The AI Cost Iceberg
Inputs
Low: Single-step, templated tasks (classify, extract a field, route a request).
Medium: Multi-step reasoning or a few tool calls (summarize and decide, draft a short response).
High: Long chains, many tools, or heavy context (multi-doc analysis, agentic workflows, complex approvals).
In-house: Custom-built on foundation model APIs. Highest fixed build cost and engineering/compliance overhead, but the best unit economics at scale.
Low-code platform: Built on a vendor agent/orchestration platform (e.g., Writer, Copilot Studio, Salesforce Agentforce). Lower build cost than in-house since the platform handles orchestration and guardrails; you still configure and integrate it.
Vendor-embedded: A packaged AI feature inside software you already run. Lowest build cost and overhead, least customization.
Flat platform subscription: Pay for the platform; you (or your team) build the agents on it. Cost barely moves with volume.
Per-agent / per-seat: Vendor builds and manages the agents, then charges a monthly fee per agent. Scales mildly with how many agents/use cases you run.
Usage-based / metered: Priced per case or per call. Cheapest at low volume, but scales up directly with how much you run through it.
Independent of your organization's overall AI experience — a carrier that's mature with AI elsewhere can still have one line of business stuck on spreadsheets with no APIs.
Low: Ungoverned, siloed, or manual (spreadsheets, no APIs). Expect real data remediation cost before AI can sit on top of it.
Medium: Reasonably structured, some integration exists.
High: Clean, governed, API-accessible data ready to connect.
Internal only: Employee-facing tool. Baseline compliance and HITL overhead.
Customer-facing: External users interact with the AI directly. Substantially higher compliance/governance burden (disclosures, bias testing, legal review) and modestly higher human-in-the-loop oversight.
Modern: API-first, cloud-native core systems. Standard integration effort.
Legacy: Mainframe, on-prem, or poorly documented core systems. Higher one-time engineering cost and higher ongoing infrastructure/data overhead to bridge to the AI layer.