Agentic AI Examples: 18 Enterprise Use Cases by Industry (2026)

How to read these examples

Every example uses the same four questions, because they are the ones a risk committee asks:

• What the agent does: the steps it plans and carries out.
• Systems it touches: the tools it calls, which is where the risk sits.
• Where the human approves: the point beyond which the agent cannot go alone.
• Governance angle: the rule or supervisory expectation that shapes the design.

We name a specific organisation only where it has described its own deployment in a primary publication. Everywhere else the example describes a common pattern rather than a particular company. If you want the underlying building blocks first (chaining, routing, evaluator loops and approval steps), read our explainer on agentic workflows and their patterns.

Agentic AI examples in banking

Banking has the most mature set of agentic AI use cases because its operations already run on written procedures, case management systems and documented approval authorities. That structure gives an agent a path to follow and gives examiners a place to look.

Supervisors are paying attention. FINRA’s 2026 Annual Regulatory Oversight Report added a section on AI agents, which it describes as systems that can autonomously perform and complete tasks on behalf of a user. It warns about agents acting without human validation, going beyond the user’s intended authority, being hard to audit across multi-step reasoning, and mishandling sensitive data. It asks firms to consider monitoring agent access, placing human-in-the-loop review, tracking agent actions and adding guardrails (verified October 2026). Those points are written for broker-dealers, but they read as a sensible checklist for any bank.

At least one large bank has described its own agent programme in public. In a joint announcement dated 8 December 2025, BNY and Google Cloud said Gemini Enterprise had been integrated into Eliza, BNY’s enterprise AI platform, to support agents that process and analyse large financial reports and run deep research for market analysis.

1. KYC and onboarding file review

• What the agent does: reads the onboarding pack (identity documents, company registry extracts, ownership charts, source-of-funds evidence), checks it against the bank’s KYC procedure, lists missing items and drafts the risk rating rationale.
• Systems: onboarding workflow, document store, screening service, company registry lookups.
• Human approval: the KYC analyst confirms the risk rating; enhanced due diligence cases go to a senior reviewer.
• Governance angle: the agent’s checklist must match the bank’s written procedure version, and every conclusion needs a document reference. See the KYC onboarding agent.

2. AML alert investigation and SAR drafting

• What the agent does: for each monitoring alert, pulls account activity, prior alerts, the customer’s expected behaviour and adverse media, then drafts a disposition note or a SAR narrative.
• Systems: transaction monitoring, core banking, case management, news screening.
• Human approval: the investigator decides to close, escalate or file. The agent never files.
• Governance angle: 31 CFR 1020.320(e) bars a bank and its directors, officers, employees and agents from disclosing a SAR, so the drafts, prompts and logs belong inside the bank’s confidentiality boundary. The AML and sanctions agent and the AML, KYC and surveillance use case show the split between preparation and decision.

3. Card disputes and chargebacks

• What the agent does: classifies the dispute reason, gathers the transaction record and merchant data, checks card scheme timelines and drafts the customer letter and the chargeback submission.
• Systems: card processing, dispute management, customer communication.
• Human approval: an operations analyst approves provisional credit decisions and outbound letters.
• Governance angle: deadlines and consumer-protection rules are unforgiving, so the agent’s job is to make sure no case misses a date. The disputes and chargebacks agent covers this pattern.

4. Credit memo preparation

• What the agent does: spreads financial statements into the bank’s template, computes ratios with a calculation tool rather than in the model, summarises collateral and guarantors and drafts the memo.
• Systems: loan origination, document store, spreadsheet tooling.
• Human approval: the relationship manager and credit officer own the recommendation and the decision.
• Governance angle: the European Commission lists credit scoring that can deny a person a loan among its high-risk examples under the AI Act, and says Annex III high-risk obligations apply from 2 December 2027 (verified October 2026). Keeping the agent in a preparation role for commercial credit, with the credit decision staying human, keeps the design simple. See the credit underwriting agent and the loan origination use case.

