'You do have to give it a good assignment'

Oracle consultant at Defence

Wilko works as an Oracle consultant in the KBM team at the Dutch Ministry of Defence. That team manages the portal that keeps track of the clothing records of all Defence personnel. The same system also drives the issues and orders for military staff.

That sounds straightforward, but under the hood it is complex. There are integrations with clothing suppliers and with SAP EWM, where the orders eventually end up. Stock levels, item data and order statuses go back and forth constantly. "If there is no stock, you get a message back. If stock becomes available later, the order has to be offered again." All the logic lives in the Oracle database and, to a large extent, in APEX.

Two worlds: the client environment and the sandbox

Wilko's use of AI takes place in two worlds. "There is a difference between what I experiment with now and then and what I can actually use. The central environment is pretty much locked down."

At the client he uses DefGPT, Defence's own AI assistant, every day. He can paste project code into it, give it instructions and have it modify code. For truly new pieces of code he sometimes works on his own PC with ChatGPT or Gemini and then brings the result into the client environment. The rule is clear: "Of course you cannot take any Defence information or code outside." So only new code goes in that direction, never the other way around.

He does notice the difference in quality. With exactly the same instruction, a current public model delivers a better result. "We are simply a bit behind the times." Even when he provides the context, such as Oracle 19c and APEX 23.1, the internal assistant still sometimes comes up with more old-fashioned solutions.

In his own sandbox, a free account on apex.oracle.com, he works with the latest versions: Oracle Database 23ai and the latest APEX release. That is where he tries out what is not yet possible at the client:

• Generative AI in APEX: having a dashboard generated from natural language.
• AI in interactive reports: asking a question such as "give me all employees who joined after 2000", after which AI turns it into a query itself.
• AI Vector Search: searching by meaning instead of exact words. Anyone searching for "boots" will then also find the laarzen (Dutch for boots) in the range.

"For that, too, you need a newer database version. I hope Defence will take a look at 23ai soon."

Small pieces, always test

His use grows along with the models. "You notice that ChatGPT or Gemini keep getting smarter, and then you start using them more and more. Two or three years ago it was less, but every year you use it more often."

It is not blind trust. Wilko mainly has short functions and separate pieces of code generated. He does not check them line by line, but he always tests them for the purpose he wants to use them for. "Once you have had ten different functions generated, you eventually end up with a component that you designed yourself." So the design remains his; AI supplies the building blocks.

The biggest gain: reviewing and testing

When Wilko looks at where AI can deliver the most, he does not point to writing code, but to what comes after. "You can have code reviewed against the standards you have described. That way you can check yourself. There is a great deal to gain in reviews and testing." With a wink: "That is sometimes the work I enjoy least, too." His ideal is a combination: reviewing it himself and having an AI tool go over it once more.

A concrete example is at hand. For a new multi-layer clothing line, all interfaces to SAP were adapted, with six different types. "Really, you want to create a test script for that which you can hook into the pipeline." Because of time pressure, the team still chose manual testing: that feels safer. "While with AI it might actually have been faster. But that is hard to estimate."

His own solution to that dilemma: "Give yourself the time. Do a 16-hour spike, see whether it works and what comes out of it. If it does not work, you still do it another way."

"You can have code reviewed against the standards you have described. There is a great deal to gain in reviews and testing."

What holds it back: tools, time and perception

Has AI fundamentally changed his work yet? "Not really at the moment. Even though I would like it to more." He is thinking of tasks such as: describe a table and let AI set up the rest, including constraints and triggers. "Unfortunately we do not yet have the right tools for that in the client environment." His wish is simple: "Ideally, Copilot gets installed everywhere and you can get going."

Besides technology, perception plays a role. According to Wilko, most people view AI positively, but some are still unfamiliar with what AI actually does. That leads to reluctance to use it.

Still, he is optimistic. Defence has its own innovation team that experiments with AI, and that is where DefGPT came from as well. The new cloud platform, where the APEX application servers are moving to, also offers opportunities: "A lot has to be automated there, and there too you can use AI to adapt scripts."

Worries about his own profession are something Wilko does not have. "I do not have the feeling that I will suddenly be cut in a few years." What does concern him is security. "That the models become so smart that wherever there is a vulnerability, it gets hacked in no time."

Looking ahead: from prompt to reusable component

The most important skill? "Being able to prompt well, that is for sure. You keep noticing that it works when you provide the right context every time." He recommends creating a standard starting prompt, for example for every new piece of PL/SQL.

For the coming years he is mainly looking at agentic AI. "That you can quickly build good components with prompts, which you can reuse to handle recurring tasks in a few instructions." He also sees applications outside his work. As organizer of a local street fair with around 120 stallholders, he wants to have the personal emails with pitch allocations generated. "You can automate all of that. That is something I want to try one day too."

Why this matters for Qualogy

Wilko's story shows that the use of AI depends not only on the consultant, but also on the client environment. At the same time, we see clients in the market increasingly asking for consultants with concrete experience in AI-assisted and agentic development. Especially for colleagues who are (still) limited at their client, it is therefore important to gain experience outside it. Think of our AI workshops, the AI literacy training and the follow-up sessions that are coming. Wilko: "I think we as Qualogy are on the right track."

Wilko's tips for colleagues

• Always provide context: versions (for example Oracle 19c, APEX 23.1), standards and the goal.
• Work in small pieces: have short functions generated and build the component yourself.
• Always test: for the purpose you are going to use it for.
• Use AI for review and testing: as a second pair of eyes alongside your own check.
• Stick to the client's rules: no client code or data in external tools.
• Experiment in your own sandbox: such as a free account on apex.oracle.com.
• Give yourself a spike: for example, plan 16 hours to try out an AI approach.

AI in one sentence

"AI is my great help in development as an Oracle consultant." With one caveat: "You do have to give it a good assignment. That is the key."

How do you use AI in your work, or why not? Let me know. I would be happy to talk with you for a next episode of 'AI in practice'.