Agentic coding tools are genuinely transformative. Point one at a greenfield codebase and it accelerates everything. Point the same tool at an undocumented legacy system and it does something far more dangerous: it guesses.
The cost of a confident wrong answer
On a side project, a hallucinated change is an annoyance. On a core banking or insurance system, a confident wrong answer is catastrophic — and the more fluent the model, the more convincing the mistake. The barrier to using AI on the systems that most need modernizing is not the model's capability. It is the absence of a trustworthy model of the system for the AI to reason against.
Grounding, not prompting
The fix is not a better prompt. It is a better substrate. When agents reason against a deterministic, source-cited representation of the system, they stop guessing and start working from facts:
- They can explain what a component does, with a citation.
- They can assess impact before proposing a change.
- They can trace data flow across the system end to end.
The discipline that makes this safe is simple to state and hard to fake: the AI explains the facts; it never invents them. Where something is unknown, it says so, rather than filling the gap with plausible fiction.
What this unlocks
With a grounded model in place, the productivity of modern AI finally reaches the systems that carry the most risk and the most value — without betting the business on a guess. That is the difference between agentic modernization that lands and agentic modernization you cannot trust.