaluation runs, prompt and tool updates when processes or APIs change, fixes for new edge cases. As an illustrative planning assumption, not an industry statistic, budget 15–25% of the build effort per year. On a €60,000 build that is €9,000–15,000 a year, or €750–1,250 a month, which matches a light retainer of 1 engineer for 2–4 days a month (16–32 hours, €720–2,080 at our rates).
5. Model and provider changes
Providers release new models, change prices and retire old versions. Each switch means rerunning the evaluation set and adjusting prompts. Keep the model behind a thin layer in your code so a switch is a tested configuration change, not a rebuild. Running open models on your own infrastructure is a separate decision with its own hardware and operations bill; our LLM development page covers it.
Find out what your agent costs before you commit
A fixed-price pilot on your real cases, with an agreed evaluation set, measured cost per task and a go/no-go decision. You get a proposal within 24 hours of the first call.
How to cut the cost of an AI agent
Both bills can come down without lowering quality. Most of these choices are made during the build.
Start with one workflow
One task, one owner, one or two systems. A working Level 1 or 2 agent teaches you more, for less, than a Level 3 agent still in testing.
Smaller models for simple steps
Classification, extraction and formatting rarely need the largest model. Keep it for the reasoning steps.
Caching
Keep instructions and tool definitions stable at the start of each prompt so cached-input pricing applies; cache repeated lookups.
Deterministic code where rules suffice
Validation, calculations and rule-based routing cost no tokens and are fully testable. Use the model only where judgement is needed.
Model routing in numbers, continuing the illustrative example with hypothetical prices: keep 2 of the 5 calls on the larger model (8,000 input and 800 output tokens: €0.02 + €0.008 = €0.028) and move 3 to a smaller model priced at a tenth, €0.25 and €1 per million tokens (12,000 input and 1,200 output tokens: €0.003 + €0.0012 = €0.0042). The task now costs €0.0322 instead of €0.07, and the monthly model bill falls from €700 to €322. Run the evaluation set before and after: a cheaper model that fails more cases can cost more in review time than it saves.
Two more levers. Send less context: retrieve only the passages a step needs and trim history between steps. And ask whether it needs to be an agent at all: if the steps are always the same, a workflow with one or two model calls is cheaper to build, run and audit, which is what our AI automation work often looks like.
The cheapest agent is a fixed-price pilot that tells you early whether the full build is worth it. A "no" at go/no-go costs a pilot budget, not a production budget.
Hidden costs and red flags in AI agent quotes
Quotes for the same agent differ mostly because they include different things. Often missing:
- Your own people's time: a process owner, experts labelling the evaluation set, IT and security reviews.
- Data clean-up before the agent can rely on your documents and records.
- Integration layers and test environments for systems without usable APIs.
- Security and legal review: penetration testing, a data protection impact assessment for personal data, provider contracts.
- Training and change management for the team working with the agent.
- Usage growth: more users, longer documents and more steps raise the token bill.
- Model changes: re-evaluation and prompt work when a version is retired.
Red flags:
- One price for "an AI agent" with no named task, systems or success metric.
- An accuracy figure promised before anyone has seen your data; no evaluation set in scope.
- No estimate of run cost or cost per task; or "usage included" in a flat fee with no stated cap.
- Write access with no approvals, spend limits or audit logs in scope.
- No mention of prompt injection for an agent that reads emails, uploads or web pages.
- Code, prompts and evaluations owned by the vendor, or running only in its account.
- A chatbot sold as an agent. Gartner warned of this "agent washing" and estimated that only about 130 of the thousands of vendors claiming agentic AI offer real agentic capabilities.
What a complete AI agent quote includes:
- The task, its owner and the success metric against today's baseline.
- Every system the agent touches, read or write, and who approves what.
- The evaluation set: how many real cases, who labels them, the pass mark for go-live.
- A pilot phase with a fixed price and a go/no-go criterion.
- Security measures: least-privilege tools, injection defences, step and spend limits, logging.
- A run-cost estimate per task and per month, with the assumptions shown.
- Maintenance after go-live: what is included, at what rate.
- Ownership: code, prompts and evaluations in your repositories and cloud accounts.
For help scoping before you ask for quotes, our AI consulting work starts there: choosing the task, checking systems and data, costing both bills.
Build vs buy an agent platform
You may not need to build. Gartner predicted in August 2025 that 40% of enterprise applications would feature task-specific AI agents by the end of 2026, up from less than 5% in 2025, so check what your existing tools now include. The three routes cost money differently.
| Agent features in your existing software | Agent platform, configured | Custom-built agent | |
|---|---|---|---|
| Upfront cost | Low: switch it on, configure | Configuration and integration work | Engineering build (ranges above) |
| Ongoing cost | Per seat, per conversation or resolution, or usage credits | Platform fees, often usage-based, plus model costs | Model usage at provider prices, hosting, maintenance |
| Fits when | The task lives inside one tool | Common patterns across a few systems, moderate customisation | Core processes where the value is in your data, rules and integrations |
| Cost risk | Pays for features that do less than promised | Usage fees that grow with volume; lock-in of flows and prompts | Build and upkeep for a task not worth it |
Compare routes at your expected volume over two or three years, not at month-one list price: per-resolution pricing that looks small in a pilot can exceed a custom agent's run cost at scale, while a custom build rarely pays off for a low-volume, standard task. Hybrids are common: a platform for the conversation, custom tools behind it via MCP. For the portfolio view, see our AI implementation roadmap.
