# Costing an agent system before building it Entry E20. Sector: Internal research. Status: internal. Delivered: 2026-03. Source: https://ai.prospicience.in/work/agent-cost-model Prospicience prices an agent system before building it. Our cost model maps twenty models from eight vendors to the agent roles that need them, budgets twelve agent types, and plans a 70 / 20 / 10 routing split across budget, mid-tier and premium models, so running cost is set by design rather than discovered on the first bill. ## The challenge An agent system's monthly cost is decided by architecture choices made in week one, and by the time the bill arrives those choices are expensive to reverse. ## Why it mattered A system that sends everything to a premium model, or lets responses run uncapped, costs far more than it needs to, and the gap only shows once real volume arrives. For a client, that is the difference between a project that scales and one that gets switched off after the first invoice. ## What we built - Priced twenty models from eight vendors, each mapped to the agent role it suits. - Built a budget planner covering twelve agent types, from orchestrator to classifier. - Set the routing split we plan against: most volume on budget models, a smaller share on mid-tier, and premium reserved for the work that needs it. - Modelled cost at three stages of growth, from first version to scale, so the architecture matches the volume it will actually see. ## The result - We price every agent role before a build, instead of averaging one guess across the whole system. - Budget ranges for three growth stages are ready to share before a client commits. - Caching and batch discounts are planned in from the start, not discovered later. - Every agent call carries an output cap by default. ## Built for trust Model choice is a cost decision made per role before the build, not a preference applied afterwards. Every agent call carries a maximum output limit, because output costs three to ten times more than input, and a capped response keeps every charge predictable. ## AI at work Plans how AI workers share a system's daily load: small, fast models take the high volume sorting, extraction and routing, and premium models are kept for the hard reasoning. Every task runs on the cheapest model that does it well. What it produces: A price for every agent role, Monthly budgets at three growth stages, A model matched to each job, An output cap on every call. - Work is planned against a 70 / 20 / 10 split across budget, mid-tier and premium models, so running cost is set by design. - Repeated context is cached and non-urgent work is batched, so the routine jobs an agent repeats every day cost less to run. - Every response is capped, because output costs three to ten times more than input, so the bill stays predictable as volume grows. How it works day to day: Before a build, each agent role is priced against the volume it will see. Once live, each task goes to the model its role was matched to, with caching, batching and an output cap on every call, and the client sees the running cost before committing. ## Figures from the delivered system - models priced: 20 across 8 vendors - agent roles modelled: 12 - routing split planned against: 70 / 20 / 10 - growth scenarios: 3 - output cost versus input: 3 to 10 times ## The technology - Work out what an agent system costs before building it - Match each job to the cheapest model that can do it well - Keep the expensive models for the work that needs them - Cap every response, because long answers cost the most ### Technical notes - Per model input, cached input and output pricing, normalised to an effective cost at a realistic input to output ratio - Agent role to model mapping with context requirements stated per role - Caching and batch discounts modelled explicitly, since repeated context and non urgent work are the two largest available savings - Maximum output limits treated as a required parameter on every call rather than an optimisation ## Questions ### How do you estimate what an AI agent system will cost to run? By pricing each role separately before building. An orchestrator, a coder, a retrieval agent and a classifier have completely different context and quality needs, so we model volume per role at the scale the system will actually reach, and the estimate holds up when real traffic arrives. ### What is the single biggest cost mistake? Uncapped output. Output costs three to ten times more than input, so one agent that occasionally writes a very long response can dominate the bill. A maximum output limit on every call is the cheapest control there is, and we set it by default. ### Does using a cheaper model mean worse results? Not when the routing is right. Most calls in a real system are classification, extraction and routing, where a small fast model does the job just as well. Reserving the expensive model for genuinely hard reasoning is what makes the whole system affordable at volume. Capabilities: ai-governance-and-guardrails, ai-agents-and-automation.