Novaheim AI
Book a readiness review
Proof, on record.

Your AI drafts the SAR.
Can you prove it’s right?

Independent monthly testing of the AI agent in your SAR workflow — for fintechs, exchanges, stablecoin firms and MSBs. On your own cases, inside your own cloud, with evidence an examiner can follow end to end.

Book a readiness reviewSee a sample finding
  • Read-only, in your cloud
  • SAR data never leaves it
  • Your BSA Officer signs every report
snapshot→sample→check→judge→review→signBackground ledger · synthetic data
ScopeBSA / AML · FinCEN SAR · OFAC sanctions · Model risk review · Sponsor-bank oversight · GENIUS Act
Compatible withUnit21 · Hummingbird · Verafin · FIS · Fiserv · In-house buildsTrademarks of their owners. No affiliation implied.
The problem

The agent is fast. The proof is missing.

650+
institutions have adopted financial-crime AI agents on one vendor platform alone.
Nasdaq
30 days
to file from detection. The clock doesn’t change because an agent wrote the draft.
FinCEN
Excluded
Generative and agentic AI, from the April 2026 model-risk guidance. Proving it works falls to you.
OCC Bulletin 2026-13
What we catch

Agents fail quietly.
These are the three ways.

F-01 · caught by code
The wrong number
An amount, date or counterparty in the narrative that doesn’t match the record it cites.
$45,800 ≠ $48,200
F-02 · caught by AI judge
The missing “why”
A narrative that describes the activity but never says why it is suspicious.
reason for suspicion · absent
F-03 · caught on re-review
The alert that should have been filed
A closed or below-threshold alert the agent let through — the costliest failure of all.
disposition · closed → escalate
What you get

Your exam binder, built every month.

CASE-2026-04125 · Harlow Trading LLC
Agent narrative

On March 2, Harlow Trading received a wire of $48,500 and within four hours sent $45,800 in USDCT4 to an external wallet. Funds reached a high-risk exchange in two hops.

Source records
03-02FIATWire in · ref 88120 · A1$48,500
03-02CHAINUSDC out · 0x7a3f…c91e · T4 ≠ narrative$48,200
03-04CHAIN2 hops → high-risk exchange · C2$47,900
A1, T4 and C2 are the record IDs the narrative cites. Synthetic data.
Check results
Amount mismatchcode
Narrative $45,800. Record T4 $48,200.
No reason for suspicionAI judge
No sentence explains why. Evidence quoted.
Awaiting BSA Officer
D1
Monthly QA report
Risk-weighted sample of drafts and dispositions, checked against source records, with error rates and trends.
D2
Missed-activity review
Closed and below-threshold alerts re-reviewed, including on-chain flows.
D3
Tuning memo
Recommended prompt and rule changes, re-tested once you make them.
D4
Model and prompt register
Every agent version, prompt change, re-test and sign-off, versioned.
D5
Exam binder
Workpapers, sign-offs and a one-page board summary.
How a test runs

Code checks. AI judges.
Your officer signs.

Following CASE-2026-04125 · run 2026-10all inside your cloud
  1. 01
    Your vendor
    Agent output
    “…sent $45,800 in USDC to an external wallet…”
  2. 02
    Code
    Snapshot
    1 of 87 sampled
    risk-weighted · read-only
  3. 03
    Code
    Check
    $45,800 ≠ $48,200
    narrative vs record T4
  4. 04
    AI
    Judge
    reason for suspicion
    absent · evidence quoted
  5. 05
    Novaheim
    Review
    2 flags upheld by a Novaheim analyst
    re-draft recommended
  6. 06
    Your BSA Officer
    Sign
    Signed · 10-14 09:12
    sealed to evidence log
The usual way
Once a year, on the vendor’s model.
1×
With Novaheim
Twelve times a year, on your cases.
12×
Security · inside your environment

SAR data stays inside your walls.

Your environment
cloud · endpoint · logs
SAR workflow
Drafts and case records
Test engine
Read-only role, your IAM
Evidence log
Write-once, yours
In: named reviewers, time-boxed, approved per session · Out: the report, to your BSA Officer
Control sheet
AccessRead-only, under your IAM
ResidencyYour cloud, your model endpoint
Copies heldNone
Model trainingNever
EvidenceWrite-once, owned by you
ExitEngine removed; evidence stays
How we work

Start with a 30-day baseline.

01 · 4 weeks · start here
Baseline sprint
Map your agent, deploy read-only, review 60–120 cases, deliver a baseline report and remediation plan.
02 · Monthly
Assurance retainer
Monthly QA runs, quarterly re-tuning and re-testing, a live model and prompt register.
03 · Annual or on trigger
Exam pack
Full evidence binder and board summary, triggered early by a model change or exam letter.
Baseline delivered within 4 weeks of data access, or work continues at no cost and your first month is free.
30-minute call · by appointment · 3 engagements per quarterBook a readiness review
About
Bradley Behan

Bradley Behan

Founder · Novaheim AI

Bradley founded Novaheim to give BSA Officers and CCOs independent proof that the AI drafting their SARs gets it right, tested every month on their own cases, inside their own cloud. He brings seven years of measuring model performance and false positives, and works directly on every engagement, from the baseline sprint to the evidence your officer signs.

LinkedIn
FAQ

Questions a BSA Officer asks first.

Anything not answered here is what the 30-minute readiness review is for.

Book a readiness review
We already have our vendor’s validation. Why add this?
Vendor validation covers the platform. Monthly testing covers how your firm uses it — your prompts, your typologies and your data.
Do you ever change a disposition or file anything?
No. Access is read-only. Novaheim never drafts, edits, files or closes anything.
If you find a missed SAR, what happens?
Your BSA Officer gets the flag the same day, with the case, the source records and the evidence quoted. What happens next (investigate, escalate or file) is your decision under your own procedures; Novaheim never makes it. Before the first run we agree with you, and your counsel if you choose, how and when flags are handed over, so the timing fits your SAR filing deadlines.
How do you handle SAR confidentiality?
Everything runs in your cloud, on your model endpoint. Nothing is copied to Novaheim or used for training. Access is time-boxed, approved per session and limited to named roles.
What do you install in our cloud?
A read-only test engine that runs under your access controls. It is the same engine for every client, maintained by Novaheim.
How can AI check AI?
Numbers come from versioned SQL against your records. The AI only judges language quality and quotes its evidence. People review every flag, and the judge’s own accuracy is measured and logged.
Which agents can you test?
Any agent whose outputs can be exported or reached through an API — vendor-built or in-house.
How much of our team’s time does it take?
A few hours to start, then two to four hours a month for BSA Officer review and sign-off.
Will this satisfy our examiner?
Novaheim produces the evidence examiners ask for — tested samples, error rates, tuning history and sign-offs — in a binder your firm owns. The examiner’s conclusion remains theirs.
Does this replace our independent BSA test?
No. It feeds into that test, and Novaheim works alongside your auditor.
Who owns the evidence?
You do. If the engagement ends, Novaheim removes the engine and the evidence stays where it is.

An untested agent isn’t an efficiency gain. It’s a finding.

Agents change every release. Examiners ask every cycle. Start with a 30-day baseline.

Book a readiness review
Baseline delivered within 4 weeks of data access, or work continues at no cost and your first month is free.
30-minute call · by appointment · 3 engagements per quarter