The Benefit of AI and the Risk of AI Are the Same Thing

Ask a dozen Access developers what a human-led, AI-staffed, development team buys you, and at least 9 or 10 will say "speed". In my opinion, that's the wrong answer; at best, it's a shallow view of a much deeper situation. Producing viable code quickly was never the bottleneck for experienced developers. We could always type faster than we could figure out what to type. Logical analysis, planning, and design were the real constraints, not raw keystrokes.

In my opinion, speed to completion is simply the wrong metric for what AI Agents do. The part of the process where AI Agents have the biggest impact is the cost of exploration. AI allows us to push well beyond the frontier, as my friend, Eric, put it in one of our rambling discussions. I like that phrase, because it fits my personality and history as an Access developer.

As some of you know, I have a penchant for calling myself a "Cowboy Coder"; that's someone who's always heading off into the wide open spaces, exploring new territory. Living by the motto, "How hard can it be?"

There's also a significant societal tradition there, too. It's the basis on which the American West was explored and settled. In one way, I guess, AI enables that tendancy, although it's also fenced me in in other ways. More on that later.

Working with AI, you ask for something, say, for example, a function that lists all emails in your inbox from your clients where the word "urgent" or the word "idiot" appears. Instead of an approach you would have reached for out of habit, your AI proposes an approach you didn't know existed. Or maybe it offers three versions to compare side by side. The 2 important points here are that:

• you didn't spend half a day researching what could be possible
• you also don't have to spend the rest of the day trying to implement one of those versions

The cost of trying a out a new idea drops very close to zero. A recursive CTE you would have had to dig out of one of your previous databases to copy becomes a five-minute experiment instead of an afternoon spent searching and studying and experimenting.

An AI agent lowers the investment of time, energy and mental effort of going past the edge of what you already know how to do.

In a word, as I see it, the most impressive benefit of AI assistance is the ability to explore beyond your existing knowledge.

The Cost of Freedom

All that comes with its own cost, though. The moment you're past the edge of what you already know, you no longer have the ability to validate the AI's work. You've entered a zone where trust can no longer be taken for granted.

The primary job of the lead developer (the human) is judgment. A team led by a human and staffed by machines works only as long as the human can still evaluate what the machines produce. AI agents are notoriously overly confident, and they are confidently wrong at a rate that can be very hard to measure.

You, the human, are the only team member who can look at an AI's suggested output and say no, that's broken. Unless you can spot the Blank() that lands in a calculation as 1900-01-01 instead of NULL, you're not leading a team. You're trusting a stranger; some might even go so far as to suggest it is a stranger which has a lot in common with con men.

Thus, we have to understand how the same benefit that makes AI Agents so valuable can erode the most important skill we need to exercise as lead developer, or project manager.

The more you let the agents implement the work autonomously, the less likely it is you will learn enough to catch their mistakes.

If you lean on AI to explore beyond your knowledge, over time your own knowledge can stop growing like it used to. You drift toward being a reviewer of code in a language you're slowly forgetting, or worse, never learned at all.

Ironically, therefore, the greatest benefit to using AI is operating past your current competence, and the greatest risk is operating past your competence.

Here's where I realized that being a Cowboy Coder has its place. It's the spirit that is never satisfied with the status quo.

But there's also a place for the Citizen Developer, the one who stops roaming and settles down to really build a solid, dependable home with well-defined borders, standards and, like it or not, rules. The Lead Developer is the Citizen Developer who has accepted responsibility for getting getting it all right.

The key to success is staying close enough to AI generated code to catch the inevitable lie before you incorporate it into your client's project. Read and understand what the agents write; don't just run it and watch it pass.

When you say to yourself, "I don't understand this yet", that's a hard stop. Exploration is a benefit only if you close the loop and actually understand and validate the thing before it goes into production.

I'm sure I'm not the only one who gets annoyed by an agent's abundant confidence. That's partly because we recognize that being positive is not the same as being right. Just because AI keeps telling you how brilliant your insight is doesn't mean the AI is telling you the truth. Verify everything it tells you that you don't already know. And not by asking the AI; by testing the actual behavior in the actual working environment with real data and real conditions.

Obviously, none of my thoughts here are new. The Lead Developer's job didn't change when agents showed up. What did change was that more of us were moved from ordinary developers into Lead Developer roles. At the same time, the role itself got harder, and perhaps a bit more ambiguous.

You didn't become the lead in your team because you know more than anyone else in your shop. You became the lead because you actually know less than the AI agents you manage. And now you are learning on the job what the task of managing an AI agent really involves.