- Market
- Saudi Arabia
- Written for
- COO, CIO
- Decision
- Apply AI to fraud, risk and document work under a supervisor who can ask why
- Languages
- EN · AR
Under supervision
What a supervised model owes a regulator and a declined customer
A financial model does not operate in private. Every decision it shapes may be questioned by a supervisor, disputed by a customer, or examined after a loss. These are the obligations that come with that, and a model that cannot meet them is not ready regardless of its accuracy.
A reason for each decision
The model is built so the main factors behind a score can be named, not left inside a number no one can unpack.
What it prevents: A decision the bank cannot defend when it is challenged.
A review path for adverse outcomes
A blocked transaction or a declined application reaches a person who can examine it and reverse it, within a sensible time.
What it prevents: A customer wrongly refused with no way to be heard.
An audit trail of every score
What the model saw, what it decided and who acted on it are logged and retained for the period supervision requires.
What it prevents: An investigation that has no record of how a loss happened.
Monitoring across customer groups
Outcomes are checked so the model does not systematically disadvantage a group the bank must treat fairly.
What it prevents: A bias the bank carries without knowing it is there.
Where the signal comes from
The many inputs behind a single fraud decision
A fraud score is not one clever trick; it is many weak signals weighed together into a decision that then splits three ways. The drawing shows the shape, because a customer who is held for review deserves a process, not a verdict.
Three input streams — the transaction, the customer's pattern, the device and session — converge on a scoring and rules layer that records its reason, then split into three outcomes: allow, hold for analyst review, or block, with a dispute path preserved.
The transaction itself
Amount, time, merchant and channel — the immediate facts of what is being attempted.
The customer's pattern
How this compares to how the customer normally behaves, which is where most fraud shows as a break.
The device and session
Signals about where and how the request arrived, independent of the account it claims to be.
What we do
Scored, with rules and a reason
The signals are weighed into a score, combined with explicit policy rules, and recorded with the factors that drove it — so the decision carries its own explanation.
Allow
The clear majority pass without friction, because a system that stops everyone stops the business.
Hold for review
The uncertain middle goes to an analyst with the reason attached, rather than being blocked or waved through.
Block
Only the clearest cases are stopped outright, and even those keep a path to dispute and reverse.
Precision and recall trade against each other; the right balance is a business decision about the cost of a missed fraud versus a blocked customer, not a setting the model owns.
The cost of a black box
The most accurate model can be the one you cannot use
It is tempting to choose the model with the best score on a test set and deploy it. In a regulated firm that instinct is a trap, because the highest-scoring model is often the one whose reasoning is least legible, and a decision you cannot explain is a decision you cannot defend. When a customer disputes a declined loan, when a supervisor asks why a segment is blocked more often, when an investigator reconstructs a loss, the answer "the model said so" is not an answer. The accuracy that cannot be explained buys nothing it does not later cost.
So we treat explainability as a requirement fixed before the model is chosen, not a report bolted on after. That sometimes means selecting a method whose reasoning a person can follow over one that scores a fraction higher and cannot be read, and it always means the decision is logged with the factors that drove it. The result is a model that is a little less clever on the benchmark and far more usable in a firm that has to answer for what it does.
Our role
What we build and what your risk function keeps owning
- 01
Explainable by construction
We build the model so a reason can be given for each decision, choosing methods your teams can defend rather than the highest score no one can read.
- 02
The human path, built in
We design the review and dispute path alongside the model, so an adverse decision always has a person who can examine and reverse it before it becomes a complaint.
- 03
Evidence your supervisor accepts
We log the decisions, monitor outcomes across groups and document the model so your risk function has what a review will ask for, not a black box it has to defend blind.
The risk appetite, the thresholds and the balance between a missed fraud and a blocked customer are yours to set; we build a model that acts within them, explains each decision, and always leaves an adverse outcome a way back to a person.
Common questions
What buyers ask
Can we use AI for credit decisions if we have to explain every rejection?+
Yes, and the requirement to explain is what shapes how it is built rather than a reason to avoid it. We use methods whose reasoning can be turned into the factors behind a decision — a set of reason codes a customer and a supervisor can be shown — instead of choosing the highest-scoring model that cannot be read. Sometimes that means accepting a model a fraction less accurate on a test set in exchange for one you can defend when a rejection is disputed. The obligation to explain does not disappear because a machine made the call; the bank still owes the reason, so the model is built to produce it.
How does an AI fraud system avoid blocking legitimate customers?+
By treating a hold as a third option between allow and block, and by tuning the balance as a business decision rather than a technical default. A system set to catch every possible fraud will block a stream of genuine customers, which costs more in lost business and complaints than the fraud it prevents. So the clear majority pass without friction, the clearest fraud is blocked, and the uncertain middle is held for an analyst with the reason attached rather than refused outright. Where that line sits depends on your appetite for a missed fraud versus an annoyed customer, and we set it with you and keep watching the outcomes, because the right balance shifts as fraud patterns do.
Does the model replace our risk and compliance teams?+
No — it changes what they spend their time on, and building a system that replaced them is a claim this page will not make. The model handles the volume: it scores the routine cases, clears the clear majority and surfaces the ones that need judgement with the reason attached. Your risk and compliance people then spend their attention on the genuinely difficult decisions, the disputes and the oversight of the model itself, rather than on manually reviewing everything. They also remain accountable for the outcomes, which is exactly why the model is built to explain itself to them rather than to act beyond their reach.
In this cluster
Next step
Scope it with us
A scoping session on one use case — fraud, credit or documents — that fixes how a decision is explained, how an adverse outcome is reviewed, and where the model hands back to a person.

