Starting picture
Software analyzes sample accounts, obligations, and spending to establish the planning context.
← Projects
AI implementation case study · Functional demo
An AI financial planning workflow that turns conversation into reviewable decisions.
Steward interprets goals, corrections, and purchase questions, invokes validated planning tools, and helps users explore trade-offs. Deterministic software calculates the amounts and projected dates; users approve changes before they become part of the saved plan.
Scope
Self-directed functional demo using synthetic account data. Opening account analysis is deterministic. No real bank activity, money movement, or financial advice.
5 toolsValidated planning actions
HumanFinal plan approval
255 / 255Automated tests · Sep 6, 2026
System flow
Where AI fits in the product.
Conversation
AI interprets goals, corrections, and purchase questions while retaining the evolving draft.
Tool calls
AI requests validated planning actions and receives application-calculated results.
Calculations
The engine calculates allocations, affordability, and projected dates for scenario comparisons.
Review & approval
The user reviews calculated figures and explicitly approves changes to the saved plan.
AISoftwareAI with human control
Responsibility
AI, software, and human control.
Interpreting intent, asking for clarification, requesting planning tools, and explaining trade-offs using their results.
Argument validation, financial calculations, scenario comparisons, draft state, approval checks, and persistence.
Goals, corrections, priorities, and final approval. Tool calls produce proposals; the user decides what becomes part of the saved plan.
Implementation choices
The decisions behind the demo.
Connect conversation to actions
Five tools let the model read context, propose updates, calculate a plan, compare scenarios, and prepare review. Application results return to the model to inform its next response.
Validate and allow bounded repair
Invalid arguments return feedback so the model can correct them within explicit limits. Unsupported tools are rejected, required steps are enforced, and requests have a timeout.
Keep the plan editable across turns
Users can revise goals, contributions, deadlines, and scheduled bill changes without restarting. For example, a future rent increase affects eligible future bills while retaining the current bill and existing goals.
Separate financial figures from model prose
Application-rendered cards supply calculated amounts and projected dates. The conversation path replaces model prose containing digits, dollar signs, or specified consequential phrases with a neutral follow-up.
Make failure and approval explicit
If the conversation service is unavailable, it reports the failure and preserves the draft. Deterministic calculations remain separate from AI availability; saved plan changes still require user approval.
Evaluate behavior, then refine the workflow
Automated checks cover rules and failure handling. An earlier live-model evaluation exposed interpretation and missing-tool failures; tool descriptions and required-step enforcement were refined before the final 18-case rerun passed.