Muhammad Ikram Let's talk

Your Fraud Filter Is Also a Conversion Filter

5.0% of orders get rejected on average, according to the Merchant Risk Council's 2025 survey of 1,082 merchant professionals. Some of those are good buyers.

Rules are readable but brittle. Machine learning adapts but needs data and is harder to explain. Most teams end up with a hybrid: the model scores, readable rules act on the score, and people review the grey zone.

Pick by your fraud history, how often you must explain a decline, and how fast attacks change.

What screens your checkout today?

01 / 07Payments

Fraud screening at checkout

Stop fraud without
blocking good buyers

Rules, machine learning or both: how to pick the screening model that fits your checkout.

RulesvsML modelvsHybrid

02 / 07Payments

The cost sits on both sides

5.0%average order rejection rateDown from 5.8%
14%see false positives above 10%Good buyers lost
3%use no AI or ML fraud toolsAI is the norm

MRC 2025 Global eCommerce Payments & Fraud Report: 1,082 merchant professionals, surveyed Oct to Nov 2024.

03 / 07Payments

Rules vs machine learning

Rules

  • Every decision readable
  • Fast fix, known attacks
  • Fraudsters learn limits
  • Lists grow and clash

ML model

  • Hundreds of signals
  • Adapts to new patterns
  • Needs labelled history
  • Less explainable

04 / 07Payments

How a hybrid screen decides

1

The model scores it

Stripe Radar scores each payment from 0 to 99.

2

Rules act on the score

Allow, block, send to review or ask for 3D Secure.

3

People review the grey zone

Elevated-risk orders go to a queue, not straight to decline.

4

Outcomes train the model

Marking fraud feeds back into the next score.

05 / 07Payments

The trade-offs side by side

RulesML modelHybrid
ExplainableHighLowerGood
Data neededLittleA lotSome
New attacksSlowAdaptsAdapts
Main costAnalyst timeVendor or data teamBoth, smaller

Author's assessment, not measured results.

06 / 07Payments

Four questions to ask first

1

Do you have labelled fraud history?

No: start with rules or a vendor model
2

Must you explain every decline?

Keep readable rules on top of the score
3

Do attacks change every week?

That is where a model earns its cost
4

Who owns the thresholds?

One named person tunes and reviews them

07 / 07Payments

Your turn

Rules, a model or both: what screens your checkout today?

Comment

Share your setup

Share

With your payments lead

Save

For your next fraud review

Sources: Merchant Risk Council 2025 Global eCommerce Payments & Fraud Report; docs.stripe.com/radar/risk-evaluation. Checked Sep 2026.

More insights

Keep reading

All posts
Carousel

E-commerce & Ops 2 min read

Black Friday Speed Is Built Weeks Before the Traffic

A common trap: a sale campaign works, traffic spikes, and every visit hits the database. Pages lag and carts stall. A cache serves popular pages from memory. The lesson: protect conversion at peak, and clear cached prices the moment they change.

Carousel

E-commerce & Ops 2 min read

Your Search Box Is the Highest-Intent Funnel You Own

A shopper who types a query already knows what they want. Zero results, a missed typo or a slow page sends that intent elsewhere. Deloitte linked a 0.1s faster mobile site to 8.4% more retail conversions. Audit the search log first.