# Why every AI agent sounds the same

> Agents built on the same models with the same defaults sound identical. In banking, insurance, healthcare and government that is a trust problem, not a style one.

Source: https://www.conversationdesigninstitute.com/blog/why-every-ai-agent-sounds-the-same

---

[ All blog articles ](/blog) 

# Why every AI agent sounds the same, and what it costs you

Agents built on the same models with the same defaults sound identical. In banking, insurance, healthcare and government that is a trust problem, not a style one.

[ ![](https://cdn.conversationdesigninstitute.com/800x800/416b1d23a1/peter-thomas.jpeg) ](/expert/peter-thomas)

[Peter Thomas](/expert/peter-thomas)

Blog 

Author

Nov 22, 2026

5 min. read

![](https://cdn.conversationdesigninstitute.com/2880x1620/386201a75c/agents-02-default-cover-2880.png/m/1800x0/)

## Chapters

[Can you tell them apart? ](#26029936-fd41-49f0-a861-2d37a86db328) [Slop is sameness ](#485b8c21-ca3b-4ed9-8a48-7598b387d954) [Why sameness is a trust problem ](#4d5579e4-8dc5-4875-863c-72023db9d62a) [A default agent makes no deposits ](#091e47f1-1345-48d4-acc1-8a0d9bf22c5a) [What the difference consists of ](#198a0963-2474-4aa2-bf9e-8f0a0d3c840b) [The model is a commodity; the conversation isn't ](#4f7a287a-16fc-4e73-ac56-8e86641a260d) 

## Can you tell them apart?

Most AI agents sound the same because they are built on the same models with the same defaults. In banking, insurance, healthcare and government that sameness is a trust problem rather than a style one: a conversation that could be anyone's is a conversation nobody is accountable for. To fix that doesn't need a better model - it needs better design.

Try an experiment. Open the assistant of your bank, then your insurer, then the airline you flew last, then whichever retailer most recently lost your parcel. Ask each the same awkward question. Then ask yourself whether, with the logos removed, you could tell which was which.

Probably not. They open the same way, apologise the same way, and offer the same three options when they get stuck. The tone is uniformly pleasant. The teams behind them are not lazy; they are all building on the same handful of models, with the same defaults.

## Slop is sameness

[Dan Shipper at Every has a useful definition](https://every.to/p/after-automation) of what has gone wrong. Slop, he says, is not a particular mistake. It is not em dashes or sentence rhythm. Slop is sameness, repeated until people notice. It is what you get when everyone has access to the same expert, who has the same default tendencies, and nobody thinks too hard about it.

His point was about writing and code. It applies to AI conversations, because conversation is where customers meet the organisation. A document that reads like every other document is a mild disappointment. An agent that behaves like every other agent is a different kind of problem: it is a conversation that could be anyone's, and that is the trust issue.

## Why sameness is a trust problem

Trust in a conversation is built on the sense that the other party knows who they are, knows who you are, and will act in a way consistent with both. When I call my bank, I trust the person partly because they are recognisably my bank's person: they know what my bank does and doesn't do, what it would never say, what it would agree to and what it wouldn't. The organisation's judgment is present in the conversation. The default agent has none of that. It has the model's judgment, which is to say the average of everyone's, and it will give my bank's customers the same advice it gives the insurer's.

Customers can tell. Not consciously, usually. But the sense that you are talking to something generic creates a sense that nobody is accountable, and that is corrosive in the moments that matter: a claim, a missed payment, a test result, a benefit that has stopped, and the question that doesn't fit the menu. In those moments the customer is looking for the organisation to show up, and a default agent is the organisation not showing up.

## A default agent makes no deposits

At CDI we think about every AI interaction as a deposit in, or a withdrawal from, a store of trust we call [Conversational Capital](https://www.conversationdesigninstitute.com/conversational-capital). The thing about a default agent is that it rarely makes large withdrawals. It is polite, it is accurate enough, it doesn't do anything alarming. What it does is make no deposits at all. On the P&L this is invisible. The agent costs what it was budgeted to cost and contains what it was promised to contain; what it isn't doing is the thing the investment was meant to buy, which is a customer more likely to stay. Nothing in the conversation tells the customer that this organisation is different from the last one, or that it would be worth coming back to. Over thousands of interactions the account drains, and the dashboard shows nothing because the dashboard is [measuring containment](https://hansvandam.substack.com/p/false-containment-how-your-best-dashboard).

The diagnostic we use is [a grid with empathy on one axis and agency on the other](https://www.conversationdesigninstitute.com/conversational-capital): how well does the agent read and respond to the person, and how much is it able to do for them. The default agent, almost without exception, sits high on agency and nowhere on empathy. It can do a great deal and it has no idea who it is doing it for. Vendors sell the agency axis because it is easy to demonstrate. The empathy axis is where the difference lives.

## What the difference consists of

What the difference consists of is less mysterious than it sounds. It is the organisation's judgment, encoded. What would we never say to a customer. Where do we give the benefit of the doubt. What does our customer already know that we shouldn't explain again. When do we stop being helpful and start being honest. Which situations get a person, every time, regardless of what the automation could handle. None of this comes from the model. All of it is design, and most of it is the kind of design that gets skipped when a deployment is treated as configuration.

In a [bank](/use-cases/financial-services), it might be the rule that any mention of hardship or a bereavement goes to a person, and that the agent never offers a product in the same conversation. In an insurer, that a claim is never declined in a chat window, only explained. In a [health service](/use-cases/healthcare), that the agent says what it cannot advise on before the patient has to ask. In a [government service](/use-cases/government), that it uses the same words for an entitlement as the letter the citizen is holding. None of these is a model capability. Each is a decision one organisation made about how it treats people, and each is something a default agent will never do on its own.

## The model is a commodity but the conversation isn't

Two banks on the same model can end up with agents that build trust at completely different rates, and the difference is entirely in what each bank decided its agent should know, refuse, notice and do. That decision is what the culture lens of CDI's [Standards Framework](https://www.conversationdesigninstitute.com/conversational-ai-for-business/standards-framework) looks for: whether an organisation has made those choices deliberately, or left them to the default. The organisations treating it as a procurement question will all end up with the same agent, and will have paid the same for it. The difference between the two banks is not in the contract.

The same cheapness that produces the default also makes the alternative affordable: it has never been easier to get an agent right for your organisation, one that is unmistakably yours. But nobody will do it for you, and if you don't, you get the default, like everyone else.

This is the second in a series on what autonomous agents ask of the organisations deploying them; the first asked [how much autonomy an agent should have](/blog/how-much-autonomy-should-an-ai-agent-have). The series, and CDI's wider work on designing agents for high-stakes settings, is collected on the [Agentic Experience Design](/topics/agentic-experience-design) page.