GenAI, AI agents and agentic AI are three different species of system — with three different risk profiles, cost structures and futures. Today we take them apart, live, using one banking chatbot.
The industry uses these terms interchangeably. They are not interchangeable. The variable that separates them is autonomy — how much of the loop between intent and outcome the system owns.
A model that generates content — text, summaries, explanations — from what it learned in training plus your prompt. It has no hands: no access to your data, no ability to act.
An LLM given tools — APIs, databases, search — that it can choose to call in order to complete a bounded task you asked for. It fetches, checks, and does — one delegated task at a time.
A system that takes a goal, decomposes it into a plan, executes multi-step workflows, adapts when steps fail, and operates across time — sometimes without a human in each loop.
Same customer. Same message. Turn the dial and watch the same banking chatbot become three fundamentally different systems. Watch the right-hand panel — that's what the system is allowed to do.
Six real-world banking scenarios. Classify each one. This is the exact judgement you'll need in vendor meetings when everything is marketed as "agentic."
Here is the idea most banks haven't priced in: within a decade, a large share of "customers" hitting your APIs will be other agents. Your client's personal finance agent will shop your rates, negotiate your fees, and switch providers at machine speed — while its owner sleeps. Watch a personal agent run a mortgage refinancing auction:
One human intent — "cut my mortgage cost" — becomes a multi-party agent negotiation.
This flips every assumption in consumer banking: loyalty becomes an algorithm, marketing must persuade machines (APIs and verifiable terms beat brand jingles), pricing becomes continuous and personal, and the moat shifts from distribution to being the most trustworthy counterparty for someone else's agent. Banks will need agent-facing storefronts — machine-readable terms, cryptographic mandates, sub-second onboarding — the way they needed websites in 1999 and apps in 2010.
A working map, not a prophecy. Each era is gated by three things maturing together: model capability, trust infrastructure, and regulation. The bank that wins each era is the one that built for it in the era before.
GenAI chat matures into tool-using agents grounded in real account data. RM copilots draft, summarise and pre-fill. Hallucination drops from headline risk to managed metric. The quiet work: permissioning, audit trails, evaluation harnesses.
Customers grant scoped mandates: "keep my cash at the best yield," "dispute fees under $100 automatically," "rebalance quarterly within my risk band." Human-on-the-loop replaces human-in-the-loop for low-stakes actions. First regulatory frameworks for AI mandates appear.
Personal agents negotiate with bank agents directly. Agent-readable product terms become a distribution channel. "Customer acquisition" splits in two: winning humans and winning their agents. Interoperability standards (agent identity, verifiable mandates) become as strategic as card networks once were.
A household runs a standing financial agent: continuous tax optimisation, insurance re-brokered monthly, bills negotiated, savings goals managed as a portfolio of sub-agents. Banks compete to be the execution layer and the trust anchor underneath other people's agents.
Most retail financial decisions are made or brokered by agents. Markets in deposits and credit clear continuously. The open question is no longer capability but governance: who is accountable when your agent and my agent agree to something neither of us would have? This is where the AGI debate stops being philosophy and becomes supervision policy.
The honest answer: agentic ≠ general. A banking agent that flawlessly runs your finances is still a narrow system with superb orchestration. AGI implies competence across essentially any cognitive domain, with transfer — a different claim entirely.
But here's the more interesting framing: banking may be one of the first domains to feel AGI-like, because money is legible. It's digital, structured, rule-governed and measurable — exactly the substrate agents excel at. A domain can approach functional generality within its own borders long before general intelligence arrives.
So the strategic question isn't "will our chatbot become AGI?" It's: at what level of autonomy does capability stop being the bottleneck — and trust becomes the product? Every era on the roadmap is really a trust milestone wearing a technology costume.
The bank that understands this sells something no model can commoditise: accountable autonomy.