AI Will Not Fix a Broken Fraud Program.

AI is frequently purchased as a solution to a problem that has never been stated precisely. "Reduce fraud" is not an objective a model can optimize. Reduce which fraud, in which segment, at what cost to approvals, measured over what horizon — those are objectives.

In the absence of a stated objective, the model optimizes whatever proxy is available in the training data, which is usually the historical decision pattern of the existing program. The result is a system that reproduces the current strategy more efficiently, including its blind spots.

It is worth being precise about the capability, because vague enthusiasm and vague scepticism cause equal damage. There are things machine learning does better than any human process, and they are genuinely valuable.

Notice what these have in common. Each is a perception task — finding, ranking, connecting, comparing. AI is exceptionally good at establishing what is happening. It is far weaker at determining what the organization should therefore do.

A model learns from the data it is given. It has no access to the commercial context in which its outputs will be used, and that context is usually where the real decision lives.

None of this is a criticism of the technology; it is a description of scope. A model can tell you that a transaction resembles previous fraud. It cannot tell you whether declining it is the right commercial decision for this business, this quarter. That judgment has to be supplied by people, encoded explicitly, and reviewed as conditions change.

Supervised models learn from labelled outcomes: this case was fraud, that one was not. The quality of those labels sets a hard ceiling on the quality of the model, and in fraud the labels are structurally compromised.

When a transaction is declined, the organization never learns what would have happened. Every declined customer is recorded as a suspected fraud that was prevented, whether or not that is true. Train on that history and the model learns to reproduce the existing decline pattern, inheriting its false declines as if they were successes.

Fraud that was never disputed, never detected or written off through another process enters the training data as legitimate activity. The model learns that a genuine attack pattern is acceptable behaviour.

If past decisions were harsher toward a particular geography, issuer, payment method or customer type — for reasons that may no longer apply, or may never have been sound — a model trained on those decisions will encode that harshness and apply it consistently at scale. Consistency is a virtue only when what is being applied consistently is correct.

The mitigations are known and mostly operational rather than technical: approve a small randomized holdout to observe outcomes on traffic you would otherwise decline, reconcile fraud outcomes from every source rather than disputes alone, review label quality by segment, and treat label improvement as a permanent function rather than a pre-launch task.

A rule is legible. Anyone can read it, date it and argue with it. A model is a different governance object: it changes behaviour as data shifts, its reasoning is not self-evident, and its failures are quiet.

The population a model was trained on stops resembling the population it now scores. Customer behaviour changes, a new market opens, an attack pattern evolves, an upstream data field changes format. Performance degrades gradually while the model continues producing confident scores. Drift is not a defect — it is the normal condition of a deployed model, and it has to be monitored deliberately.

Someone will eventually ask why a specific customer was declined — a regulator, an auditor, an executive or the customer. "The model scored it above threshold" is not an answer. The requirement is not that every model be simple; it is that the organization can reconstruct and articulate the basis of a decision after the fact.

The value of automation is that decisions happen without human intervention. That is also the risk: an error is applied to the entire population at machine speed before anyone notices. Automated decisioning needs the controls any high-throughput process needs — thresholds on change velocity, alerting on distribution shifts, and a rollback path that can be executed quickly.

The most effective arrangement we see is not humans versus models, and not humans replaced by models. It is a division of labour along the line each is actually good at.

The model handles scale, consistency and detection. People handle objectives, trade-offs, edge cases, novel attacks with no historical precedent, and the accountability that has to attach to any consequential decision. Experienced analysts also remain the fastest source of insight into what is new — they notice the thing the model has no examples of yet.

This is why analyst-assistance applications frequently deliver better returns than full automation projects. They compound the capability of the people who already understand the business, rather than attempting to encode that understanding before it has been articulated.

These are not preconditions imposed by caution. Each one directly determines whether a model can produce a business result.

Data must be complete, consistently defined and joined across the systems where fraud, payments, customer and dispute information live. If the same customer cannot be identified across those systems, no model can reason about them coherently. Fragmented data is the single most common reason AI projects underperform.

State what the model is meant to improve, in business terms, before selecting a technique. Loss rate in a specific segment. Review volume at held accuracy. Approval rate on returning customers at a stated loss tolerance. An undefined objective guarantees an unevaluable result.

Every automated decision needs an accountable human owner and a path to explanation. This is both a governance requirement and a practical one — unowned models are never retired, never challenged and never improved.

Outcomes must flow back into the system: what happened to approved transactions, what review confirmed, what disputes arrived, and — through a holdout — what happened to some of what would have been declined. Without a feedback loop, a model cannot learn and the organization cannot tell whether it is working.

Deploy against a control population, compare on matched segments, and change one thing at a time. Then monitor continuously — inputs, outputs and business results — with a defined response when any of the three moves.

With those foundations in place, several applications are consistently worth the investment. All of them are in production somewhere today, and none requires removing humans from the decision.

Learning a baseline of normal behaviour and flagging deviation, which catches emerging patterns before anyone has written a rule for them. Particularly valuable against attacks that are novel rather than high-volume.

Determining that multiple accounts, devices, addresses and payment credentials belong to the same actor. This is foundational: most other analysis improves once the entity view is correct.

Mapping relationships between actors to reveal coordinated activity that appears unremarkable transaction by transaction. Organized abuse is a structural problem and responds to structural analysis.

Ranking the review queue so limited analyst capacity is spent where it changes the outcome. Often the fastest return available, because it improves results without changing any customer-facing decision.

Surfacing recurring structures in confirmed fraud that inform new controls — using the model to generate hypotheses that experienced people then evaluate and codify.

Assembling case context, summarizing history and highlighting the relevant signals so an investigator starts from an informed position. It shortens decision time and improves consistency while leaving accountability exactly where it belongs.

The question worth asking is not whether to use AI. It is whether the system around it is ready to convert its output into a better decision.

AI is one of the most valuable tools available to a modern fraud program. It is simply not a substitute for knowing what the program is trying to achieve. If it would be useful to assess whether your decision system is ready to get real value from it, we are happy to have that conversation.