The AI-Native Bank: Six Processes to Start With
Six banking processes worth evaluating for AI assistance, with evidence from published cases and clear boundaries for human review.
Slow onboarding has a visible business cost. In Fenergo’s August 2025 survey of 600 senior decision-makers at financial institutions in the UK, US, and Singapore, 70% reported losing clients because of inefficient onboarding. That is a vendor survey, not an Israeli benchmark, but it describes a problem worth measuring locally.
For a bank CTO, a practical question is where to start using AI without weakening controls. Camunda’s article on six banking use cases offers a useful shortlist. The examples combine possible agent roles with reported implementation results, which need to be read in context. Suitability in an Israeli bank depends on its systems, data, policies, and applicable requirements.
1. Customer onboarding and KYC/KYB
Agents can help collect identity information, check completeness, and request missing documents, while analysts review cases requiring judgment. In Camunda’s Mauritius Commercial Bank case study, onboarding time fell from 70 to about 25 minutes, a reported 64% reduction. The implementation combined orchestration, intelligent document processing, AI/ML, and automation. The case study describes agentic AI as an area for future exploration, so the result should not be attributed to autonomous agents.
For a pilot, define which checks can run automatically, which require human review, and what evidence must be retained. Record the policy version, data sources, reviewer, and outcome where relevant. Orchestration can coordinate that flow; logging and approval paths still need to be implemented and tested.
2. Lending document processing and income verification
Classification and extraction components can sort documents, pull income information, and flag inconsistencies. Camunda’s Harmoney case study reports that broader loan-origination orchestration reduced average decision time from four days to about 17 minutes. It separately attributes a 20–30% reduction in drop-off among applicants whose standard income checks failed to BankBot, an AI agent used for complex income verification. These are results from one lender, not a forecast for every bank.
A useful design allows eligible, well-supported cases to proceed automatically and routes exceptions to an underwriter with the relevant documents and checks. The proportion that can safely take each path needs to be established from the bank’s own data.
3. Payment exceptions and investigations
Agents could help interpret payment messages, gather context, and propose resolutions, while deterministic checks validate required fields and a person reviews ambiguous cases. Swift’s research with 30 institutions estimates that its Case Management service could reduce investigation time by up to 80% and save the industry more than $600 million annually in operational and liquidity costs. Those are projected benefits of Case Management, not measured savings from AI agents.
For banks using the relevant Swift services, the migration milestones are specific: receiving camt.110 investigation requests with in-flow translation by November 2026, followed by ISO 20022 investigation requests and responses through Case Management in November 2027. This is not a blanket November 2026 deadline for every ISO 20022 payment flow.
4. Fraud detection and AML alert triage
An agent-assisted triage flow can gather transaction and customer evidence, prioritize alerts, and draft a rationale for an investigator. Treat that as a proposed operating model to test. A faster closure rate is not enough: reviewers need to check missed risks, unsupported explanations, and the quality of the evidence.
For an Israeli bank, agree with financial-crime and compliance teams which alerts can be handled automatically, which require investigation, and what must remain traceable. Measure both efficiency and detection quality.
5. Financial-crime investigation and SAR filing
Agents can help assemble evidence, reconstruct fund flows, and draft a suspicious-activity report for review. The investigator needs to validate source records, distinguish facts from inference, and approve the report through the bank’s authorized process. A fluent narrative is not evidence that the investigation is complete.
Reporting duties, deadlines, and accountability depend on the jurisdiction, institution, and report type. A reporting deadline from another jurisdiction should not be assumed to apply to an Israeli workflow. Have the compliance team define the applicable reporting rules, escalation points, and retention requirements before automating the work.
6. Credit decisioning and underwriting
Agents can gather and reconcile information and present a recommendation from approved risk models. The bank should define which decisions may be automated and which require an underwriter. The EU AI Act’s Annex III covers systems that assess the creditworthiness or credit score of natural persons, with an exception for financial-fraud detection. It does not classify every banking AI tool as high-risk.
As of September 30, 2026, the European Commission’s updated timeline sets December 2, 2027 for Annex III high-risk rules. For an Israeli deployment, confirm whether the EU regime applies and assess local requirements separately. Validate accuracy, fairness, explanations, and overrides against the actual credit policy; an orchestration platform does not establish those properties by itself.
What these six processes have in common
The proposed pattern is the same across these six areas: define what the agent may do, where fixed rules apply, when a person intervenes, and how the team can reconstruct the outcome. A named owner remains accountable for the process. Human approval should be meaningful, with enough information and authority to challenge the recommendation.
What this means for your organization
A bank that wants to be AI-native can start with one bounded process, human review at the right points, and measures of business value and control quality. At NG Workshop, we help Israeli banks and financial firms select that starting point and build it with Camunda, alongside their technology, risk, and compliance teams.
Further reading: AI Agents Need a Conductor and Human-in-the-Loop as a Governance Mechanism.
Sources
Summary
- KYC, lending documents, payment exceptions, AML triage, investigations, and underwriting are candidates to evaluate—not equally mature or equally suitable for agentic AI.
- Published results need context: conventional automation, agentic components, and projected savings are different kinds of evidence.
- Start with one bounded process, explicit authority, and measures that test both value and control.