Everyone's talking about AI transforming enterprise software delivery. Some of it holds up. A lot of it doesn't survive contact with a real project.
My honest take: it depends entirely on the type of system and the type of work. AI earns its keep where there's repetitive work, solid documentation, and a fast feedback loop. It falls apart where the real knowledge lives in someone's head, not in a doc anywhere.
Where it's actually good
- Test generation. Writing unit tests, catching edge cases, generating test data, filling coverage gaps. Works best when a developer already understands the system and just needs help covering every scenario by hand.
- Legacy code and migrations. Understanding old code, converting patterns, explaining dependencies, helping with framework upgrades. This used to be weeks of a senior engineer just building context. AI shortens that a lot.
- Documentation that doesn't exist. Most enterprise systems have documentation that's outdated or scattered across tools nobody opens. AI is good at reading the actual codebase and explaining what's really going on.
- Debugging. Feed it logs, get plausible causes back, narrow things down faster.
- ERP configuration, with a catch. It can help with config guidance and documentation. But it still needs someone who knows the business rules, because ERP systems run on company-specific logic that only exists in people's experience. This is where the gap between something that sounds right and something that is right shows up most.
Where it falls apart
Every failure I've seen traces back to the same thing: AI treated as a replacement instead of an accelerator.
- Output that's technically fine but wrong for the business, because the model doesn't know the real context
- Piles of generated code or tests that look like progress but become maintenance debt
- More time spent reviewing and fixing AI output than it would've taken to just write it
- Complex systems where the important knowledge isn't written down anywhere for the model to find
- False confidence. A polished-looking wrong answer doesn't get double-checked the way a shaky one would
What people keep getting wrong
AI doesn't remove the need for judgment. The value isn't "it writes code." It's that it clears out low-value grunt work so experienced people can focus on architecture, decisions, and solving the right problem in the first place.
This shows up in the broader data too.
A December 2025 RGP survey of 200 U.S. CFOs found that only 14% reported seeing a clear, measurable impact from their AI investments to date.
The issue appears less to be that AI models do not work, and more that many organizations are still figuring out where AI creates durable business value, how to redesign workflows around it, and how to measure returns properly.
What the good ones do differently
They're not bolting AI onto everything. They're fixing their processes, data, and documentation first, so AI has something solid to work from.
Less exciting than "AI cuts your ERP timeline in half." But it's the version still true six months after the pilot ends.
What's your experience been? Curious where others have seen this land or fail in their own projects.
Top comments (0)