AI agent cost table and an ROI sanity check
All figures here are illustrative: our rates of €45–65 per engineer-hour and the hypothetical model prices above.
| Cost line (illustrative) | Type | How to estimate | Illustrative figure |
|---|---|---|---|
| Pilot | One-off | 2 engineers × 3–4 weeks | €10,800–20,800 |
| Build, Level 1 / 2 / 3 | One-off | Engineer-weeks × 40 h × rate | €14,400–31,200 / €32,400–62,400 / €54,000–93,600 |
| Model tokens | Monthly | Tasks × tokens per task × provider price | €700 at 10,000 tasks (€322 with routing) |
| Hosting, vector database, monitoring | Monthly | Cloud and vendor price lists | €800 (assumption) |
| Human review | Monthly | Reviewed tasks × minutes × loaded cost | €2,500 |
| Maintenance | Monthly | 15–25% of build per year ÷ 12 | €1,000 (20% of €60,000 ÷ 12) |
Now the sanity check. Take a Level 2 agent handling 10,000 tasks a month, built for €60,000 including the pilot, with the monthly lines above (without routing): €700 + €800 + €2,500 + €1,000 = €5,000 a month.
| ROI check (illustrative) | Figure |
|---|---|
| Time saved per task (review counted separately as cost) | 1.5 minutes |
| Hours saved per month | 10,000 × 1.5 = 15,000 minutes = 250 hours |
| Value per month at an assumed €50 loaded cost per hour | 250 × €50 = €12,500 |
| Net benefit per month | €12,500 − €5,000 = €7,500 |
| Payback on the build | €60,000 ÷ €7,500 = 8 months |
| ROI over two years | Value 24 × €12,500 = €300,000; cost €60,000 + 24 × €5,000 = €180,000; (€300,000 − €180,000) ÷ €180,000 ≈ 67% |
Then stress it. At 1 minute saved per task instead of 1.5, value falls to about €8,333 a month, net benefit to about €3,333 and payback to 18 months. The floor: covering the €5,000 run cost alone takes €5,000 ÷ €50 = 100 hours, or 6,000 minutes, so 0.6 minutes on each of the 10,000 tasks. If the pilot cannot show a saving well above that, the build is not worth it at this volume.
Count only hours that are redeployed or avoid a cost, measured against a recorded baseline. McKinsey's 2025 survey shows why: 62% of organisations are at least experimenting with AI agents and 23% are scaling an agentic system somewhere, yet only 39% report any enterprise-level EBIT impact from AI. For a range on your own scope, try the cost estimator.
Frequently asked questions
How much does it cost to build an AI agent?
Illustratively, at €45–65 per engineer-hour: a single-task assistant with tools €14,400–31,200, a workflow agent with approvals €32,400–62,400, a multi-system agent €54,000–93,600. A pilot of €10,800–20,800 comes first and decides whether the rest is built.
How much does an AI agent cost to run per month?
Add model tokens (tasks × tokens per task × your provider's current price), hosting, vector database, monitoring, human review and maintenance. In our illustrative example of 10,000 tasks a month, tokens are €700 of a €5,000 total.
Why do quotes for AI agents vary so much?
They include different things: integrations, data preparation, approvals, evaluation, security and compliance explain most of the gap. Ask each vendor for the task, systems, evaluation set and run-cost estimate in writing.
Is buying an agent platform cheaper than building?
For standard tasks inside one tool, usually yes. For core processes across several systems, compare total cost at real volume over two or three years; usage-based fees can overtake a custom agent's run cost.
How can we reduce the cost of an AI agent?
Start with one workflow, route simple steps to smaller models, use caching, send less context, use code where rules suffice, and begin with a fixed-price pilot with a go/no-go decision.
Can an AI agent be built at a fixed price?
Yes, when the task, systems and quality bar are written down. Our AI pilots are fixed price with an agreed evaluation set and a go/no-go decision; MVPs start from €15k at a fixed price, and squads from €15k a month suit open-ended work. Rates are on our pricing page.
Get both bills on paper before you build
Tell us the task and the systems involved. We estimate build and run cost and propose a fixed-price pilot with an evaluation set and a go/no-go criterion, within 24 hours of the first call. You own the code, in your repositories and cloud accounts.
Sources
- Gartner, press release, 25 June 2025: Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027
- Gartner, press release, 26 August 2025: Gartner predicts 40% of enterprise apps will feature task-specific AI agents by 2026
- McKinsey, November 2025: The state of AI in 2025: Agents, innovation, and transformation
- OWASP: Top 10 for LLM Applications (2025)
- Model Context Protocol: modelcontextprotocol.io
