In my essay about measuring the value of Generative AI (GenAI) for knowledge work, I argued that organizations should focus their AI ROI measurements on processes and workflows where value is created. I’d now like to examine this a bit further by answering the next obvious question:
“How do we measure AI ROI in the places where it matters?”
My answer is that it depends less on the use of AI than on the nature and maturity of the underlying business process, and how that process generates value.
Let’s unpack this a bit by considering three common use cases where many organizations look to integrate AI into core business processes: software development, IT service management, and proposal writing.
Software development
For software development, ROI can’t be measured in terms of lines of code or even the amount of code generated with AI. The meaningful measure is how efficiently a team creates production-ready software that passes testing, code review, security review, and acceptance. Additional effort needed to review, correct, and validate AI-generated code must be included in the equation. And opportunities for AI to make other elements of the software development lifecycle more efficient should be as well.
Also, anyone who’s serious about using agentic AI coding tools like Claude Code or Codex understands that token consumption can quickly become very expensive. That means that for ROI calculations to be meaningful and actionable, AI costs need to be captured at the same level at which productivity gains are being measured.
IT service management
In IT service management, ROI is not about tickets touched by an AI system. It’s about the rapid resolution of IT disruptions with minimal escalation. Reducing support costs in lower tiers (without driving traffic to higher tiers) can be an obvious form of value to capture and measure.
Still, an even more important benefit often comes from reducing IT friction across the enterprise. Every minute AI gives back to employees is a minute they can spend doing something more valuable than waiting on a help desk.
In other words, if you are measuring only the reduction in cost within your IT organization, you’re probably failing to measure—and therefore understating—the true ROI.
Proposal writing
When it comes to AI-assisted proposal writing, AI ROI isn’t measured through pages drafted or the percentage of content generated by AI. It’s about producing compliant, compelling proposals that require minimal review and rework. It’s about increasing organizational throughput. Ultimately, it’s about improving win rates and contributing to backlog growth.
Measure what you understand
These examples all have some very important aspects in common.
First, they’re mature business processes with defined workflows, established gates, and measurable outcomes. That means the impact of AI is not evaluated in isolation; it is measured in terms of how it improves value generation across the overall workflow.
Second, because they’re typically important processes, well-run organizations already understand their baseline performance (including how they currently generate value) without AI.
And this leads to a broader point. I’ve been told by many people that AI ROI is difficult to measure. In many cases, I do not think this is really an AI problem at all.
AI does not create measurement problems. It exposes them.
If you cannot describe a workflow, identify where value is created, or establish the current baseline, how can you credibly claim to measure the ROI when AI is introduced?
The practical corollary is straightforward: If demonstrating AI ROI is important to your organization, process maturity (and associated measurability) should become one of your use-case selection criteria.
Do not start with the most exciting AI application, at least if you want to demonstrate meaningful ROI. Start with the business processes you already understand well enough to recognize—and measure—real improvement in AI-enabled value generation.