Recommended Free Tools
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Banks can use AI to process fraud signals at scale, but automation alone does not decide what to do when a signal is ambiguous or an action could seriously affect a customer. In a May 25, 2026 interview, Vittesh Sahni, Senior Director of AI at Coherent Solutions, argues for a division of labor: machines handle speed and volume, while people retain judgment over consequential decisions. That is a strategy recommendation, not proof that any particular system improves fraud outcomes.
What Sahni means by human-augmented AI
Sahni’s proposal is not to replace conventional controls with an AI that acts without supervision. It is to combine automated analysis with human review and clear institutional responsibility. AI can sift high volumes of transactions and flag patterns; investigators can assess cases where context is incomplete, the model’s signal is uncertain, or the next step carries a substantial cost for the customer.
In the interview, Sahni summarizes the division this way: “Machines handle the speed and the volume. People stay in charge of the judgment calls.” The principle is useful as a design goal, but it does not specify a universal staffing ratio or dictate which decisions may be automated under every jurisdiction’s laws.
Where rules, AI, and hybrid systems differ
Sahni contrasts fixed rules with models that can identify less obvious or changing patterns, while favoring a combination of the two. A bank’s choice should reflect its data, operating constraints, and the consequences of a mistaken alert—not a simple assumption that one method is always superior.
#1 Best Overall
| Approach | Potential strength | Important consideration |
|---|---|---|
| Rules-based | Explicit conditions can be easier for staff to inspect and explain when the rule is clear and stable. | Rules need review as circumstances change; Sahni argues that static controls can struggle to keep pace with shifting fraud patterns. |
| AI-driven | Can analyze large volumes of signals and identify patterns that are less straightforward to encode as fixed rules. | Performance depends on suitable data and validation. Staff need enough explanation and context to assess model outputs. |
| Hybrid | Can use rules for fixed requirements and AI to surface evolving or less obvious patterns, with people reviewing selected cases. | Requires defined ownership, escalation paths, monitoring, and a way to feed investigation outcomes back into the system. |
This is a practical comparison, not a guarantee that a hybrid setup will outperform either alternative. The Bank of England and Financial Conduct Authority’s 2019 UK survey reported machine-learning use in fraud detection and anti-money-laundering work, alongside safeguards including alerting and human-in-the-loop mechanisms. The survey collected 106 responses and the regulators cautioned that its findings were not statistically representative of the entire UK financial system. Read the Bank of England and FCA report.
Match human review to the impact of the action
Not every alert calls for the same intervention. Sahni recommends a risk gradient: a reversible request for verification may be suitable for automation, while a payment hold or account freeze deserves more scrutiny as its potential harm and difficulty of reversal increase. He puts the principle succinctly: “The bigger and harder-to-undo the decision, the more a human needs to be involved.” Treat this as his design recommendation, not a universal legal rule.
- Verification: Ask for additional confirmation when it can resolve uncertainty without immediately blocking access.
- Temporary review or hold: Escalate for assessment when a transaction warrants investigation and delay is proportionate to the risk.
- Account freeze or relationship-ending action: Apply heightened review and ensure an accountable decision-maker understands the evidence and consequences.
Governance remains the institution’s responsibility even if software produced the recommendation. In a speech delivered in Mumbai on September 11, 2026, Reserve Bank of India Deputy Governor Shirish Chandra Murmu said: “Responsibility rests with the regulated institution, and boards and senior management must understand the models they deploy, their limitations and the consequences of their use.” His remarks concern Indian institutional accountability; they do not establish a blanket prohibition on automated decisions in every country or for every type of rejection. Read Murmu’s speech hosted by the BIS.
Reducing false positives without weakening protection
A false positive can mean a legitimate customer faces a declined payment, extra checks, or a disrupted account. Sahni recommends improving transaction context, using investigation outcomes as feedback, and choosing graduated interventions rather than defaulting to a hard block. These measures can help teams distinguish suspicious activity from unusual but legitimate behavior; they still require testing and monitoring to ensure that changes do not simply shift risk elsewhere.
Rank #3
Sahni also cites performance figures in the interview: up to 80% fewer false alarms in some recent deployments; two to four times more suspicious activity found and more than 60% lower overall alert volume for AI-based anti-money-laundering tools. The interview does not identify the underlying deployments or studies, so these are attributed claims, not independently verified industry benchmarks. They should not be used as expected results for a bank’s own implementation.
What a bank should establish before deployment
The Bank of England and FCA’s 2019 survey is historical and UK-specific, but it underscores why controls around a model matter alongside its detection capability. A practical implementation should establish:
- Data quality and context: Identify relevant data sources, gaps, and integration limits before treating model scores as reliable evidence.
- Validation and monitoring: Check whether the system performs as intended, and watch for changes in data or fraud patterns that undermine its usefulness.
- Explainability for action: Give investigators enough information to assess why an alert was raised and what uncertainty remains.
- Escalation and ownership: Specify which cases require human review, who may authorize consequential actions, and how decisions are recorded.
- Feedback from investigations: Use outcomes to improve controls while checking that feedback does not encode mistaken assumptions or incomplete labels.
Coherent Solutions’ related finance research page describes a broader implementation roadmap spanning agentic systems, biometrics, graph analytics, data strategy, governance, and lifecycle management. It names Sahni and Chief Strategy Officer Shawn Torkelson as contributing experts. This is the company’s own research and should be read as its perspective, not as an independent assessment of the approach. See Coherent Solutions’ overview of AI in financial-services fraud detection.
What the interview does—and does not—show
The interview makes a case for pairing automated scale with human judgment, particularly when decisions are uncertain or difficult to undo. It includes illustrative figures, but no named deployment details or underlying studies sufficient to verify the reported performance improvements. It also mentions an unnamed Midwest regional bank’s onboarding experience without quantified results; that example cannot establish a measurable benefit or be attributed to a specific institution.
Sahni gives an illustrative split in which AI handles 80% of easier alerts and people handle 20% of trickier ones. That is an example from the interview, not a measured universal allocation. A bank would need to set its own thresholds and review arrangements based on its data, risk tolerance, operational capacity, and applicable rules.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
