Живой финансовый прогноз, построенный на ваших собственных книгах.

УВЕРЕННО

• Driver-based forward model auto-derived from QB + Stripe + HubSpot. Revenue model auto-shapes to your data: cohort × ARPU × retention for SaaS, repeat-purchase rate × AOV for e-commerce, recurring-contract retention for services. OpEx model from headcount × loaded cost + tools + marketing efficiency. 3-statement linked through AR-DSO, AP-DPO, deferred revenue. Driver-derivation runs nightly; operator can override per-driver in plain English.
• Append-only versioned assumption ledger with audit trail. Every assumption ('hiring pace 2/quarter', 'Q3 enterprise close rate 22%', 'churn 3.5%') is a row in an append-only ledger — never updated, never deleted. Each edit creates a new version with author, timestamp, prior value, and the rationale you typed. The model rebuilds from the current assumption-version set on every refresh. CLAUDE.md principle #2 made operational for the Financial Planning & Analysis surface.
• Cohort retention curve auto-fit from Stripe actuals. Reads Stripe subscription events, computes per-cohort retention curves (monthly cohort × month-since-signup → retained %), fits a decay model, and uses the fitted curve as the forward retention driver. Non-SaaS equivalents: repeat-purchase rate decay from QB invoice patterns for e-commerce; recurring-contract retention for services. Operator can override the fitted curve.
• Macro auto-overlay — Fed rate + wages + sector benchmarks. Daily refresh from public sources: Fed funds rate (cost-of-capital input for runway-extension decisions), Carta + Levels.fyi wage benchmarks by role/level/region (headcount cost assumption sanity), SaaS Capital + OPEXEngine sector growth / burn-multiple / NRR norms for SaaS, NFIB SMB index for non-SaaS. The model surfaces every assumption against its macro-overlay range — 'your Q3 enterprise close-rate assumption 28% sits at the 90th percentile of SaaS Capital benchmark — verify or revise.'
• Scenario library — base / conservative / aggressive + custom. Three pre-built scenarios run on every model refresh (base = your assumptions; conservative = −20% growth / +20% costs; aggressive = +20% growth / +20% costs). Custom scenarios accept natural-language input ('what if I raise $2M at 15% dilution and hire 4 engineers') — the scenario builder parses the prompt into assumption deltas you preview before the rerun. Every scenario saves with a name + creation timestamp + cited assumption changes.
• Sensitivity analysis — top-3 assumptions driving runway. On every refresh, the model computes which 3 assumptions would shorten runway most if they moved by ±10% — surfaced as a ranked list with delta-runway in weeks per +10% / −10% move per assumption. The founder sees not just 'runway 47 weeks' but 'runway is 64% driven by your Q3 enterprise close-rate assumption — if that drops 5pts you lose 8 weeks.'
• Monthly AI-narrated assumption-vs-actuals drift. On the day books close (≤35 days post month-end), the Tier-3 cross-family supervised worker narrates the drift between every assumption and what actually happened in plain English. 'CAC payback slipped 14→17 because Q1 cohort retention dropped 2pts; the macro benchmark for your stage is 14 — your prior assumption was at the benchmark, the actual is below it.' Every claim cites a source GL transaction or Stripe event. Lands in /app/fpa + Slack #finance digest channel.
• Slack /runway slash command + Slack-native model Q&A. Real-time read from the latest model state. /runway returns the current runway + the macro context + the top-3 sensitivity-driving assumptions. Free-form Slack Q&A ('what's our gross margin trend over the last 6 months?', 'how much would runway extend if I cut $50K headcount?') returns answers grounded in the live model + actuals. Quoted answers always cite the relevant model row + the source transaction chain.
• 13-week rolling cash forecast with P10/P50/P90 confidence bands. Pulls cash actuals from QB + Stripe + Mercury + Brex + Ramp (read-only); projects the next 13 weeks of cash-in / cash-out from the driver model + AR-DSO + AP-DPO + scheduled payroll. Confidence bands derived from scenario library (P10 = conservative, P90 = aggressive). Surfaces in /app/fpa + emits a weekly Slack digest every Monday 07:00 local-time with the next-13-week chart.
• Append-only audit log + Tier-3 supervisor on monthly narrative. Every assumption change, scenario rerun, model refresh, and narrative emission is logged to the append-only audit_log table with tenant_id, user_id, timestamp, metadata. The monthly drift narrative routes to Tier 3 cross-family supervisor (Anthropic worker → OpenAI reviewer → Gemini tiebreaker on disagreement) — the narrative is the externally-visible artifact, so its guardrails sit at the strictest tier in the marketplace.
• Books-quality gate — refuse-to-run if books not closed in 35d. At onboarding and on every monthly refresh, the worker checks whether the customer's QB/Xero books are closed within the last 35 days, reconciled, and categorized above a minimum quality bar. If not, refuse to run the model — surface the specific gaps + recommend a bookkeeper (Pilot / Bench-refugee / Puzzle partner list). Brand-defense discipline: a narrated drift on garbage books is a credibility bomb, not a deliverable.
• Read-only by API-scope across every connected vendor. QuickBooks / Xero / Stripe / HubSpot / Mercury / Brex / Ramp tokens carry NO write scopes. Cannot post a journal entry, cannot recategorize a transaction, cannot modify a subscription, cannot move a deal stage, cannot move money. Read-only at the API-scope layer, not the prompt layer — schema-enforced refusal.