---
title: Models may change. The control boundary does not.
description: AI models will keep changing. What trading desks need to hold fixed is the boundary around them, and who stays responsible for the trade.
image: https://avarr.ai/hubfs/models-may-change-the-control-boundary-does-not.png
---

[Skip to content](https://avarr.ai/insights/models-may-change#main-content)

[![Avarrai logo with signal pulse symbol](https://avarr.ai/hubfs/avarrai-logo-cream-full-1.svg)Homepage](https://avarr.ai)

- [Home](https://avarr.ai)
- [Platform](https://avarr.ai/platform)
- [Insights](https://avarr.ai/insights)
- [About](https://avarr.ai/about)
- [Contact](https://avarr.ai/contact)

- [Home](https://avarr.ai)
- [Platform](https://avarr.ai/platform)
- [Insights](https://avarr.ai/insights)
- [About](https://avarr.ai/about)
- [Contact](https://avarr.ai/contact)

![A trader rests his hand on a closed orange gate in a glowing intelligence layer, while teal streams flow only through the open channels.](https://avarr.ai/hs-fs/hubfs/models-may-change-the-control-boundary-does-not.png?width=1672&height=941&name=models-may-change-the-control-boundary-does-not.png)

Execution intelligence Governed AI

# Models may change. The control boundary does not.

![Abdullah Hiyatt](https://avarr.ai/hs-fs/hubfs/Abdullah800.jpg?width=48&height=48&name=Abdullah800.jpg)

 Abdullah Hiyatt

November 9, 2026

AI has arrived on the trading desk, and heads of trading are asking vendors sharper questions than they were a year ago. Which models sit behind the tool? What data do they use? And when something goes wrong, who is responsible for the trade?

At FILS Europe this autumn, one head of fixed income trading put the last point very simply. At the end of the day, the trader is responsible for the trade. [Coalition Greenwich](https://www.greenwich.com/blog/european-fixed-income-market-structure-focus) hears the same from desks across Europe. They welcome AI that removes manual analysis, but they are cautious where it starts to replace judgement, and they insist on being able to challenge what it produces.

Those are the right questions. The answers they get are often the wrong ones.

## The model is the part that will change

Ask most AI vendors how their product can be trusted and the answer starts with the model. It is the latest one, it has been fine-tuned, and it performs well on a benchmark.

The trouble is that the model is the least permanent part of the system. The models available today will be replaced within months, and then replaced again. Any vendor that is serious about quality will keep upgrading.

So if a desk's confidence in an AI tool rests on the model, that confidence has to be rebuilt every time the model changes. The same goes for any review, sign-off or oversight that was tied to it. That is not a sound foundation for anything that touches execution.

## What should stay fixed is the boundary

A better question is not which model a tool uses, but what that model is allowed to do.

In a well-governed system, the boundary around the model matters more than the model itself. It sets out what AI may do, such as interpret, rank, recommend and explain. It sets out what it may not do without a person deciding. It also determines what has to be recorded every time, so that the decision can be reviewed later.

When that boundary is designed properly, a model can be upgraded without the desk having to re-learn what the system is permitted to do. The controls, the approvals and the audit trail stay where they were. Only the quality of the advice improves.

That is why we put it the way we do. Models may change. The control boundary does not.

## Responsibility is a design requirement

"You're responsible for the trade" is often heard as a disclaimer. It should be read as a specification.

If the trader carries the responsibility, the system has to make that responsibility possible to exercise. The trader needs to see why a recommendation was made, in terms they recognise. They need to be able to disagree with it, easily and without friction. And there needs to be a record showing what was suggested, what was decided and by whom.

An AI tool that cannot do those things is asking the trader to accept accountability for reasoning they cannot see. Few experienced traders will do that, and they are right not to.

## An explanation has to be evidence

There is a subtler risk too. Language models are very good at producing explanations that sound convincing. But an explanation the model has composed for itself is not evidence. It is a story.

For an explanation to be worth anything on a trading desk, it has to be tied to the facts the recommendation was actually based on. A trader, a compliance officer or a portfolio manager should be able to follow it back to those facts and check it.

## Five questions worth asking any AI vendor

For desks evaluating AI tools for execution, these questions tend to separate governed systems from impressive demonstrations.

