The AI ROI Problem Isn't What You Think
Recent research from Forrester points to a completely different reality: AI's ROI challenge is an operational and measurement problem, not a technology bottleneck.
The AI ROI problem isn't what you think. Recent research from Forrester points to a completely different reality: AI's ROI challenge is an operational and measurement problem, not a technology problem.
Debates about AI ROI will persist regardless of model performance until we fix how we track and define success.
Think of it like this: The surgeon is skilled. The OR is equipped. But the patient wasn't properly prepped, the instrument tray was incomplete, and the procedure notes haven't been updated since the pre-op consult. Nobody blames the surgical technique. They look at the protocol that was never followed.
We do the opposite with AI.
The Upstream Failure Loop
When an agent underperforms, we immediately rush to fix the technical parameters:
- Upgrade the model.
- Code a better prompt.
- Open a bigger context window.
- Fetch a newer version.
But the actual failure happened upstream. There was no clear definition of success, data pipelines that nobody validated, and evaluation baselines that silently drifted while everyone watched the initial demo.
That's not a model quality problem. That's a definition problem: nobody agreed on what good looks like. And in most organizations, it's the one role nobody's chartered.
Who owns accountability for your AI agents—IT, the vendor, or the business? That's the question worth answering before the next deployment.