In banking, insurance, and healthcare, the chatbot is the new IVR.
Customers will do what they always did: endure it, then demand the human.
One concession first, because it decides the scope of everything after it. The AI assistant you chose works. ChatGPT, Muse, whichever one you picked: you invited it in, you set what it may touch, and you can fire it by closing the tab. Chosen AI earns trust. This article is about the other kind, the agent an institution places between you and itself.
We have seen this wall before
IVR was sold as containment. Leadership heard deflection rates and cost per contact. Customers heard something else: a wall between them and a person who could actually decide something.
So customers adapted, the way customers always do. They mashed zero. They repeated "representative" until the machine gave up. A generation learned to treat the menu as an obstacle course with a human at the end of it.
The chatbot inherits the wall and improves the grammar. It parses full sentences now. It apologizes in a warmer tone. Underneath, the shape is the same: a system whose job is to handle you, standing where a person used to stand.
Chosen and imposed
The difference between the assistant people love and the chatbot people endure is who did the choosing. Your own assistant runs on consent, control, and an exit: you granted the access, you can narrow it, and you can walk away. The institution's version offers the customer none of the three. Its refusal is the institution's refusal, delivered politely, at scale.
That is why imposed AI reads as avoidance. The customer cannot tell where the agent's design ends and the company's reluctance begins. Every "I can't help with that" lands as the sound of a door being held shut by someone who could open it.
Nobody mashes zero on their own assistant.
The proof from banking
This month, the argument got its proof from the other direction. Axios reported that banks are moving to block third-party AI agents from transacting on customer accounts when those agents have not been vetted by the bank.
That is simply good risk discipline. An unvetted agent with transaction power over other people's money is a fraud surface with an API. Blocking it is the prudent move, and I would want my own bank to make exactly that call.
But hold the two decisions next to each other. The customer's own agent, chosen by the customer, authorized by the customer, acting for the customer, gets zero trust and no access. The institution's chatbot, chosen by nobody the customer knows, gets zero authority and all of the conversation. Banks trust agents so little they block the ones customers chose, and ship the one customers did not, with no power to decide anything.
The authority ceiling
The obvious fix is to give the chatbot real authority. In trust industries, that fix is the one thing nobody can ship.
Look at what customers actually come to decide. Waive the fee. Reverse the charge. Settle the claim. Approve the exception. Clear the patient. Each is a decision with a dollar sign or a diagnosis on it, and handing them to a customer-facing agent means the first generous or mistaken call multiplies. One wrong fee waiver is a correction; ten thousand, overnight, is a financial event. One wrong clearance is a clinical event. One screenshot is a reputational event. No risk committee approves that, and on the merits, no risk committee should.
The courts have started pricing the statements too. In 2024, a Canadian tribunal ruled against an airline whose chatbot told a passenger he could claim a bereavement refund after booking, when the real policy required applying before travel. The airline argued its chatbot was a separate legal entity, answerable for its own words. The tribunal rejected that and made the airline pay the refund difference. This year, a German appeals court reached the same conclusion at a higher level, ruling against a healthcare company whose website chatbot gave its doctors specialist titles they did not hold, and brushing aside the idea that customers should know bots make mistakes. Neither agent had the authority to decide anything. Each only had to say something wrong, once, in public, and the institution owned every word.
So the agent ships under a ceiling. It may answer questions. It may collect details and summarize them. It may apologize, fluently, in forty languages. Then it escalates. Everything the customer came for lives on the far side of that ceiling, with a person.
Here is the test I apply to any agent pitch in a regulated industry. Ask what the agent is allowed to decide. If the answer is nothing with a dollar sign on it, you bought a menu. An articulate menu that never sleeps is still a menu, and customers will treat it the way they treated the last one.
What survives
What survives is the AI that does not announce itself. The strongest consumer AI removes steps instead of adding conversations: it uses what the institution already knows and pre-fills what never needed retyping. The chat window is the loudest tier of AI and the weakest. Above it sits the ping, where the system acts and surfaces only the decision that needs a person. Above that sits silence, where the product simply feels faster and more considerate.
The metric that sorts the tiers is time-to-purpose: how long it takes the customer to accomplish the thing they came for. A chatbot never wins that metric, because it adds a conversation to a journey that wanted an outcome.
This is also where the economics actually work. Removing steps lowers cost per completed outcome, the number leadership wanted from the chatbot in the first place.
And when the journey reaches a genuine decision, the goal flips. Surface the human early, carrying the full context, so the customer never retells their story to the one person who can decide it. In a trust industry, the agent's finest moment is a warm handoff. Deflection is the failure mode wearing a success metric.
None of this argues against replacing the IVR itself. Swap the phone menu for an agent that understands plain sentences, carries the context, and gets a caller to the right human faster, and you have improved a feature customers already use. The wall gets a door in it. That is honest, measurable progress.
The mistake is the other rollout: shoving a chatbot into every channel and calling the coverage a strategy. The app gets one. The website gets one. The phone tree gets a second one. Each ships under the same authority ceiling, so the institution has multiplied the wall and taught its customers to mash zero in four places instead of one.
IVR never really died. It taught two generations of customers that the menu is not the destination, and they have carried that lesson into every chat window since. So here is the question I would put to every institution shipping a customer-facing agent this year. Is yours a door, or is it a wall with better grammar?