The four things that move the number
How many systems the agent has to touch. Each integration is real engineering, and the price is set by the worst one, not the average — a carrier with a flat-file upload costs more than five with modern REST APIs. On EZRater we integrated more than forty insurance carriers whose systems ranged from REST to legacy SOAP to flat files, and the adapter framework was most of the work.
How expensive a wrong decision is. A drafting assistant needs light review. An agent that moves money, approves a claim or changes a customer record needs permission scoping, an audit trail and a human in the loop at defined points. That is not a feature you add later; it shapes the architecture.
Whether your data is ready. Scattered documents, inconsistent records and three systems that disagree about who a customer is are usually the long pole. This is the cost most proposals leave out, and the one that surprises people.
How much evaluation you need. Proving an agent behaves is work, and it does not end at launch. A regulated process needs a test set that grows every time you find a new failure.
Build cost is not the whole cost
An agent that is live carries four ongoing costs. Inference: an agent loops, so it makes several model calls per task rather than one, and the bill scales with how often it second-guesses itself. Evaluation: the suite that proves it still behaves has to grow. Model churn: providers deprecate and reprice models, and an agent tuned to one does not always behave the same on its successor.
The fourth is human time. Someone reviews the edge cases the agent escalates. If that queue is ignored, the guardrails stop meaning anything and you are running an unsupervised system you believed was supervised.
Why a pilot costs less than you expect, and production more
A narrow pilot around one measurable outcome is weeks of work, because it touches one system, handles the happy path and runs with a person watching. That is a genuinely useful thing to buy: it tells you whether the decision can be automated at all before you commit a budget.
Production is where the multiplier lives. The same agent connected to real data, real permissions, real failure modes and a real support rota is a multi-quarter commitment. Most of that is not AI work — it is integration, access control, logging and the operational plumbing any serious system needs.
What we do instead of quoting blind
A short assessment: what the decision is, which systems hold the pieces, what has to be provable afterwards, what regulation applies, and what happens today when it goes wrong. That last question usually reveals the real requirement.
The output is a written scope and estimate with stages, costs and risks, based on our seven-stage delivery process, before any build starts. If a workflow automation solves it for less, that is in the document too.
The cheapest agentic AI project is the one you do not build
If your process has few exceptions, a workflow automation is cheaper to build, cheaper to run and far easier to defend when something goes wrong. If the task is reading documents and pulling out facts, that is LLM integration rather than an agent.
We would rather tell you that at the assessment than after a proposal. An agent that was never needed is the most expensive outcome on this page.
