# The AI Pilot Scoping Starter Kit

From "Applied AI for Marketing Ops" (appliedaiformops.com) by Lily Luo. Companion to my post "How to Actually Get AI Working at Work." This is the framework I use to start with a real problem, choose the right method, and define the smallest pilot that can prove value.

**Pick your level:**
- **Level 1: Manual.** One worksheet, 30 to 45 minutes, no tools.
- **Level 2: The full setup.** A folder for documenting the problem, approach, pilot, measurement, and final decision.
- **Level 3: Hand it to your agent.** Paste this whole file to your AI agent and say "help me scope this pilot." The last section is written as instructions for the agent.

The principle at every level: choose the work before choosing the tool.

---

## Level 1: Scope one pilot by hand

### Step 1. Find the challenge

Start with friction you can feel. Choose one recurring process and answer:

```
PROCESS OR WORK:
WHO DOES IT:
WHAT TRIGGERS IT:
HOW OFTEN IT HAPPENS:
HOW LONG IT TAKES:
WHERE THE INPUTS LIVE:
WHERE THE OUTPUT GOES:
WHAT VARIES EACH TIME:
WHAT BREAKS OR CREATES DELAY:
WHO WOULD NOTICE IF THIS IMPROVED:
```

Look for work that takes too long, varies too much, produces inconsistent output, or cannot scale. Describe the problem without naming an AI product.

### Step 2. Choose the approach

Match the work to the simplest method that fits:

- **Automation:** repetitive steps with predictable rules. The same input should produce the same output.
- **Workflow with AI:** a structured process where one or more steps require synthesis, interpretation, or contextual generation.
- **Agent:** continuous work that requires judgment, tools, memory, and action inside a defined scope.
- **Keep it in the chat:** a one-time question, analysis, or artifact that does not need to become a shared system.

Write down your choice and the reason:

```
APPROACH:
THE STEPS THAT SHOULD BE DETERMINISTIC:
THE STEPS THAT REQUIRE SYNTHESIS OR JUDGMENT:
WHY THIS APPROACH FITS:
WHAT WOULD MAKE A SIMPLER APPROACH SUFFICIENT:
```

### Step 3. Scope the pilot

Define the smallest version you can test in two weeks:

```
PILOT NAME:
PROBLEM:
THREE TO FIVE PILOT USERS:
TRIGGER:
INPUTS:
OUTPUT:
WHERE THE OUTPUT WILL BE DELIVERED:
DATA ALREADY AVAILABLE:
ACCESS OR APPROVALS NEEDED:
WHAT IS OUT OF SCOPE:
START DATE:
REVIEW DATE:
```

Keep the pilot close to where people already work. A useful output in an existing system is easier to test than a new destination everyone has to remember to open.

### Step 4. Define success

Choose one primary metric and one adoption signal:

```
PRIMARY METRIC:
CURRENT BASELINE:
TARGET:
HOW IT WILL BE MEASURED:

ADOPTION SIGNAL:
CURRENT BASELINE:
TARGET:
HOW IT WILL BE MEASURED:
```

Examples include time saved, conversion lift, output quality, adoption rate, reduced rework, or faster delivery. Use a measure tied to the problem you started with.

### Step 5. Make the decision

At the review date, record what happened:

```
DECISION: CONTINUE / REVISE / STOP
WHAT THE PILOT PRODUCED:
WHAT USERS ACTUALLY DID:
WHAT THE METRICS SHOWED:
WHAT FAILED OR CREATED FRICTION:
WHAT WE LEARNED:
NEXT STEP:
DECISION OWNER:
DATE:
```

Stopping is a valid result. A small pilot that shows the approach does not fit has still done useful work.

---

## Level 2: The folder-based setup

Create one folder for AI pilots:

```
ai-pilots/
  intake/
  active/
    pilot-name/
      problem.md
      approach.md
      pilot.md
      measurement.md
      decision.md
  archive/
```

Use the Level 1 worksheets as the contents of each file. Keep the folder in version control if you can so changes in scope, access, and success measures remain visible.

Before building:

1. Confirm the input data exists and can be used for this purpose.
2. Identify any security, privacy, legal, or brand review needed.
3. Separate deterministic steps from the steps that need synthesis or judgment.
4. Confirm where the output will land and who will use it.
5. Name the person responsible for maintaining the workflow after the pilot.

During the pilot, log user feedback and changes in the relevant file. Do not silently expand the scope. At the review date, complete the decision record and move stopped or completed pilots to the archive.

---

## Level 3: For your AI agent

*Human: paste this whole file to your agent and say "help me scope this pilot." Agent: follow this section literally.*

1. Ask the human to name one recurring process that creates visible friction. Do not recommend a tool yet.
2. Interview the human using the Step 1 worksheet. Do not invent missing facts.
3. Classify the work as automation, workflow with AI, agent, or keep it in the chat. Show your reasoning and let the human approve or change the classification.
4. Identify which steps should be deterministic and which genuinely need synthesis or judgment.
5. Ask where the data lives, whether the human has access, and which approvals may be required.
6. Scope the smallest version that can be tested in two weeks with three to five users. Keep everything else explicitly out of scope.
7. Help the human choose one primary metric and one adoption signal with current baselines where available.
8. Create the folder and files shown in Level 2 only after the human approves the pilot.
9. Do not build, purchase, connect, publish, or deploy anything without separate approval.
10. At the review date, bring the evidence back to the human and ask them to choose continue, revise, or stop. Record their decision verbatim.

---

Questions, or a pilot that taught you something useful? Contact lily@lilyluo.ai.