Case Study: Agent Unit Economics for a Wealth Advisor Assistant

Asset & Wealth Management

Agent Unit Economics for a Wealth Advisor Assistant

The Challenge: A Successful Pilot That Cost Too Much to Scale

A wealth manager had piloted an agentic assistant with 400 advisors. It answered client and portfolio questions by working through research, account data, product documentation and CRM notes, and advisors rated it highly. Each query involved many model calls and a large amount of retrieved context, and the pilot was costing $2.40 per query. At that price a rollout to all 6,000 advisors would have cost more than the business case allowed, and the program was at risk of stopping at the pilot.

Our Solution: Price Each Step, Then Redesign the Expensive Ones

Lydatum instrumented the assistant to show what each step of a query cost, then reworked the steps that accounted for the bulk of the spend:

• Step-Level Cost Tracing: Every model call and tool call was traced and priced, which showed that two of the assistant's nine steps accounted for 64% of the cost.
• Model Routing: Planning and final synthesis stayed on the most capable model, while classification and extraction moved to smaller models that passed the same evaluation suite.
• Context & Caching: Retrieved context was trimmed to the passages each step needed, and prompt caching was applied to the instructions and reference material repeated on every query.
• Cost Guardrails: Each query type was given a cost ceiling with alerts, and unit cost now appears on the same dashboard as usage and advisor satisfaction.

Each change was tested against a set of 1,200 real advisor questions with graded answers, so the team could confirm answer quality held before anything was released. The wealth manager's own platform team took part in every step and now runs the evaluation and cost reviews itself.

The Impact: A Rollout the Business Case Could Support

Ten weeks after the work began, the cost of a query had fallen from $2.40 to $0.70.

Answer quality on the evaluation set stayed within one point of the pilot baseline. The firm approved the rollout to all 6,000 advisors and completed it in the following quarter, within the budget set for the program. Cost per query is now reported monthly next to adoption, and any new assistant feature has to state its expected unit cost before it is built.

Services Delivered: AI Unit Economics & ROI, Model Routing & Cost Optimization, AI Cost Visibility & Attribution, Agent Evaluation & Observability

Technologies Used: Anthropic Claude on Amazon Bedrock, LangGraph, OpenTelemetry, Langfuse, Snowflake, Python