Fraud detection
AI-driven fraud detection saved a start-up $2M/year
We built an AI-driven fraud detection system for a start-up, and it saved the business $2M a year. This page keeps to what is on the record: the kind of problem it solves, how we approach a build like this, and that result.
The context
A start-up losing real money to fraud.
The client was a start-up with a fraud problem large enough that stopping it was worth $2M a year. At a company that size, a loss like that is not a rounding error. It is money that was meant for hiring, product and growth.
The answer was AI-driven detection. That choice matters, because it is the difference between a system that can pick up new patterns and one that waits for someone to write them down.
The challenge
Fixed rules fall behind the people they are meant to stop.
Most young companies start fighting fraud with rules: block this country, flag that amount, review anything that looks odd. Rules are quick to write and easy to explain, and for a while they work.
The trouble is that fraud adapts and rules do not. Each new pattern needs someone to spot it, write a rule and ship it, and by then the pattern has often moved on. Tightening the rules to compensate catches more genuine customers too, which costs sales and fills the manual review queue.
That is the kind of problem AI-driven detection is for: patterns that change faster than anyone can write rules for them.
What we built, and how
An AI-driven fraud detection system.
We built the system that detects fraud for the start-up. We are not publishing the client’s details or the system’s internals here, so rather than a week-by-week account, this is how we approach a build of this kind.
-
01
Agree the time budget first
A fraud check that finishes after the payment has gone through is a report, not a control. For a system like this we agree how long a decision may take before any modelling starts, and design everything else to fit inside it.
-
02
Train and score from the same data
We compute the signals a model relies on in one place, for both training and live scoring. It removes the most common reason a model that looked good in testing disappoints in production.
-
03
Make every decision explainable
A score should carry the reasons behind it, so a support agent can answer a customer and a compliance lead can answer a regulator. A model nobody can explain is hard to keep in service.
-
04
Keep rules for the certainties
Some things should always be blocked, and a plain rule is the right tool for those. The model is there for the patterns nobody thought to write down.
The result
Fraud stopped costing what it did.
The system saved the start-up $2M a year. That is the figure recorded for the project and the only one we publish for it.
We have not broken it down or added numbers around it. We would rather give you one result we can stand behind than a page of figures we cannot.
What changed
For a start-up, fraud losses come straight out of the plan.
Money lost to fraud comes out of the budget a start-up uses to grow: people, product and the months before the next round. Getting it back is about as direct a return as software can offer.
It also shows where AI is worth paying for: a system with one clear job and a result you can measure in money.
Related services
More work
Next step
Talk to an engineer about the problem costing you money.
A senior engineer reads every enquiry and replies within one business day. Within a week you have a written approach with a cost and a time range, and a plain answer if we think you should not build it.