The full set of banking agents, from servicing to treasury, is in the banking agents catalogue, and the banking AI use case hub lists further workflows.

Insurance

The NAIC adopted its model bulletin on insurers’ use of AI systems on 4 December 2023. It expects an insurer to keep a written AI systems programme covering governance, risk controls and internal audit, and it states that decisions supported by AI must still comply with unfair trade practice and unfair claims settlement laws (verified October 2026). Agent designs in insurance follow from that.

5. First notice of loss and claims triage

• What the agent does: reads the claim notification, photos and policy, checks coverage basics, flags missing information and routes the claim by complexity and fraud indicators.
• Systems: claims platform, policy administration, document intake.
• Human approval: the adjuster decides coverage and payment.
• Governance angle: a triage score must never become a silent denial; the claims settlement rules still apply to the human decision. See the claims processing agent and the FNOL triage use case.

6. Commercial underwriting submissions

• What the agent does: extracts exposure data from broker submissions, loss runs and schedules, compares them with appetite guidelines and drafts the underwriting summary.
• Systems: submission inbox, rating workbench, appetite guide.
• Human approval: the underwriter prices and binds.
• Governance angle: the bulletin’s third-party expectations apply when the model or data comes from a vendor. See the insurance underwriting agent.

Healthcare

7. Prior authorization packets

• What the agent does: for a planned procedure, gathers the clinical notes, test results and the payer’s documentation requirements, assembles the request and tracks the response.
• Systems: electronic health record, payer portals or APIs, document management.
• Human approval: clinical staff confirm the packet; any appeal is clinician-led.
• Governance angle: the CMS Interoperability and Prior Authorization final rule (CMS-0057-F) requires impacted payers to decide urgent requests within 72 hours and standard requests within seven calendar days from 2026, to give a specific reason for denials, and to support a Prior Authorization API by 1 January 2027; drugs are excluded (verified October 2026). Agents on both the provider and payer side work to those clocks. See the prior authorization use case.

8. Clinical documentation support

• What the agent does: drafts visit notes, referral letters and discharge summaries from the encounter record for the clinician to edit.
• Systems: EHR, dictation, template library.
• Human approval: the clinician signs every note.
• Governance angle: protected health information should stay in the provider’s controlled environment. See the clinical documentation agent.

Manufacturing

9. Predictive maintenance work orders

• What the agent does: when sensor analytics flag an anomaly, reads the asset history, the maintenance manual and spare-parts stock, then drafts a work order with likely causes and parts.
• Systems: condition monitoring, maintenance management, inventory.
• Human approval: the maintenance planner schedules the job; nothing changes machine settings automatically.
• Governance angle: operational technology networks are often isolated, which favours models running on site. See the predictive maintenance agent.

10. Quality deviation and root cause analysis

• What the agent does: groups defect reports, links them to batches, suppliers and process changes, and drafts a root-cause hypothesis with supporting data.
• Systems: quality management, manufacturing execution, supplier records.
• Human approval: the quality engineer accepts the root cause and corrective action.
• Governance angle: corrective actions in regulated manufacturing need traceable evidence, so every claim in the draft should link to a record. See the root cause analysis use case and the quality inspection agent.

Supply chain and procurement

11. Shipment exception handling

• What the agent does: watches carrier updates, detects delays that threaten a delivery date, checks alternative routes and inventory at other sites and proposes options with costs.
• Systems: transport management, carrier feeds, warehouse and ERP.
• Human approval: the logistics planner chooses and books the option.
• Governance angle: booking freight commits spend, so the agent proposes and a person commits. See the exception and disruption use case.

12. Purchase requests and supplier checks

• What the agent does: turns a free-text purchase request into a structured requisition, checks preferred suppliers and contract terms, and runs basic supplier risk checks.
• Systems: procurement suite, contract repository, supplier risk data.
• Human approval: budget holders and buyers approve under the normal delegation of authority.
• Governance angle: the agent must not change supplier bank details or approve its own requisitions. See the procurement agent.

