GenAI Brings A New Kind Of Agile Computing That Needs Controls Even More

October 19, 2026 Michel Mouchon

As if trying to make use of a whole new computing architecture and coming up with the budget to experiment and roll features into production were not enough, there are a slew of issues when bringing GenAI to the IBM i datacenter.

There are those who say that people cannot learn from the experience of others. We have never believed that. Which is why we want to share some of the hard-won wisdom we have developed four years after the ChatGPT moment that rocked the world and that is changing many things. But not everything needs to change or will change. This is an important fact that needs to be remembered to put business adoption of GenAI into proper perspective.

Let’s go over a few points for you to consider as you ponder the future of GenAI at your company.

First, you need to think of GenAI as a junior programmer. Some customers are hypnotized by the magic of GenAI and they want to do everything with it. No matter what, I cannot convince them to go another, more conservative way to control how GenAI helps develop code and who it does – and does not – get put into production. This is their choice, and we have no problem with that.

We are practical and careful and thoughtful about how we deploy GenAI ourselves here at ARCAD Software. We all know that the IBM i platform has a skills depletion problem as people retire from the market. So having a new junior programmer – whether it is Claude Code or OpenAI Codex or IBM Bob – that is aware of a lot of different languages and that can understand your code base is useful. This GenAI junior programmer can make suggestions on how code could be changed, but you have to double check its work.

I will give you an example of how we trust GenAI and how we don’t. In our ARCAD Transformer Microservices tool, which automates code refactoring, we definitely use GenAI to search for redundant code and for similar code that executes functions. This is obviously a statistical, probabilistic approach, and to be honest, GenAI is very good at this and it is better and faster than deterministic approaches that are less flexible. (More on this dichotomy in a moment.)

GenAI is a new guy, one that learns very quickly and reads even faster. But that is just speed. That is not necessarily understanding.

One of our colleagues is a senior IBM i developer and is very strong in development and the ARCAD tools, and she has started using GenAI for Java programs, She was originally an IBM i developer but some of the company’s applications deliver features in Java. She knows both C and Java, and when the GenAI tool she uses proposes some code to implement a function, she is able to judge and to say “Yes, it’s good” or “No, that is not good.” This is how you must approach it.

Second, the code base on which GenAI models have been trained to understand RPG is very thin. This is a known fact, and that is why IBM was soliciting RPG code from partners and customers two years ago when the IBM i division was building its own code assistant before IBM standardized on Bob for all of its customers. We have done some trials with this GenAI code assistants to see how well they can modernize RPG code. And one out of two times, it is wrong. Moreover, often the behavior of the program has been changed. Maybe this is a question of prompting. Before AI, all bugs were created by developers. Now, we have a GenAI machine creating code – and also creating bugs – as well as people creating code and also creating bugs.

Three, managing risk with GenAI is even more important than it was with the move to agile computing. If your developers are working quicker and quicker, they probably will generate more bugs because if in every 1,000 lines of code they produce one bug and now they work with GenAI to produce 10,000 lines of code in the time they produced 1,000 before, they will probably produce 10X more bugs at least, if not more. I don’t see how the code could be less buggy when it is done fast. And that makes it even more important to have a very strict protocol for testing code and promoting it to production.

So unit testing, regression testing, deployment with validation steps, and so on, must be done to ensure that what you deploy to production is not straight out from this newbie GenAI junior programmer. You would treat a human junior programmer no differently, by the way. And even senior programmers have to be part of the DevSecOps process to ensure application viability, reliability, and security.

This new bottleneck can explain we are seeing an uplift in our quality and test automation tools CodeChecker, iUnit, and Verifier, which ensure that bugs are caught as early as possible the process.

And four, use AI where a probabilistic answer is good enough and use deterministic tools where it isn’t. That split is how ARCAD itself uses GenAI tools. It is not a battle of deterministic versus probabilistic approaches, but rather using these complementary approaches together in a way that makes sense. We just have to be conscious of making this distinction when creating new code or modernizing existing code. If you validate these points, you can use AI securely.

This is also how we position ourselves as an enabler of the usage of GenAI. AI must be framed into different processes that will ensure that everything is done properly and works as expected.

For the past three decades, we have developed and delivered tools that use deterministic processes to analyze production RPG code, manage the DevOps process, modernize the source code and database, regression test the changes, just to name a few. Now with our 200+ MCP services, these same tools can expose deep context to the GenAI and agentic AI tools so as to prevent risky mistakes and reduce token cost. That’s a realistic approach that enables IBM i customers to adopt agentic and GenAI safely and realize real advantages for their business.

Michel Mouchon is chief technology officer and also chief information security officer at ARCAD Software.

This content is sponsored by ARCAD Software.

This Issue Sponsored By

Table of Contents

Content archive

Recent Posts

Subscribe

Pages

Search