AutoFlow
Independent verification for financial AI
Verify financial AI before it influences a credit decision.
AutoFlow checks eligible AI-generated financial claims against source evidence, financial definitions, compatibility constraints, and deterministic calculation rules. Every result is returned with a reproducible verification certificate—or blocked for review.
Source page
borrower_pkg.pdf · p.14
Consolidated EBITDA
$42,100,000
FY2024 · USD
Evidence ev_14b
Calculation trace
Net leverage = 3.2x
Typed claim · FY2024
// exact decimal
NetDebt / EBITDA
= 3.200000
assertion match
Finding
- ✓Definition bound
- ✓Period / currency OK
- ✓No conflicts
- ✓Replay metadata set
Fact IDs
f_ebitda · f_netdebt
- 12 deterministic rules
- Claim-level evidence
- Exact decimal arithmetic
- Reproducible certificates
The problem
The right documents do not guarantee the right answer.
Financial teams can give an AI the right documents and still receive the wrong number, wrong period, wrong definition, or an answer that hides conflicting evidence. Retrieval provides context; it does not independently verify how the model used that context.
How AutoFlow works
From evidence to certificate—outside the model.
Build the evidence layer
Extract facts, definitions, periods, currencies, and source anchors.
Compile the AI response
Convert the answer into discrete typed financial claims.
Execute approved checks
Run exact calculations and compatibility rules outside the model.
Release or escalate
Verify, contradict, block, or route the claim for human review.
Build the evidence layer
Extract facts, definitions, periods, currencies, and source anchors.
Compile the AI response
Convert the answer into discrete typed financial claims.
Execute approved checks
Run exact calculations and compatibility rules outside the model.
Release or escalate
Verify, contradict, block, or route the claim for human review.
First use case
Credit Evidence Engine
A read-only verification sidecar for underwriting and covenant monitoring. It checks borrower-package facts, reconciles cross-document differences, recalculates covenant metrics, and generates exception records without making the lending decision.
Explore use casesClaim
DSCR = 1.45x
Package A
CFO = 18.2M
Package B
CFO = 14.7M
Conflicting source anchors for the same period. Calculation blocked; routed for human review with both evidence IDs attached.
Engineering checkpoint
Runtime reliability under stress—not a hallucination benchmark.
12 rules
AutoFlow currently supports 12 deterministic financial and reconciliation rules.
1,000 / 1,000
In a synthetic no-network stress test, 1,000/1,000 jobs completed successfully.
Byte-identical replay
The tested workload produced byte-identical replay artifacts.
Certificate-schema validation and provider-gateway resilience under simulated rate limits, timeouts, retries, and invalid outputs are covered in the stress tooling.
These results demonstrate runtime reliability and fail-closed behavior. A real comparative benchmark is still required to quantify hallucination reduction. AutoFlow is designed to prevent unsupported financial claims from being released; a comparative real-model benchmark is in progress.
Read benchmarks methodologyDesign partners
Bring us one repeated credit workflow.
We are working with design partners who handle recurring borrower packages, covenant calculations, and document reconciliation. We will map one workflow and show exactly which claims AutoFlow can verify, block, or escalate.