ProSight Financial Association surveyed 125 North American banking professionals in 2026, and 85% said fraud is changing faster than they can respond. Only 3% said banks are ahead of fraudsters in the use of AI. For a lender approving loans in minutes, every new model has to prove it can cut losses without burying analysts in more alerts.
Quick Answer
AI moves the needle most in four places: scoring applications at onboarding, checking documents for tampering, linking applications that share devices or contact details, and ranking alerts for review. It helps least when data quality is poor, alert queues are already overloaded, or a lender works alone. Survey data suggests data quality and shared intelligence limit results more than the choice of model.
Key Takeaways
- AI is most useful for risk scoring, document-tampering checks, network or link analysis and prioritising alerts.
- Rules remain useful for known patterns and policy controls; models add value where fraud patterns change or combine many signals.
- A lender should measure confirmed fraud caught, false-positive rate, review time, alert backlog and prevented loss rather than count model alerts.
- Data quality, external intelligence, reason codes and analyst capacity determine whether an AI layer improves outcomes.
Fraud Is Scaling Faster Than Detection
Deloitte’s Center for Financial Services projected in 2024 that generative AI could push US fraud losses from $12.3 billion in 2023 to $40 billion by 2027. That is a 32% compound annual growth rate. Although a relatively old forecast, it sets the scale lenders plan against.[SOURCE]
ProSight’s survey shows the defender’s side of that trend. Of respondents, 73% said fraud is growing, and 87% said AI is raising fraud faster than it is raising prevention. Another 33% said fraudsters are substantially ahead in AI, and 54% said somewhat ahead.[SOURCE]
Money is not the main constraint. Some 53% raised fraud budgets by 5% or more this year, yet 85% still said the threat outpaces their response. The top barriers, rated acute or moderate limits, were alert overload (66%), staffing (65%), and data quality (63%).
Those barriers determine whether AI helps. A model adds value only when it improves detection or prioritisation without creating an alert queue the control team cannot clear.
Where AI Moves the Needle in Lending
Respondents could name more than one use, so the shares below add up to more than 100%. ProSight found AI is cited most often in transaction monitoring (67%) and onboarding (52%). Risk scoring (43%), customer decisioning (32%), and case management (29%) follow. The table maps those uses to a lending workflow.
| Function | What AI adds | Main limit |
| Onboarding checks | Scores applications on device, contact, and behavior signals alongside ID verification | Weak when the identity data itself is manufactured |
| Document review | Flags tampering through image and metadata analysis | Needs regular updates as generated documents improve |
| Link analysis | Connects applications that share devices, phone numbers, addresses, or accounts | Needs data from outside one lender’s own records |
| Behavioral monitoring | Flags unusual session behavior | Can raise false positives for genuine applicants |
| Alert triage | Ranks cases so analysts review the highest risk first | Does not fix poor input data |
| Disbursal monitoring | Scores payments in real time before funds leave | Instant payments leave little room for review |
Survey comments show what this looks like in practice. A community bank compliance officer described a check monitoring system that reviews item images and auto-clears legitimate items. That frees analysts for suspicious cases.
A mid-tier bank reported that behavioral biometric monitoring gives it real-time alerts against account takeover.
Where AI Adoption Stalls
ProSight also asked what limits AI results. Data reliability led at 48%. Talent and expertise followed at 37%, then technology integration at 36% and regulation or explainability at 36%.
Only 18% of respondents said technology spending has substantially changed how they fight fraud. Another 60% reported a moderate change.
Explainability matters in lending because applicants and regulators often ask why a loan was declined. A model that flags fraud without a usable reason code creates review work and compliance risk. Lenders should ask for reason codes before they buy any scoring model.
Alert volume is the other practical test. If a model doubles flags and the team cannot double reviews, detection gains turn into backlog.
Shared Intelligence and Human Review Still Count
Fraud rarely stays at one lender. Yet only 24% of ProSight respondents share incident information through informal channels, and 20% use formal networks. Nearly two-thirds said they want deeper collaboration.
ProSight notes that institutions increasingly see fraud as a network problem. One Kansas bank fraud manager launched a regional group with 21 other institutions and 12 law enforcement agencies. It gives her more than 75 contacts to alert quickly.
A model trained only on one lender’s history cannot see a fraudster working through ten lenders. Consortium and bureau data fill that need.
People remain part of the control. Respondents named customer judgment errors (62%) and employee judgment errors (49%) among the most preventable losses. Lenders should keep trained reviewers on exceptions and escalations.
Evaluation Checklist: Is Your AI Fraud Layer Earning Its Place?
- Does the model use device, contact, and behavior signals alongside ID checks?
- Can the system inspect documents for tampering rather than only reading their text?
- Does it link applications across shared devices, phone numbers, and accounts?
- Does it connect to consortium or bureau data from outside your own portfolio?
- Does every flag come with a reason code that analysts and applicants can understand?
- Is alert volume measured against analyst capacity before go-live?
- Can the fraud layer be updated without changes to the loan origination system?
Bottom Line
AI helps most with onboarding scores, document review, link analysis and alert triage. Results depend on data quality, lawful access to external intelligence and a team that can act on high-risk cases. Lenders should measure confirmed detection, prevented loss, false-positive rates, decision latency and reviewer workload.
Explore how a configurable origination layer can combine identity, device, document and network signals while keeping fraud decisions reviewable.
FAQs
Can AI replace rule-based fraud checks in lending?
Rules catch known patterns cheaply and are easy to explain. Models find patterns that no one has written a rule for. Using both is the more practical setup.
How much can AI reduce fraud losses for a lender?
Independent data on loss reduction is thin. In ProSight’s survey, 18% of respondents said technology spending substantially changed how they fight fraud, and 60% reported a moderate change. Lenders should set their own baseline before any pilot.
What data does an AI fraud model need?
It needs application data, device and session signals, bureau or consortium data, and labeled outcomes from past fraud cases. Data quality or availability was a moderate or acute constraint for 63% of ProSight respondents.
