Skip to content

Financial System — money + AI-usage economy

Guide

Financial System — money + AI-usage economy

Plan-phase design (not built). Owner: Finance (planned unit — stood up when we move Plan → Build). Captures the principle that AI utilization is a first-class economy alongside money.

Principle

AI utilization is an economy unto itself. We run two parallel, closely integrated ledgers — a Money ledger and an AI-Usage ledger — sharing one interface and reconciled through a cross-rate.

The two ledgers

Money ledger

  • Tracks: revenue, costs, budgets, spend approvals (subscriptions, tools, sponsorships, payouts).
  • Unit: USD.

AI-Usage ledger

  • Tracks: input/output tokens, requests, model, agent/unit, venture & phase, tool calls, cache hits, wall-clock.
  • Units: tokens (derived: requests, compute-time).

Cross-rate (the integration)

A configurable $/unit rate (e.g. $ / 1M tokens, by model) converts AI usage → money so the two ledgers reconcile into a single “cost of work” view and a per-venture / per-agent P&L (money + AI).

Data model

Both ledgers share the same 12-column transaction shape so they reconcile; usd_equiv is the bridge (money rows carry the dollar directly, AI rows compute it qty × rate from finance/rates.csv). The column order below is frozen — it is the committed header row of each ledger CSV. Freeze it before multiple authors transcribe, or the two ledgers drift in shape (knowledge: ledger-field discipline). A row missing source, ts, or usd_equiv is a reliability defect, not a minor gap.

transaction { id, ts, ledger: money|ai, unit_or_agent, venture, phase,
              type, qty, unit, usd_equiv, source, note }
budget      { scope: unit|venture|phase, scope_id, money_cap, ai_cap, period, note }
rate        { model, usd_per_unit, unit, effective_from, note }
approval    { id, ts, requestor, amount, scope, status, rationale }

The committed CSV header rows (source of truth for column order):

File Header row
finance/ledger-money.csv id,ts,ledger,unit_or_agent,venture,phase,type,qty,unit,usd_equiv,source,note
finance/ledger-ai.csv id,ts,ledger,unit_or_agent,venture,phase,type,qty,unit,usd_equiv,source,note
finance/rates.csv model,usd_per_unit,unit,effective_from,note
finance/budgets.csv scope,scope_id,money_cap,ai_cap,period,note
finance/approvals.csv id,ts,requestor,amount,scope,status,rationale

Spend-materiality threshold (Principal-approved 2026-06-01)

Finance auto-approves spend below the threshold and escalates to the Principal at or above it:

  • One-time spend: auto-approve if under $50 (USD). At or above $50, Principal approval is required.
  • Recurring spend: auto-approve if under $20/month (USD). At or above $20/month, Principal approval is required.
  • Non-USD / credit spend (e.g. Vercel deploy credits, token spend) converts to USD via the cross-rate (finance/rates.csv) before the threshold is applied.

This is the numeric boundary for the “never self-approve material spend” guardrail. It lives canonically here and is recorded as a policy row in finance/budgets.csv (scope org, scope_id materiality-threshold).

Budget approach — track-first (Principal-approved 2026-06-01)

No hard budget caps are set yet, and that is deliberate, not missing. The org stands up the ledgers and logs real actuals first — Vercel deploy credits (ADR-0009), AI token spend (GR-1 top-tier models), Neon Postgres (ADR-0014) — and sets hard caps after ~1 month of accumulated data (review target: ~2026-07). Until then:

  • finance/budgets.csv carries no money/AI caps (the materiality-threshold row is a policy marker, not a spend cap).
  • “Burn vs. budget” reporting states “no cap on file — track-first phase” rather than inventing a denominator.
  • The spend-materiality threshold above is the only active spend control during the track-first phase.

Reporting

  • Daily (in the PA’s day’s report → Notion): spend + AI usage, burn vs budget.
  • Per-venture P&L (money + AI), per-agent/unit usage, model mix.

Governance

  • Finance owns this; budgets are approved by the Principal; spend approvals follow the staffing/finance policies.
  • AI-usage data is operational — it never contains the Principal’s personal information (PA-only; see the confidentiality policy).

Status

Track-first build, 2026-06-01. The five finance/ ledger files are stood up with frozen header rows only (no fabricated data) and the Principal-approved spend-materiality threshold recorded as a policy row in finance/budgets.csv. Real actuals are logged as they accrue; hard budget caps are deferred ~1 month (see Budget approach — track-first). Later: a Finance module in the ERP (see docs/erp-plan.md).