Why we're looking at this
“AI will save you time” is not a business case on its own. We care about where, specifically, agent-native software changes a client's cost structure or revenue, because that's what determines whether a project is worth building at all — see our related notes in small business automation ROI.
What we're seeing
- The clearest ROI shows up in work that is high-volume and rules-heavy but currently requires a trained human to apply judgement — clinical intake, testing validation, first-pass documentation. This matches what Upfreq Robotics reports for robotics testing: the win is compressing a slow, expert-gated loop, not replacing expertise entirely.
- ROI is easiest to measure, and easiest to defend internally, when the agent's output is reviewed by a human before it has real-world consequences — which also happens to be the safer rollout path.
- The most overestimated case is open-ended “creative” automation with no clear success criteria — without a way to measure whether an agent's output is actually good, the ROI conversation stalls indefinitely.
Open questions we're still chasing
We're building a simple framework to score a proposed agent project on volume, rules-clarity, and reviewability before we scope it — partly to protect clients from projects that sound exciting but won't pay back, and partly to sharpen our own intake process.