Thinking about AI

The greatest focus last week (2025-W21) at the Microsoft Build event in Seattle is certainly AI usage and agentification — or using agents — of it.

Creating specialized agents that will perform specific operations and that can interact with each other, work in parallel, act on behalf of the user that runs them, etc. is something that is becoming a very interesting exercise.

As usual the first, and most common reaction we have and see on other people, is enthusiasm. The will on using these new technologies and improving products, time to market and people’s quality of life.

The AI is certainly something that is here to stay. And we need to learn how to use it, how to implement guardrails, and how to monitor its actions.

Implementing AI

The first step to using AI is actually not doing any implementation of it, but learning how to use it to deliver a business need / goal.

The second step is defining the processes, teams, and tasks needed to deliver that business need. Here it is an opportunity to review what is done so that we can write it down and be more efficient. Eliminate some steps and create a more optimized process. Deliver services to new infrastructure and you'll also be thanked for…

The third step is bringing in the IT teams and defining an implementation process. From here, it is executing it.

It is important to think about who will interact with the AI as there are costs associated with that. These will either be tied to each person or to some other metric such as the number of tokens or the number of words processed.^[1]

People have to be trained, and the interface (options in a menu, open text interaction) designed to allow the correct usage.

It is a nice project, with enough details and options to justify an assessment, some tests (PoCs — proofs of concept) and a project. In a big company, some specialized consulting will surely help.

Attention points

• Scope: what is the problem AI should address
• Integrating with information sources: from where the AI will learn or consume information to interact with your users
• Access restrictions: the answers have to be tailored to each user, if they don't have access to the source themselves then they can't receive an answer using that information
• Interface: how your users will interact with the AI
• Costs: it will have to be part of your budget

And remember that AIs today can hallucinate and based on that you'll have to see very clear rules to work with information obtained from it.

Good luck on your journey!