The honest way to read any AI agent quote.
Nearly all custom software cost comes down to one equation: engineering effort, measured in person-weeks, multiplied by a blended rate, plus infrastructure and model usage. Everything else is a discussion about how many person-weeks the scope really implies. When a quote looks surprisingly low, it is almost always because the scope excluded evaluation, error handling, or the integration work against your existing systems — the three things that consume the majority of the effort on a real agent. Rates vary enormously by region and seniority, so the figures worth arguing about are the effort estimates below. Apply whatever rate you already use for engineering work and you will have a defensible order of magnitude before you speak to anyone.
Scope 1 — Assisted workflow agent: roughly 4 to 8 person-weeks.
A single well-defined task, one or two tool integrations, and a human confirming every action. Drafting responses from a knowledge base, triaging inbound requests, or turning documents into a structured record. The effort goes on retrieval quality over your content, prompt and tool design, an evaluation set, and the interface where a human reviews output. The model work is a small fraction of it. This is the right starting point when a task is repetitive, judgement-light, and already has a human checkpoint you intend to keep.
Scope 2 — Multi-step agent with real tool access: roughly 10 to 20 person-weeks.
The agent plans across several steps and calls real systems — a CRM, a ticketing system, an internal API. It still asks before consequential actions, but it does meaningful work unattended. Effort concentrates in three places teams underestimate: authentication and permissions against systems never designed for a non-human caller; failure handling, because a half-completed multi-step action is worse than one that failed outright; and observability, since you cannot debug an agent whose decisions you cannot replay.
Scope 3 — Production autonomous agent: roughly 20 to 40 person-weeks.
Operates continuously, handles meaningful volume without per-action human review, and has defined authority limits. Cost rises steeply here, and not because of the model. The additional work is approval gating for high-impact actions, audit logging that satisfies whoever owns the risk, rollback paths, rate limiting, cost controls on the model calls themselves, and a regression suite that catches behavioural drift when you change a prompt or upgrade a model. Teams that budget Scope 3 as Scope 2 plus a little extra are the ones whose projects stall at the pilot boundary. The gap between a working demo and a system you would let run overnight is most of the project.
Scope 4 — Multi-agent platform: 60+ person-weeks.
Several specialised agents coordinating, shared memory and state, a routing layer, and per-agent permissions. Justified when you have already run Scope 2 or 3 successfully and the bottleneck is breadth rather than depth. Very few organisations should start here. Almost every multi-agent platform we have seen succeed was grown from a single agent that had already proven its value.
The costs that estimates leave out.
Model usage varies by orders of magnitude with design, not just volume. An agent that re-reads a long context on every step can cost many times one that maintains state properly; prompt caching and routing simple steps to a smaller model are the two changes with the largest effect. Budget for usage measured in a pilot rather than a per-token price from a pricing page. Evaluation is ongoing: an agent is not done when it works once, and skipping the test set does not save money — it defers the cost to whoever discovers the failure in production. Model upgrades arrive frequently and behaviour changes with them; without a regression suite every upgrade is a manual re-verification. And every system your agent touches will eventually change its API, permissions model, or rate limits, which is ordinary software maintenance that does not go away.
Where the money is genuinely well spent.
Evaluation infrastructure, early — it is the difference between improving the agent and guessing. Observability, meaning full traces of what the agent saw, decided and did, which is non-negotiable at Scope 3. Human approval gates on consequential actions, which are cheap to build and are the thing that lets the system be trusted with more over time. And a narrow first scope, because the most expensive AI agent is the ambitious one that never reaches production.
Where it is usually wasted.
Fine-tuning before trying retrieval and prompt design is occasionally correct, usually premature, and always expensive to maintain. Building a framework instead of the product produces an abstraction layer written before the second use case exists, which is almost always wrong. And chasing full autonomy in version one misreads what human-in-the-loop is for: it is not a failure state, it is how a system earns the authority to act alone.
How to get a number you can rely on.
Any credible estimate requires four answers: which task, which systems it touches, what happens when it is wrong, and what volume it handles. If a vendor quotes before asking all four, the number is a placeholder. Run the effort estimates above against your own engineering rate and you will have a defensible order of magnitude before you speak to anyone — usually enough to know whether the project is worth scoping properly.