The Checklist You Must Run Before Your First AI Project

Most AI projects do not fail because the technology is wrong. They fail because nobody wrote down how the process actually works before the build started. Discovery is where projects are won or lost, and it is the step most teams skip.

Pick One Workflow, and Only One

Before anything else, name the single workflow you want to automate. Not a department. Not a category of work. One process, with a clear start and a clear end.

The right first candidate is repetitive, rule-based, and annoying. Someone on the team does it with a sigh. The inputs arrive in roughly the same shape each time. If you cannot describe the workflow in two sentences, it is not ready to automate.

Do not pick the most complex or most strategic process first. Pick the one where a mistake is recoverable and where a win will be visible quickly.

Write the Workflow Understanding Document

At Focus AI, every engagement starts with a Workflow Understanding Document. This is a written map of how the process works today, before any AI is designed. It is not a flowchart. It is a description of reality.

The WUD answers these questions:

• Who does this? Name the person or role, not the team. There is always one person who actually knows all the edge cases.
• What are the real inputs? PDFs, emails, spreadsheets, verbal instructions? What format are they actually in, versus what format they are supposed to be in?
• What are the exceptions? Every process has a "but sometimes" buried inside it. Find them now or find them broken in production. Ask: what happens when the input is wrong, late, or missing?
• Where does the output go? Which system, which person, in which format?
• How often does a human override the normal path? If the answer is "fairly often," that is a signal the process is more complex than it appears.

You cannot write this document by interviewing a manager. You write it by sitting with the person who actually does the work.

Audit Your Data Before You Commit

An AI system is only as good as the data it can legally and reliably access. Find out who can touch the data before you promise anything to anyone.

Data readiness is the most common hidden blocker. Run this audit before any technical work begins:

• Where does the data live? On-premise, a cloud provider, a SaaS tool with API restrictions?
• Who can access it? Not in theory. In practice, today, with current permissions.
• What are the GDPR implications? If personal data is involved, you need a lawful basis to process it through an automated system. This is not optional in the EU.
• Is the data structured? Structured data in a database is straightforward. Handwritten notes scanned to PDF are a project in themselves.

Cost the Manual Version Honestly

Before you budget for AI, write down what the manual process actually costs: time per run, frequency per week, loaded hourly rate, and what errors cost to fix.

This number does two things. First, it tells you whether automation is worth the investment at all. Some processes are cheap enough to stay manual. Second, it gives you a baseline to measure against after launch. Without it, you will never know if the project worked.

Discovery is not glamorous. It is writing, asking uncomfortable questions, and resisting the urge to jump to a solution. But it is the only way to build something that survives after the handoff.