Essay

Why fraud models should only propose. By Amir Shachar.

Read the essay
Skip to content
Riskernel

Blog · Field notes

Real-Time Fraud Scoring Latency: What a 47 ms Median Actually Means

A practical guide to fraud scoring latency: what the median, 95th and 99th percentile mean in production, why tail latency matters, and how to test vendors.

Futuristic blue 3D low-latency fraud scoring scene with timing rings, fast data streams, and a central risk engine

Fraud vendors love to say they are fast. The problem is that “fast” usually means one cherry-picked number with no context.

If you are evaluating real-time fraud scoring for checkout, instant payments, payout approval, or account takeover flows, the only latency number that matters is the one your customer actually feels when the system is under load.

That is why “47 ms median” is useful only if you understand what sits behind it, and why averages by themselves are usually the wrong way to compare vendors.

Average latency is the easiest number to game

A product demo can produce a beautiful average. Production traffic usually does not.

  • P50 tells you what the middle of the distribution looks like.
  • P95 tells you what happens when the system is a bit stressed.
  • The 99th percentile tells you whether the tail is ugly enough to affect real users and downstream systems.

A vendor can still claim “sub-100ms average” while giving you a 99th percentile that spikes into the high hundreds. That may be fine for a batch workflow. It is not fine for a hot path.

Why tail latency is the real product metric

In fraud, the decision rarely happens alone. It sits inside a longer chain: device signals, enrichment calls, policy logic, auth windows, orchestration, and the user waiting on the other side.

Once the tail gets bad, every other service has less room to breathe. Review queues start to absorb more borderline cases, fallback rules become more aggressive, and the user experiences friction that has nothing to do with model quality.

Scoring target: 47ms median Card network + orchestration budget: ~200-300ms Enrichment provider timeout: 120ms Risk engine tail spike: 800ms at the 99th percentile Result: The average still looks fine. The user experience does not.

Fast enough depends on the flow

Not every workflow needs the same latency budget.

  • Card authorization and instant bank transfers need tight tails.
  • Payout approval has slightly more room, but delays still affect conversion and support load.
  • Manual review routing can tolerate slower decisions, but only if the output is much richer.

The right question is not “is this system fast?” It is “is this system fast enough for our exact flow when traffic is real, inputs are messy, and dependencies are imperfect?”

What to ask a vendor besides one number

If you want an honest answer, ask for these things together:

  • Median, 95th percentile and 99th percentile latency on production traffic.
  • Whether enrichment calls are included in the published number.
  • What happens when one upstream provider times out.
  • How the system degrades when signals are partial or missing.
  • Whether you can test latency in a shadow run against your own traffic shape.

If the answer stays vague, the number is usually marketing, not engineering.

Latency and explainability are connected

Teams often compare speed and explainability as if they trade off automatically. Sometimes they do. Often the real issue is bad product architecture, not physics.

A fast score that no one can trust creates queue work. A beautiful explanation that arrives too late breaks the flow. The useful system is the one that gives you both inside the budget the business can actually support.

If you are evaluating the full stack, the broader checklist is in Fraud Detection API: What to Look For in 2026. If you care about what the ops team sees after the decision lands, read SHAP Explainability for Fraud Ops.

The practical standard

A 47 ms median (Riskernel-reported; single calls vary) is useful because it leaves room for the rest of the transaction path. In an Algogens architecture, this speed is achieved by keeping generative AI out of the live synchronous loop: models propose hypotheses and candidate rules offline, while only pre-validated, deterministic algorithms execute at runtime.

About Riskernel

Riskernel is built on Algogens: generative intelligence with algorithmic certainty. Generative models explore fraud patterns offline, and deterministic validators check every proposal before anything runs live, delivering scoring at a 47 ms median (Riskernel-reported; single calls vary) with per-decision SHAP explanations. If you want to compare latency and decision quality against your current stack, start with a shadow test. Request a shadow test.