AI Assistant
How does Finance AI track AI usage costs?
Understand the admin Agent Costs dashboard, agent tags, per-user totals, and cache-aware cost estimates.
Quick answer
Finance AI records each user-triggered AI call with a required user ID, an enum agent tag, token usage, cached input tokens, Evlog metadata, and Finance AI's own cache-aware cost estimate.
Steps
- 1Open the Finance AI admin dashboard.
- 2Choose the Agent Costs tab.
- 3Select a range such as 1d, 7d, 30d, or 90d.
- 4Review tracked cost, average cost per user, users with usage, Evlog estimate, cache delta, and token totals.
- 5Use the By Tag, By Agent, By User, and By Model tables to isolate cost drivers.
- 6Select a user in the By User table to see that user's cost breakdown by agent and enum tag.
- 7Open Recent Calls to inspect the latest agent, tag, model, user, tracked cost, and Evlog estimate.
- 8For onboarding comparisons, open Agent Evals, choose the OpenAI, Grok, or DeepSeek model, and run the same fixture for each model before comparing accuracy, tokens, and cost.
What gets recorded
Usage rows are tied to a user ID and a typed agent tag, such as Web Chat, Mobile Onboarding, Nightly Sync, Insights, Memory, Support, or Context Compaction. This lets the admin dashboard break cost down by both product surface and individual user.
Average cost per user is calculated from Finance AI's tracked cost divided by the number of users with usage in the selected range. The user breakdown lets admins click into one user and compare which agents and tags drove their cost.
Finance AI stores model token usage, cached input tokens, uncached input tokens, output tokens, reasoning tokens, local cost totals, Evlog AI metadata, and Evlog's estimated cost when it is available.
Why tracked cost can differ from Evlog
Evlog estimates model cost from input and output tokens. Finance AI also stores its own cost estimate because cached input tokens can be billed differently from ordinary input tokens.
The Cache delta metric compares Evlog's estimate with Finance AI's tracked cache-aware cost for the selected range. A larger delta usually means the model reused cached input.
Comparing onboarding models
The Agent Evals model selector runs the same onboarding fixture through OpenAI GPT-5.4, xAI Grok 4.3, DeepSeek V4 Flash, or DeepSeek V4 Pro. Each run keeps its selected model for the full live eval chat so the accuracy and cost comparison is consistent.
An eval's Total cost is the complete end-to-end cost, including context compaction when it runs. The model cost breakdown separates the primary onboarding model from supporting calls so admins can verify cached input, uncached input, output tokens, and cost for each provider and model.
Finance AI can configure a provider experiment globally for text agents, including web chat and onboarding, nightly categorization, Weekly AI Insights, long-chat context compaction, and small structured extraction helpers. Mobile onboarding and regular mobile chat can override that global selection independently, and each user stays on one stable configured variant.
When chat or onboarding runs on Grok or DeepSeek, Finance AI keeps that provider as the primary agent and exposes web search as an ordinary function tool. Only the isolated search helper uses OpenAI GPT-5.4 mini with hosted web search, then returns a concise sourced result to the primary model. Weekly AI Insights on alternative providers remain grounded in Finance AI's transaction packet without optional merchant web research.
Speech transcription and semantic memory embeddings remain on OpenAI because the configured DeepSeek chat provider does not expose compatible transcription or embedding models. If no provider override is configured, each text feature keeps its prior OpenAI model as the fallback.
Development builds show account-test controls directly on the mobile Connect your accounts onboarding page. Admins can add selected production-source test accounts to the current onboarding user or open Plaid Sandbox and add a sandbox account in place. These actions do not reset authentication, restart onboarding, or change onboarding progress, and the controls remain hidden from production builds.
Production-source accounts added to a development onboarding are local snapshots with cloned balances and transactions. Finance AI treats them as already synced and never sends their placeholder access tokens to Plaid.
Before enabling an alternative provider for real users, configure its production API key and confirm the vendor data-processing terms, retention settings, and Finance AI privacy disclosures are approved for transaction data.
Review categorization accuracy, action accuracy, missed rows, retries, completion status, elapsed time, and total tracked cost before replacing the baseline. Lower token price alone does not prove equivalent onboarding behavior.
Frequently asked questions
How does Finance AI track AI usage costs?
Finance AI records each user-triggered AI call with a required user ID, an enum agent tag, token usage, cached input tokens, Evlog metadata, and Finance AI's own cache-aware cost estimate.
What gets recorded?
Usage rows are tied to a user ID and a typed agent tag, such as Web Chat, Mobile Onboarding, Nightly Sync, Insights, Memory, Support, or Context Compaction. This lets the admin dashboard break cost down by both product surface and individual user. Average cost per user is calculated from Finance AI's tracked cost divided by the number of users with usage in the selected range. The user breakdown lets admins click into one user and compare which agents and tags drove their cost.
Why tracked cost can differ from Evlog?
Evlog estimates model cost from input and output tokens. Finance AI also stores its own cost estimate because cached input tokens can be billed differently from ordinary input tokens. The Cache delta metric compares Evlog's estimate with Finance AI's tracked cache-aware cost for the selected range. A larger delta usually means the model reused cached input.
What should I know about comparing onboarding models?
The Agent Evals model selector runs the same onboarding fixture through OpenAI GPT-5.4, xAI Grok 4.3, DeepSeek V4 Flash, or DeepSeek V4 Pro. Each run keeps its selected model for the full live eval chat so the accuracy and cost comparison is consistent. An eval's Total cost is the complete end-to-end cost, including context compaction when it runs. The model cost breakdown separates the primary onboarding model from supporting calls so admins can verify cached input, uncached input, output tokens, and cost for each provider and model.