Short answer: building a production AI agent typically takes from two to twelve engineer-weeks, depending on how many systems it touches, how much autonomy it has and how strict the quality bar is. Running costs range from a few cents to a few dollars per completed task. The ranges below are typical ranges drawn from how these projects are usually scoped — not a quote. Your actual cost depends on your workflow, and we’ll give you a fixed price after a short discovery.
Typical effort ranges by complexity
| Agent type | Example | Typical build effort |
|---|---|---|
| Single-task assistant | Classifies inbound emails and drafts replies for review | 2–4 engineer-weeks |
| Integrated task agent | Reads orders from email, matches them in the ERP, creates drafts, escalates exceptions | 4–8 engineer-weeks |
| Multi-step autonomous agent | Researches leads across sources, enriches CRM, schedules outreach with approvals | 6–12 engineer-weeks |
| Multi-agent system | Planner, specialists and reviewer for complex document or research work | 8–16+ engineer-weeks |
To turn effort into money, multiply by the day or week rate in your quote. Rates for experienced AI engineers vary widely by region and seniority, so treat any figure you see online — including ours — as specific to that provider.
What the build effort actually pays for
It’s tempting to think an agent is “just a prompt”. In practice, the prompt is a small fraction of the work. A production agent needs:
- Discovery and workflow mapping — understanding the real process and its exceptions (10–15% of effort).
- Integrations — authenticated, reliable connections to each system the agent reads from or writes to. This is usually the largest single cost, and it grows with every additional system.
- Tools and guardrails — typed tools, validation, idempotent actions, approval thresholds.
- Evaluation — a test set from real cases and automated scoring, so you know it works and can change it safely.
- Observability — tracing of every run, costs and outcomes.
- Deployment and handover — hosting, monitoring, documentation and training.
What drives the price up
- Number of integrations. Each system adds authentication, data mapping, error handling and testing. Systems without APIs (requiring browser automation) cost more to build and maintain.
- Autonomy level. An agent that only drafts for human review is cheaper than one allowed to act, because acting safely requires more guardrails, testing and approvals.
- Quality bar. Going from 85% to 97% accuracy can cost as much as getting to 85% in the first place.
- Input messiness. Clean, structured inputs are cheap; scanned documents, mixed languages and free-form emails need more work.
- Compliance requirements. Audit logging, data residency, self-hosted models and formal documentation add effort, and they’re worth it when required.
What brings it down
- Start with one narrow, high-volume task, not a general “do everything” agent.
- Use existing APIs and workflow tools where they fit; not everything needs custom code.
- Keep humans in the loop at first — draft-for-review agents ship faster and earn trust.
- Have a process owner available; agents get built faster when someone can answer questions about exceptions quickly.
Running costs: the maths
Running cost per task is roughly:
(input tokens × input price + output tokens × output price) × number of model calls per task + tool and infrastructure costs
As an illustration with stated assumptions: an agent that makes four model calls per task, each with about 3,000 input tokens and 400 output tokens, uses roughly 12,000 input and 1,600 output tokens per task. At a hypothetical price of $1 per million input tokens and $4 per million output tokens, that’s about $0.012 + $0.0064 ≈ $0.02 per task. With a frontier model priced ten times higher, it would be about $0.20. Check current provider pricing — it changes frequently and usually downwards.
Hosting, databases, monitoring and vector storage typically add a modest fixed monthly amount for most business workloads.
Maintenance
Budget for ongoing work: APIs change, models are updated, new edge cases appear. Many teams keep a small monthly retainer for monitoring, tuning and small improvements, or train an internal engineer to own it.
How to get an accurate number
The most reliable way to price an agent is a short discovery sprint: map the workflow, prototype on real data, measure accuracy and cost per task, and then fix the build price. If the prototype shows the agent won’t pay for itself, you’ve spent a little to avoid spending a lot.