Case Study · May 2026

How AI.FO ran 2,000+ deterministic financial assertions per close — without an LLM touching the books.

And why that pattern became Axiom.

AI.FO was a production CFO-reasoning system. 2,000+ assertions. Sealed snapshots. Audit-independent verification. No hallucinations in the audit trail. This is how it worked — and what Axiom learned shipping it.

Production deployment $12M nonprofit spin-out, Tides Foundation Federal Single Audit required Multi-entity: US · UK · France · Spain
The situation
$12M nonprofit spin-out from the Tides Foundation
Multi-entity treasury: US, UK, France, Spain — each with independent bank accounts, local counsel, and GAAP requirements
Federal Single Audit required annually — full GAAP compliance, ITGC controls, board reporting
Non-technical founder. No engineering team. Built the entire system using Claude + Cursor.

Separating auto-generated artifacts from human commits

When Claude writes a journal entry, it produces an artifact. That artifact needs to be auditable — but it should not be mixed into human-authored records without a clear boundary.

AI.FO's sealed snapshot architecture generates a cryptographic receipt at the moment every automated task completes. The receipt is sealed before the artifact enters any human-facing system. Human commits and auto-generated artifacts are kept in separate git scopes, with the snapshot serving as the provenance record.

[Claude writes artifact]
        ↓
[Artifact sealed at generation time]
        ↓
[HMAC receipt generated]
   skill | timestamp | inputs_hash | output_hash
        ↓
[Artifact enters human-audit scope — provenance verified]
        ↓
[Human commits only provenance metadata, not raw data]

Nightly pressure tests. No auditor involvement required.

Most audit processes are retrospective — you run the audit after the quarter closes. AI.FO ran a continuous audit harness: every night, the system executed the full close checklist against live data and produced a signed Engine Check Audit report.

The report was structured for a Federal Single Audit reviewer. If a reconciliation failed, it failed publicly and in real-time. By the time the auditor arrived, the evidence was already sealed and verified.

Assertion Logic Failure condition Last run
Balance sheet zero-check Assets − Liabilities = 0 Non-zero balance 2026-05-26 23:00
AP aging under 30 days Invoices > 30 days = 0 Any balance > 30 days 2026-05-26 23:00
Grant expenditure ≤ budget Σ grant spend ≤ budgeted Overage 2026-05-26 23:00
JE sequence unbroken JE count = last − first + gaps Gap detected 2026-05-26 23:00
Foreign entity FX rate valid Rate date ≤ statement date Stale rate 2026-05-26 23:00

2,000+ assertions. Every one deterministic. Every one traceable. Every one signed.

LLM never touches financial data. Protected by design, not by policy.

AI.FO's architecture never sent financial data to the LLM. Not a policy choice — a structural one. The LLM received structured metadata and produced structured instructions. Financial data moved through deterministic logic paths only.

The LLM operated on
Which assertions to run
How to format the close checklist
Which accounts needed attention based on prior patterns
The LLM never saw
Account balances
Grant amounts
Payroll figures
Any raw transaction data
Financial Data
        ↓
[Deterministic assertion layer — no LLM]
        ↓
[Structured metadata only → LLM]
        ↓
[LLM reasoning on metadata, not data]
        ↓
[Result sealed → receipt generated → audit trail]

What deterministic assertions look like in practice

Every assertion is a provable statement about the data. No guesswork.

Balance sheet integrity
Assets − Liabilities − Net Assets = 0 — fails if any posting is unbalanced
Beginning balance + Net income − distributions = Ending balance — cross-period consistency
Foreign currency accounts revalued at period-end rate — FX validation
Grant compliance
Σ grant expenditures ≤ grant amount — no over-spend without approval
Grant draws ≤ eligible spend + retained overhead — cash draw validation
Expense allocation % matches grant budget % — allocation consistency
Payroll controls
Gross pay − deductions = net pay ± $0.01 — arithmetic integrity
Payroll tax withholdings match current tax tables — rate validation
PTO accrual ≤ policy max — benefit cap enforcement
Intercompany reconciliation
Receivable entity A = Payable entity B — bilateral matching
Intercompany eliminations sum to zero — consolidation integrity
FX translation at period-end rate for all entities — consistency check

The pattern that became the product

Every AI.FO primitive has a direct Axiom equivalent. Here's the mapping.

AI.FO primitive What it did Axiom equivalent What it does
Sealed snapshot Receipt generated at artifact creation, before human review HMAC receipt Cryptographic attestation generated at invocation time, before output is returned
Audit-independent harness Nightly signed test run — verifiable without auditor present /verify endpoint Any compliance team or regulator verifies receipts without Axiom's infrastructure online
Deterministic assertion layer Financial data never reaches the LLM — structural isolation Timing-safe HMAC verification Receipt signature verified with constant-time comparison — no timing oracle exposure
Regulated DNA Federal Single Audit, GAAP, multi-entity treasury, ITGC controls EU AI Act / SOC 2 / HIPAA mapping Structured regulatory control mappings: CC7.2, CC7.3, Art 12, Art 13, §164.312(b)
2,000+ assertions Every close-cycle check provable and traceable Per-receipt execution trail Every skill invocation produces a signed, queryable receipt

4 signed LOIs from organizations across regulated sectors — before Axiom existed.

AI.FO's approach attracted inbound from organizations that needed what Axiom now delivers as a product:

Nonprofit healthcare foundation — Federal Single Audit + HIPAA exposure
European financial services firm — DORA + EU AI Act requirements
Legal technology company — attorney-client privilege audit requirements
Regulated education nonprofit — Federal funding + state compliance requirements

Names withheld under NDA. Letter counts and sector exposure are the only specifics that can be disclosed.

Get started

Run this pattern in your stack.

90-day pilot. Full cryptographic receipt infrastructure from day one.

If you're deploying AI agents in a regulated environment — finance, healthcare, legal, critical infrastructure — you need this. Not as a feature. As a structural requirement.