How to Scope an AI Project in 90 Minutes

Most AI projects fail at scope, not execution. The team picks the wrong workflow to automate, builds for a use case nobody asked for, or signs up for an open-ended engagement that drags into next quarter. The fix is a tighter front door.

This is the 90-minute conversation we run before every engagement. Run it on your own project before you build, or hand the spec to whoever is building for you.

Block 1 (0-15 min): The job

Write down, in a single sentence, the job the AI system is hired to do.

Bad: "Use AI to improve our customer experience." Good: "Automatically triage 100 percent of inbound support emails into the right category and surface a draft reply for the rep."

If you cannot write it in one sentence, you do not have a job. You have a vibe.

Block 2 (15-35 min): The workflow as it exists today

Walk through the current process step by step. Who does what, in what tool, with what input, producing what output. Time-box each step. Be specific.

This often takes longer than you expect. Do not skip it. Every shortcut you take here becomes a build problem later.

By the end, you should have a flowchart on paper or in a Notion doc. Each step labeled with: actor, tool, input, output, time. The AI candidates are the steps where the actor is a human and the work is repetitive or cognitive.

Block 3 (35-55 min): The leverage matrix

Plot each candidate on two axes:

• Vertical: impact. How much time or money does this step cost per month? Use real numbers.
• Horizontal: AI fit. How well does current AI handle this? Score from 1 to 5.

The top-right quadrant is your roadmap. The top-left is "AI cannot do this yet, revisit in six months." The bottom-right is "AI can do it but the impact is small, skip." The bottom-left is "neither relevant."

Focus on no more than three top-right candidates. More than three and you are building a platform instead of solving a problem.

Block 4 (55-75 min): The build sequence

For the top three candidates, decide order. The order rule:

1. First: the one with highest impact AND lowest technical risk.
2. Second: the one that unlocks the most downstream value (data, workflow, customer experience).
3. Third: the moonshot, only if the first two ship cleanly.

Most teams reverse this. They build the exciting thing first and run out of momentum before the boring-but-valuable thing ever ships. Resist.

Block 5 (75-90 min): The deliverable spec

For the first candidate, write a one-page spec:

• One-sentence job.
• Input (where the data comes from, what shape).
• Output (where it goes, what shape).
• Reasoning layer (which model, which prompt, which guardrails).
• Human override (when the AI hands back to a person).
• Success metric (one number that defines done).
• Cut list (what is explicitly not in scope).
• Timeline (in weeks, not months).

If the spec does not fit on one page, the scope is too big.

What this prevents

Run this and you avoid the four most common AI project failure modes:

1. Wrong workflow. Building for the workflow that sounded cool instead of the one that loses money.
2. Scope drift. No clear deliverable, no clear stop point, no clear success metric.
3. Wrong model. Picking the tool before defining the task. Always backwards.
4. No human override. Shipping an AI system with no graceful fallback when it gets something wrong.

Want this run on your project?

We run this exact conversation as the front door to every engagement. Two paid scoping calls inside our AI Audit + Roadmap product, or the first call of any larger build.

Run the free AI Audit for a 60-second version, or join the Build Queue for the full thing.