Most companies prioritize AI initiatives with a value-vs-feasibility matrix, a use-case scoring model, or a process-mining tool. These are popular because they produce a tidy quadrant chart everyone can point to in a meeting. They share the same flaw: they rank hypothetical initiatives. They tell you where to look. They don't tell you which dollar figure is real today.

Why a scoring model points at the wrong thing

A scoring model rates an idea by how good the pitch sounds, not by whether the number behind it is real. "Automate customer service triage" can score high on a feasibility matrix and still be worth close to nothing, because nobody checked whether reps are quietly bypassing the tool that was supposed to do the automating. The matrix has no way to catch that. It was never reading the system, only the proposal. Gartner's 2026 survey of 782 infrastructure and operations leaders found only 28% of AI use cases meet ROI expectations, and traced the gap to exactly this: use cases picked for how exciting they sound, not how feasible or provable they are.

Start with what your systems already show

Instead of scoring hypothetical initiatives, read what your own operational systems already show about where work actually breaks down:

  1. Read the operational systems already in place, CRM, support, contracts, finance, for where handoffs break, automation is bypassed, or a paid capability goes unused.
  2. Quantify every finding in dollars against a source log, not an estimated productivity multiplier applied to a headcount number.
  3. Rank findings by what's actually provable, not by how well they fit a quadrant.

None of these three steps ask what an AI agent could theoretically do. They ask what your own systems are already telling you, if anyone bothered to read the logs instead of running a workshop.

What this catches that a matrix misses

A scoring matrix only ranks what people think to propose. Reading the systems directly surfaces what nobody would have thought to propose in the first place, because the waste is invisible unless you look at the gap between what a tool produced and what a human actually did with it. A CRM can show a lead-routing automation as "active" while a third of leads are still being reassigned by hand every morning. No use-case matrix generates that finding, because nobody would think to write "manual lead reassignment" onto a roadmap slide. The system already knows about it. Nobody read the log.

One real scan of a fintech's CRM stack, walked through in detail in a later post, found exactly this pattern: four findings, none of which would have shown up on a feasibility matrix, worth $77,235 a year.

Where to go from here

The result of reading the systems instead of scoring a matrix is a ranked list where every dollar figure traces back to a log line instead of a workshop estimate. That's a different kind of prioritization than a quadrant chart produces, and it's the only kind that survives being questioned by a board.

Frequently asked questions

What's wrong with a value-vs-feasibility matrix for prioritizing AI initiatives?

It ranks how promising an idea sounds, not whether the underlying dollar figure is real. Two initiatives can score identically on a matrix while one is backed by an actual log-verified number and the other is a guess.

What should I read instead of building a scoring model?

The operational systems already in place, CRM, support, contracts, finance, for where handoffs break, automation gets bypassed, or a paid capability goes unused. That's where the real, provable dollar figures already live.

How do I quantify a finding once I've found one?

Against a source log, not an estimated productivity multiplier applied to a headcount number. A number traceable to a real system event is defensible in front of a board; an estimated multiplier is not.

← Back to Blog