Public sector

In the United States, OMB memorandum M-25-21 of 3 April 2025 requires agencies to apply minimum risk management practices to “high-impact” AI, including pre-deployment testing, an AI impact assessment, ongoing monitoring, human oversight with a fail-safe where practicable, and timely human review and appeal for people affected by AI-enabled decisions (verified October 2026). Agent designs in government work backwards from those duties.

13. Public records requests

• What the agent does: logs each request, searches the records systems, de-duplicates results and proposes redactions with the exemption cited.
• Systems: request tracking, document repositories, email archives.
• Human approval: the records officer approves every redaction and release.
• Governance angle: releases are legally binding and public, so approval is mandatory. See the public records request use case.

14. Permit application review

• What the agent does: checks applications for completeness against the published checklist, compares site details with zoning data and drafts a deficiency letter.
• Systems: permitting system, GIS, document intake.
• Human approval: the case officer decides and signs.
• Governance angle: applicants need a route to human review of any adverse outcome. See the permit processing use case.

Legal

15. Contract review against the playbook

• What the agent does: compares each clause with the organisation’s positions and fallbacks, proposes redlines and drafts a summary of open issues.
• Systems: document management, clause library, matter management.
• Human approval: a lawyer accepts or rejects each change.
• Governance angle: client and matter confidentiality decides where documents and logs may live. See the contract review agent.

16. eDiscovery first-pass review

• What the agent does: classifies documents for relevance and privilege, groups near-duplicates and drafts a privilege log entry for each withheld item.
• Systems: review platform, custodial data stores.
• Human approval: reviewing attorneys make the final relevance and privilege calls.
• Governance angle: the review protocol and sampling plan should be documented so the process can be defended. See the eDiscovery review use case.

IT operations

17. Incident investigation

• What the agent does: on a new alert, gathers recent changes, logs, metrics and the matching runbook, forms a hypothesis and proposes a remediation.
• Systems: observability stack, ticketing, code repositories, runbook wiki.
• Human approval: the on-call engineer approves any change to production; read-only diagnostics can run unattended.
• Governance angle: OWASP’s “excessive agency” risk for LLM applications applies directly, so diagnostic tools should be read-only and remediation tools separately granted. See the incident response agent.

18. Service desk triage

• What the agent does: classifies tickets, answers known questions from the knowledge base, gathers diagnostics and routes the rest to the right resolver group.
• Systems: IT service management, identity directory, knowledge base.
• Human approval: access grants and account changes go to an approver.
• Governance angle: identity changes are high-impact even when they look routine. See the service desk agent and the ticket triage use case.

What the strongest AI agent use cases have in common

Across all 18, the use cases that survive a risk review share five traits:

The weak candidates are the reverse: open-ended goals, no procedure, nobody accountable and broad write access. Those are the projects that stall in review, or worse, pass review and then surprise someone.

Running these agents on your own infrastructure

Many of the examples above involve data that cannot leave the organisation: SAR material, health records, privileged legal files, operational technology data. VDF AI is designed for that situation. It runs on the customer’s own servers, in a private cloud or fully air-gapped, and VDF AI Cloud is available where hosted deployment is acceptable.

Agents are built in Agent Hub, a no-code builder whose tools come from an administrator-governed MCP tool registry with grants assigned per role. Retrieval over connected systems such as SharePoint, Confluence, Jira and Google Drive respects each user’s permissions. Human approval steps pause an agent before consequential actions and record the approver, and every model call, retrieval and tool call is logged for audit, which also supports EU AI Act evidence. The AI Router decides which model, local or cloud, may handle each step under the organisation’s policy.

To compare platforms for building these agents, see the AI agent platform overview, and for governance roles and evidence, the agent governance guide. For the procedures behind each example, browse the agents directory.