Live benchmarks · Updated hourly

Live benchmarks.
Full dataset open-source.

Most KYC vendors hide behind "contact sales for accuracy." We publish ours hourly — on 1,000 labeled synthetic cases anyone can download and verify.

Last run: Pending first run 1,000 synthetic cases 30 jurisdictions
Next run in: —
FlowAI — current run
—
Overall accuracy
—
OFAC catch rate
—
False positive rate
—
p50 latency
—
p95 latency
$—
Per 1K decisions (Growth)
4-vendor comparison
Metric FlowAI ★ Best Sardine Unit21 Persona
Overall accuracy — Not disclosed Not disclosed Not disclosed
OFAC catch rate — Not disclosed Not disclosed Not disclosed
False positive rate — Not disclosed Not disclosed Not disclosed
p50 latency — 3000ms 5s 2500ms
p95 latency — 8s 15s 6s
$/1K decisions at 50K/mo $— ~$1,450 ~$2,000 ~$300

Competitor latency from published SLAs. Competitor accuracy not publicly disclosed — they don't publish it. Cost estimated from public pricing or contact-sales minimums. Full methodology →

FlowAI per-category accuracy
Clean Approvals
600 cases
—
Low-risk applicants that should APPROVE
OFAC Sanctions
50 cases
—
Matches against SDN / OFAC blocked lists
PEP Matches
50 cases
—
Politically exposed persons — ESCALATE
Document Fraud
100 cases
—
Altered or counterfeit identity docs
Velocity / Structuring
50 cases
—
Structuring and velocity-based AML signals
Geographic Risk
100 cases
—
FATF grey-list + high-risk jurisdictions
Edge Cases
50 cases
—
Multi-signal ambiguous — hardest category

Download the dataset

1,000 labeled synthetic cases. CC-BY-4.0. Use it to test any vendor, retrain a model, or verify our numbers yourself. No account required.

✓ Your download is starting now. Check your email for a copy of the link.
Click here if it doesn't start automatically →

Or grab it directly (no gate):

Direct CSV Direct JSON
Run it yourself
Quickstart docs Node & Python SDKs Get sandbox key
Methodology

How this works

Every hour, FlowAI's triage agent runs against all 1,000 cases in the synthetic dataset. Each case has a ground-truth label (APPROVE / ESCALATE / REJECT) that was assigned based on documented compliance criteria — not by the model being tested. The agent's output is compared to that label and accuracy computed per category.

The dataset

1,000 synthetic KYC cases built for breadth and realism:

Cases span 30 jurisdictions and 5 ID types. Dataset is fully reproducible (deterministic seed). Download to inspect →

Where competitor numbers come from

⚠️ Disclosure: Competitor accuracy numbers are not publicly disclosed by Sardine, Unit21, or Persona. We have not tested their APIs against this dataset — we don't have access, and claiming otherwise would be dishonest. What you see in the table for competitors is: latency from their published SLA documentation, and cost estimated from their public pricing pages or published minimums. Sources:

Why publish this

Most KYC vendors hide accuracy numbers behind an NDA. We think that's backwards. Compliance teams are making high-stakes decisions with these tools — they deserve verifiable numbers. Publishing ours publicly, with the dataset to reproduce them, is how we earn trust. If our numbers are wrong, we want to know.