A focused agent pilot using one tool can begin around ₪8K to ₪15K. A business system with RAG, several integrations, permissions, an interface, and monitoring may range from ₪25K to ₪80K or more. Actions and operational risk influence the price more than the model name.
Level one: an assistant
An assistant reads information, summarizes, classifies, or suggests an answer. It does not modify external systems, making it easier to build and test. The team still needs clear data boundaries, sources, and quality metrics.
Level two: an agent that acts
When an agent creates a lead, updates an order, or schedules a meeting, it needs integrations, permissions, idempotency, approvals, and audit logs. Every external system adds edge cases and maintenance.
Level three: a critical workflow
Financial actions, medical information, customer decisions, and multi-step workflows require extensive evaluation, human review, monitoring, and security. Most of the cost now sits in system engineering and control rather than the prompt.
Recurring costs
Include model usage, embeddings, vector storage, third-party services, hosting, monitoring, and support. Token cost may be small compared with the human time spent resolving exceptions, so measure cost per correctly completed task.
How to reduce cost safely
Begin with one workflow, limit tools, use a smaller model when it performs well, and require approval before adding autonomy. A good pilot proves savings before anyone builds a general agent platform.
What a proposal must include
Ask for use cases, connected systems, an evaluation dataset, acceptance metrics, permission design, estimated monthly costs, code ownership, and failure handling. A proposal listing only a model and prompt is not a production proposal.