- What exactly is the AI allowed to do, and what is it not allowed to do?
- If you change the model tomorrow, what changes for my desk?
- Can my traders see why a recommendation was made, and override it easily?
- Is there a record of what was recommended, what was decided and by whom?
- Where does my data go, and who else benefits from it?

A vendor with clear, specific answers has thought about governance from the start. A vendor whose answers all lead back to the model probably has not.

## Governance is part of the infrastructure

In capital markets, intelligence cannot simply be capable. It has to be explainable, controlled and accountable. That cannot be added at the end as a compliance layer. It has to be part of how the system is built.

That is the principle behind [Avarrai's approach to governed AI](https://avarr.ai/platform#governed_ai). The institution sets the limits, the platform works within them, and the boundary holds whichever model sits inside it.

We're looking for a small number of buy-side desks to work with as design partners. If these are the questions your desk is asking, [we'd like to hear from you](https://avarr.ai/contact).

## Share this post

<https://www.facebook.com/sharer/sharer.php?u=https%3A%2F%2Favarr.ai%2Finsights%2Fmodels-may-change><https://twitter.com/intent/tweet?url=https%3A%2F%2Favarr.ai%2Finsights%2Fmodels-may-change><https://www.linkedin.com/shareArticle?mini=true&url=https%3A%2F%2Favarr.ai%2Finsights%2Fmodels-may-change><https://pinterest.com/pin/create/button/?url=https%3A%2F%2Favarr.ai%2Finsights%2Fmodels-may-change>[mailto:https%3A%2F%2Favarr.ai%2Finsights%2Fmodels-may-change](mailto:https%3A%2F%2Favarr.ai%2Finsights%2Fmodels-may-change)

## Keep reading

### [![A flood of teal market data rises from a city into a glass intelligence layer, where a trader stands at the start of one clear orange path.](https://avarr.ai/hs-fs/hubfs/the-tape-shows-you-everything.png?width=1672&height=941&name=the-tape-shows-you-everything.png) Execution intelligence Market structure The tape shows you everything. Then what?](https://avarr.ai/insights/the-tape-shows-you-everything)

### [![A trader stands on a glowing intelligence layer beside an orange stream of light held at its surface, above a teal city of data.](https://avarr.ai/hs-fs/hubfs/not-trading-is-a-decision-3.png?width=1672&height=941&name=not-trading-is-a-decision-3.png) Fixed income Execution intelligence Not trading is a decision. Nobody records it.](https://avarr.ai/insights/not-trading-is-a-decision)

Avarrai is building the execution intelligence infrastructure for capital markets.

Avarrai Limited is a company registered in England and Wales. Company number 16556164.

Registered office: 124 City Road, London EC1V 2NX, United Kingdom.   

 

© 2026 Avarrai Limited. All rights reserved.

 

[Home](https://avarr.ai) | [Platform](https://avarr.ai/platform) | [About](https://avarr.ai/about) | [Contact](https://avarr.ai/contact)

[Privacy Notice](https://avarr.ai/privacy) | [Cookie Policy](https://avarr.ai/cookies) | [Website Notice](https://avarr.ai/terms) | [Data Privacy Request](https://eu1.hs-data-privacy.com/request/AACOgqcenBvBTAycx3psqg)

 

<https://www.linkedin.com/company/avarrai/><https://bsky.app/profile/avarr.ai>

```json
{
  "@context" : "https://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Abdullah Hiyatt",
    "url" : "https://avarr.ai/insights/author/abdullah-hiyatt"
  },
  "dateModified" : "2026-11-20T23:30:24.832Z",
  "datePublished" : "2026-11-09T19:46:13.000Z",
  "headline" : "Models may change. The control boundary does not.",
  "image" : [ "https://avarr.ai/hubfs/models-may-change-the-control-boundary-does-not.png" ],
  "mainEntityOfPage" : {
    "@id" : "https://avarr.ai/insights/models-may-change",
    "@type" : "WebPage"
  },
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "url" : "https://avarr.ai/hubfs/avarrai-logo-cream-full-1.svg"
    },
    "name" : "Avarrai"
  }
}
```