---
name: ai-use-case-matrix
description: Identify and prioritize AI use cases with a task inventory, a three-criteria scoring matrix (value, risk, feasibility) and a one-page pilot charter with a scale-or-kill decision date. Use when the user asks where to apply AI, which AI project to do first, how to prioritize AI ideas, or how to run an AI pilot.
---

# AI use-case matrix

Turn a wish list of AI ideas into two measured pilots. An AI use case is a
business task with an owner, a monthly volume and a measurable current cost,
never a technology: "classify the month's 900 support emails, service team,
45 hours" qualifies; "use a chatbot" does not.

## Step 1: inventory tasks, not technologies

Interview each function owner with three fixed questions:
- Which repetitive task consumes the most hours of your team each month?
- What information do you look up again and again in different places?
- What work piles up when someone is out?

Add an anonymous shadow-AI question: which AI tools do people already use on
their own, for what? Clandestine use is a demand-validated candidate; it gets
regularized, not punished.

Write every candidate in one line: verb + object + monthly volume + owner.
Then attach three numbers: current cost (hours x hourly cost), volume, and
variability (identical every time, or every case a world of its own?). This
silently kills half the list, which is its job.

## Step 2: knockout risk filter

Two questions before any scoring. Does the task touch personal or regulated
data without an approved circuit? Would an error reach a customer or a
regulator without human review in between? A yes without mitigation parks the
case until governance exists. Safety first, enthusiasm second.

## Step 3: score with three criteria, 1 to 5 each

Total = value + feasibility - risk. Three criteria and not one more: every
extra column is a new place to hide a personal preference.

| Criterion | 1 point means | 5 points means |
|---|---|---|
| Annual value | Under 50 hours/year, or value impossible to express in hours or euros | Over 1,000 hours/year, or direct revenue or regulatory impact with a defensible figure |
| Risk (subtract) | Error is internal, visible, cheap to fix; no personal data | Error reaches customers or regulators, or touches regulated data |
| Feasibility | Scattered or paper data, task different every time | Accessible digital data, clear pattern, pilotable with general tools in two weeks |

Expect the surprise: the committee's glamorous case (usually a customer-facing
chatbot) loses to the boring internal case with volume.

## Step 4: pilot charter, one page or no pilot

Fields: task and owner (a business owner, not the AI lead); baseline measured
BEFORE touching anything; one success metric with its threshold; estimated
cost in three lines (tokens/subscriptions, platform, human supervision);
governance rules (what runs alone, what needs review, where failures are
logged); and a decision date 4 to 8 weeks out where scale or kill are the
only two valid outcomes. Both outcomes are success: one produces a system,
the other produces cheap learning.

## The five decision rules

1. Nothing enters the matrix without its three numbers.
2. Two pilots maximum at once; the third steals the supervision the first two need.
3. No pilot without a prior measured baseline.
4. Every decision date is honored; "keep piloting" is the forbidden answer.
5. Rescore the whole matrix every quarter: feasibility moves with each model
   generation, and whoever rescores first captures first.

## Source and method

Full blueprint with worked examples and FAQ:
[y8y.ai](https://y8y.ai/es/casos/como-identificar-casos-de-uso-de-ia/).
Author: Sebastián Ocampo ([LinkedIn](https://www.linkedin.com/in/seb-ocampo/)).