Skip to content
CM System Designs, business systems for South East SMEs
Open navigation Menu

Who controls your AI Harness?

By Chris Palmer · 17 November 2026

Artificial intelligence is moving from answering questions to getting things done.

A few years ago, asking an AI system to summarise a document felt remarkable. Today, an AI agent can examine an enquiry, look up a customer’s details, check stock, prepare a quotation and move the work on to the next stage. Increasingly, it can do much of that without someone prompting it at every step.

The potential for businesses is enormous. Agentic AI could take on repetitive work, help teams make better use of their information and give smaller companies capabilities that would once have required much larger teams.

But as you begin to put AI agents into the heart of your business, there is a question worth asking early:

Who controls the system that controls your AI?

The answer could have a lasting effect on your costs, resilience and freedom to choose how your business operates.

What is an AI harness?

An AI model is the technology that can interpret information, reason about a task and generate a response. An AI agent uses a model, together with tools and instructions, to work towards a goal.

The harness is the surrounding system that puts that agent to work. It connects the agent to your business information and software, gives it tasks, controls what it is allowed to do and determines what happens when it needs help or something goes wrong.

Think of an AI agent handling an incoming customer order. The model might interpret the customer’s request and identify what needs to happen. The harness can then govern how the order is processed: which customer records can be accessed, how stock is checked, whether a discount needs approval and what gets recorded along the way.

That surrounding system matters. If AI is going to play a part in running your business, the harness determines much of how it does so.

The convenient route

The major AI providers are working hard to make their technologies easy for businesses to adopt. Many offer platforms that bring models, agents, connections and workflow management together in one place.

That can be a sensible way to experiment. It may get an initial project up and running quickly, and a managed platform can provide valuable security, support and operational features.

The question changes when an experiment becomes an essential part of the business.

Imagine that, over time, you use one provider’s platform to handle customer enquiries, prepare quotations, update your CRM, manage purchasing and coordinate internal approvals. Your team comes to rely on it. Your processes are built around it. Your operating knowledge and history accumulate inside it.

At that point, changing provider means more than choosing a different AI model. It may mean rebuilding the machinery that carries out your daily work.

You could find yourself dependent on that provider’s prices, product decisions, terms and availability. Even if an alternative offers a better model or better value, the cost and disruption of moving may make a change difficult.

The risk is easy to overlook because it develops gradually. Each individual automation may seem like a small, sensible decision. Taken together, they can become the operating fabric of the business.

A lesson from the Microsoft outage

The Microsoft 365 outage in November 2026 offered a reminder of how disruptive a dependency on one provider can be. Problems affecting services including Outlook and Teams disrupted communications for businesses and other organisations around the world. Microsoft said services had fully recovered after more than 24 hours of disruption.

That incident was not primarily an example of an AI agent making a poor decision. It illustrates a broader point: when important functions depend on a common provider, a problem outside your control can quickly become a problem inside your business.

Now consider what happens when the affected service is involved not only in communication, but also in progressing orders, updating records, preparing payments or coordinating customer work.

Owning your AI harness would not prevent every outage. Your business might still depend on external models, cloud hosting, internet connections and other software. But designing the system with those dependencies in mind gives you more opportunity to prepare: to pause safely, switch where practical, or provide a manual route for essential work.

Who checks what the agents are doing?

Availability and cost are only part of the decision. There is also the question of oversight.

An AI agent may be able to follow instructions such as “refer unusual discounts to a manager” or “do not send sensitive information outside the business”. But an instruction given to a model is not the same thing as a control enforced by the surrounding system.

What happens if the agent misunderstands the instruction? What if it encounters a situation the original designer did not anticipate? What if information in a customer document or email attempts to persuade it to ignore the rules?

A business should be able to decide which actions an agent can take, which require approval and which it cannot take at all. For example:

  • An agent might prepare a quotation, but a person must approve a discount above an agreed limit.

  • It might reconcile an invoice against an order, but a person must authorise payment.

  • It might draft a response to a complaint, but a member of the team must review it before it is sent.

  • It might identify a record that appears incorrect, but it cannot delete that record itself.

Those boundaries should be enforced by the system wherever possible, rather than relying solely on the agent to remember them.

There should also be a useful audit trail. When an agent takes an action, you need to be able to establish what it did, what information it used, which rules and approvals applied and who authorised any consequential step. That matters when something goes wrong, but also when you want to understand whether the process is working as intended.

AI can assist with a decision. Accountability for how the business operates still rests with the people running it.

What happens when the system changes?

Agentic systems may also retain information between tasks. Some use stored instructions, documents or memory files; others use databases or services that retrieve relevant information when needed.

This can make them more useful. An agent that knows your products, processes and customers may be able to do much more than one starting from scratch each time.

But it raises another question: who can see and change the information that shapes its behaviour?

If an agent can add to its own memory or operating instructions, a seemingly small change may influence later decisions. Over time, it could become difficult to understand why the system behaves as it does, particularly if nobody can see what changed or when.

A well-designed harness can put controls around that information: restricting who or what can make changes, keeping a history, requiring approval for changes to important rules and allowing those changes to be reviewed or reversed.

The aim is to make the way the system operates visible and manageable as it evolves.

An alternative: a harness your business controls

One approach is to build a harness around your own workflows, with the AI models providing capabilities inside it.

In this arrangement, your business can retain control of the important elements: the workflow, permissions, approval points, records and connections to your systems. You can decide which model is used for a particular task and review that choice as technology changes.

For example, you might use:

  • A powerful frontier model for a difficult planning or analysis task.

  • A smaller, less expensive model for routine classification or extraction.

  • A suitable model running on your own infrastructure for certain sensitive tasks, where keeping data within that environment is a priority.

The right mix depends on the task, the data and the quality required. A smaller or locally hosted model will not automatically be cheaper, more secure or capable enough for every job. It still needs testing and proper management. But having the choice allows you to make those decisions for your business.

A harness designed for more than one provider can also make it easier to change models over time. That does not mean every model can be swapped instantly: they differ in their abilities, interfaces and behaviour, so a replacement must be integrated and tested. It does mean that your core processes need not be built entirely around one provider’s way of doing things.

You do not have to build everything at once

A custom harness does not have to begin as a large, complicated project.

Start with one business process. Choose something valuable enough to improve, but contained enough to understand and test. Map how the work happens today, identify where an agent could help and decide where a person must remain involved.

Then ask some practical questions:

  1. What information does the agent need, and what should it be unable to access?

  2. Which actions can it take, and which need human approval?

  3. What record will we have of its actions and the decisions around them?

  4. What happens if the AI provider or another connected service is unavailable?

  5. Who can change its instructions, permissions or stored memory?

  6. How would we move to another model or provider if we needed to?

These questions are easier to answer before an agent becomes responsible for a critical part of the business. They can also help you judge whether an existing platform gives you enough control, or whether a bespoke approach would serve you better.

Make the decision while you still have a choice

Agentic AI has the potential to transform how businesses operate. The companies that make good use of it could reduce repetitive work, respond faster and give their people more time to focus on customers and decisions that need human judgement.

But the decisions made at the beginning will shape what is possible later. The model you choose matters. The system that governs how it is used may matter even more.

At CM System Designs, we help businesses build bespoke systems around the way they work. We can help you identify where AI agents could add value and build a custom harness with the controls, audit trails and human approval points your business needs, while keeping your options open as AI technology develops.

If you are considering how to bring agentic AI into your business, talk to us. You do not need a finished brief. We can start with the process you want to improve and work out what a controlled, dependable approach could look like.

What could work better
in your business?

You don’t need a finished brief to start a conversation.