Agentic AI is set to have a bigger impact on finance than the internet-here's what to do about it
A new Citi report argues that AI-especially Agentic AI-will shape finance more than the internet era did. This isn't about smarter chatbots. It's "do it for me" technology that sets goals, plans tasks, acts, and adapts in real time.
2025 is the tipping point because the pieces finally fit: stronger models, enterprise controls, cloud scale, and guardrails that meet financial standards. For finance teams, this means real workflows, measurable ROI, and fewer manual loops.
Agentic AI vs. generative AI-why it matters in finance
Generative AI analyzes and drafts. Agentic AI goes further: it decides and executes. Think of an agent that identifies a payment error, digs through unstructured data, cross-checks records, amends the entry, submits for approval, and logs the full audit trail-without hand-holding.
That's not theory. This year, major players started deploying it: OpenAI released Operating and Deep Research features, Microsoft Copilot introduced agent-like behaviors, and banks such as Scotiabank let agents perform autonomous tasks in commercial banking.
From ecommerce to productivity, personalization, and autonomous workflows
The internet added ecommerce to finance. Agentic AI adds three levers: productivity gains, deeper personalization, and end-to-end workflows that run themselves within policy and control.
In practice, this looks like automated case resolution, reconciliation, exception management, compliance triage, credit ops support, portfolio rebalancing assistants, and SME cash-flow agents-each with permissions, approvals, and audit logging.
What's next in 2026: ROI, not demos
The center of gravity shifts from "this is cool" to "this pays." Expect broader deployment and clearer commercial impact across cost-to-serve, cycle times, error rates, client response times, and revenue per banker.
We'll also see agent-led commerce: agents transacting on accounts, moving funds within set limits, renewing mortgages, and handling renewals and documentation with a clear approval path. That's a material shift for retail and commercial operations.
Jobs: fewer repetitive roles, more oversight and higher-value work
Agentic AI changes the work mix. Repetitive operational roles shrink; new roles grow: AI agent supervisors, AI ethics officers, and governance leads. The net effect is higher productivity and a push toward higher-value tasks that still require human judgment.
The teams that win will pair agents with human review, clear escalation paths, and post-action verification-especially in regulated workflows.
Risk: bot traffic, security, and hallucinations
Half of global internet traffic is already bots, and many are malicious. That pressure will rise as illicit actors adopt new tools quickly and share methods. Independent research has tracked similar trends for years (Imperva Bad Bot Report).
On model reliability, hallucinations still happen. Governments are moving to set stronger rules, such as the EU AI Act. Enterprises should move first with their own controls-don't wait for regulation to catch up.
Practical playbook for finance leaders
- Pick 3 high-friction workflows: payment exceptions, KYC refresh, loan docs, reconciliations, or client onboarding. Define success metrics (AHT, error rate, time-to-close).
- Stand up an agent policy: scope, allowed actions, approval thresholds, logging, and immutable audit trails.
- Permissions and money movement: use a "pre-paid AI wallet" with spend caps and daily limits. Separate test and prod wallets.
- Authentication: add an automated auth layer for agents (rotating credentials, short-lived tokens, device fingerprints).
- Controls: human-in-the-loop for any irreversible action; dual approval for transfers above set thresholds.
- Data access: minimize privileges, mask sensitive fields, and route unstructured data through classification before agents touch it.
- Validation: require source citations and verification steps for any agent-generated claim, then auto-log the proof.
- Security: bot defense at the edge, anomaly detection for agent behavior, and red-teaming for prompt/agent exploits.
- Procurement: require vendors to show isolation, data retention policy, SOC 2/ISO proofs, and incident response timelines.
- Change management: train staff for new roles (agent supervisors, control owners) and create a fast feedback loop with ops.
Where the gains show up first
- Operations: exception handling, reconciliations, break resolution, dispute workflows.
- Risk and compliance: document parsing, evidencing, policy mapping, model attestations.
- Front office: prep for client meetings, tailored outreach based on portfolio events, loan renewal packaging.
- Treasury and payments: routing optimization, FX netting suggestions, cash forecasting agents with alerting.
Checklist to start this quarter
- Pick one line-of-business workflow and one enterprise workflow. Timebox to 8-12 weeks.
- Define clear guardrails: what the agent can do alone vs. what needs approval.
- Instrument everything: timestamps, actions taken, approvals, and outcomes for audit and ROI tracking.
- Run an A/B: agent-assisted vs. legacy. Keep the variant that beats on cost and quality.
Useful resources
- AI tools for finance teams: curated options by use case Explore tools
The takeaway: Agentic AI is moving from headlines to balance sheets. Start with tight scopes, strong controls, and measurable outcomes-then scale what actually works.
Your membership also unlocks:
