Picking the One Metric That Matters for AI ROI

The first question to ask isn't which model — it's which metric you're actually trying to move. As David Arnoux frames it, "the first question to ask yourself is really, like, what's the one metric that matters at the moment that we're trying to optimize for?"

From there, ROI measurement hinges on defining evaluation up front. Before building anything, Abraham Gomez (Google) recommends asking: how would you evaluate the system at scale? If the current evaluation depends on one domain expert eyeballing outputs, that won't scale — so the problem becomes building the AI system plus replicating your evaluation person.

Concretely, that means documenting the domain understanding of whoever judges quality today, then turning that into a repeatable eval. Without it, you can't prove the system is working, let alone calculate ROI.

It depends on the strategic orientation of the organization. Ops teams look for operational efficiency. Customer-facing leaders — CROs and CMOs — typically want to increase pipeline, increase conversions, increase ARPU, increase retention.

The trap is wanting all of them at once. As Arnoux puts it bluntly: "everybody wants to optimize everything at the same time." That's precisely what the audit phase is for — deciding where the biggest bang is.

The sequencing rule: kick off with a high-impact, relatively high-ease use case. Once that proof of concept or quick win is established, move into the very high-impact, slightly more complicated use cases. Which funnel layer you start with — awareness, demand gen, or deeper conversion — follows from the motion you've chosen to focus on. (See also: beating-analysis-paralysis-enterprise-ai-projects.)

The audit's job is to answer "what should we focus on? Where's the biggest bang?" — because left to their own devices, teams try to optimize everything simultaneously.

A practical prioritization frame, from ScorePlay's founder: focus on the most time-consuming, smallest value-add work first. In their case, that meant adjusting Canva templates for proposals — high time spent, low value from the sales team. Handling objections, by contrast, stayed human because that's where you most need to understand how the prospect thinks.

Anticipate the next bottleneck before you ship the first automation. Once a workflow is automated you become so much more productive that "the bottleneck becomes the next step after that" — automate proposal creation and suddenly you're handling 20x more objections.

A simple impact-vs-ease matrix keeps the audit honest. (For non-technical operators running this themselves, see learning-ai-tools-as-non-technical-operator.)

In one 400-person European unicorn Arnoux references, the GTM team runs a three-to-one ratio of developers to marketers — AI-native engineers working mostly on observability of what's being released and maintenance to keep workflows running, alongside one strategist per three builders.

That said, the marketers can build themselves up into automation experts: "it's not that complicated anymore." Arnoux notes he has a weak dev background but runs roughly 25 workflows simultaneously across different ventures, with technical help only when something breaks.

The implication for ROI: you don't need a research team training models. You need one strong GTM strategist, builders for observability and maintenance, and a clear definition of which Gemini, Claude, or off-the-shelf entry point you're actually using.