Claude Financial Services: what the public plugins actually provide
Measuring Financial Services cost: include review and rejected drafts
Build a per-accepted-work-product budget that separates model use, data access and human correction.
What you will learn
- Define the accepted unit
- Separate the bills and the delays
- Investigate failure before scaling
Before you start
- Command-line and JSON familiarity
- A disposable workspace with synthetic financial records
A proposed learning project turns validated requests into reviewable records without giving documents control of routing.
Key takeaways
- Accepted work products are the useful denominator.
- Provider subscriptions are separate.
- Unmeasured cells stay empty.
Define the accepted unit
Use one accepted reconciliation note or meeting brief as the unit of work. Record whether it passed evidence checks without correction, needed revision or was rejected. Counting generated documents alone rewards fluent but unusable drafts.
Keep the synthetic dataset and acceptance checklist fixed across comparisons. A newer prompt that produces fewer words may still increase review time if it drops source-row references. Measure the operator workload as well as generated output.
Separate the bills and the delays
Model calls, provider subscriptions, orchestration retries and human review are different cost sources. The README explicitly says MCP providers may require their own subscription or API key. Repository licensing does not remove those costs.
For each trial record input and output usage when the host exposes it, provider requests, time to first usable draft and reviewer minutes. Leave prices and performance blank until measured under the accounts and versions you actually use.
Investigate failure before scaling
The inspected handoff parser rejects the nested example before routing. Increasing agent concurrency would not repair that extraction path. Diagnose absent handoffs and validation failures before attributing throughput to model speed.
Compare repeated runs on the same sample and report spread, rejected results and configuration changes. No cost or speed benchmark was run for this series, so the worksheet below is an experiment design rather than a savings claim.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Define a draft acceptance checklist.
- 2
Track model, connector and reviewer costs separately.
- 3
Include rejected trials in the final report.
Copy-ready example
trial,revision,accepted,model_usage,provider_calls,review_minutes,total_cost
example,574ed36,,,,,Frequently asked questions
Can I claim a percentage saving from fewer tokens?
Not without measuring total comparable work, including correction and review.
Are the sample cost fields measured values?
No. They are empty fields for your controlled trial.
Sources
- Financial Services / README.mdSource checked 2026-09-23
- Financial Services / scripts/orchestrate.pySource checked 2026-09-23