Who gets the keys to your cloud environment: your IT team or an AI agent?

Tools that provide insight into your cloud environment, after which you take action? You are used to that by now. But this way of working is changing rapidly: AI agents are increasingly analyzing problems in order to propose optimizations and implement changes. Microsoft now speaks of agentic cloud operations: cloud management that moves from insight to action.

Attractive, because this saves a lot of manual work and your response times go down. At the same time, this raises a new question around AI governance: 'How much authority do you give an AI agent?'

100% AI agent... or human in the loop?

Suppose your AI agent sees that a certificate needs to be renewed and performs that action independently. That feels quite comfortable, because this action is predictable and involves few risks.

But when an AI agent adjusts firewall rules or scales your production environment without your intervention, the cards are different. These are grander actions with many implications. That is why you may prefer not to leave it completely to an officer.

With every action, you want to ask one crucial question: 'Do I entrust this task to a cop or do I keep the human in the loop?'

You won't find the answer in what is technically possible.

Record what an officer is allowed to do and log what he does

First of all, you can give an AI agent autonomy just fine. As long as you know at any time who did what and when.

This starts with frameworks that fit within your organization's policy, drawn up with all stakeholders who have something to say about this. Then you make sure your AI agents know what they are allowed to do. You determine when agents are allowed to take action themselves and when an administrator has to approve something. You also determine how the escalation path to people runs: which colleague needs to know something? And at what time?

It is important to log everything. Both the changes made and the decisions behind them. Why did an officer escalate or not? Under what mandate does the agent act? Without this trace, you can never explain exactly what happened and who was responsible for an action. So if something goes wrong, you can't trace the problem.

But where? In your policy, the regulations you are bound by and the agreements you have made with customers.

How do you control the cost of an AI agent?

With traditional cloud costs, you can measure what an application, user or environment costs reasonably well. FinOps for AI is a different kettle of fish. All kinds of costs are added. Think of tokens, model choice, API calls, processing time, data, orchestration and monitoring.

That's why it's wise to keep the cost aspect in mind when developing your AI solution. Have the agent log what he does and put data on it, so you can see how many tokens are being used. Settle that against the current price and you have insight. You then process this information in your FinOps dashboards.

Then the real decisions will be on the table. Do you pass on the costs to your customer? Do you give everyone a fixed number of tokens per month? You want to make these kinds of choices consciously, before the bill comes in.

Can you demonstrate the cost savings?

How much has an AI agent earned you and where exactly is the saving? You want to be able to answer that question at the end of the month. This only works if you have a clear idea in advance of what pain you are solving. Therefore, take a good look at your business case before you determine what you will have an agent do.

Do you want to know how you can best set up the frameworks for your organization? Or how to keep a grip on your AI costs? Feel free to contact us to discuss your options.