<aside> 📌 Checklist | 10 min | Intermediate

</aside>

<aside> 📋 Most AI pilots don't fail because of the model or the strategy. They fail because the infrastructure underneath can't support them. Work through these 12 questions before your team spends another quarter finding that out the hard way.

</aside>

<aside> 💡 Before you start: Pull up your last AI pilot's post-mortem (or the Slack thread where the conversation quietly died). Keep it open. You'll need it.

</aside>

Start Here: The 10-Minute Diagnostic

Do this first. Open your data pipeline diagram, or ask your lead engineer to share their screen and walk you through where your core operational data lives right now. Not where it's supposed to live. Where it actually lives. If that conversation takes longer than 10 minutes to get started, or if nobody can find the diagram, that's already your answer to several questions below.

S&P Global found that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% in 2024. The infrastructure gap is the most consistent reason. You can diagnose yours in 10 minutes. You just have to be willing to look.

Section 1: Your Data Is the Model's Ingredient List

• [ ] The data your AI pilot needs to consume exists in one place, with one owner, and it doesn't require a manual export to access it. (If someone has to run a script or open three different tools to pull it, check this off as a no.)
• [ ] Your team can describe the data's lineage: where it originates, what transforms it, and what shape it's in when it arrives. Not at a whiteboard-theory level. They can actually show you.
• [ ] The last time someone checked your training or input data for quality issues, they found the data was clean enough to use without a cleanup sprint first.
• [ ] Your data doesn't live primarily in spreadsheets maintained by one person who has been at the company for seven years.

Section 2: Your Systems Actually Talk to Each Other

• [ ] The core systems your AI feature needs to read from or write to have documented, stable APIs. Not "we have an API" but "here's the spec, here's the auth model, here's the rate limit."
• [ ] When your team connects a new service to your stack, it doesn't require a two-week custom integration every single time because each system communicates differently.
• [ ] Your team shipped something in the last six months that required two or more internal systems to exchange data, and it worked without a human manually transferring that data at any step.

Section 3: Your Team Knows What They're Working With

• [ ] A new senior engineer joining your team today could understand why your most critical module was written the way it was, without having to ask the one person who built it.
• [ ] If your most system-knowledgeable engineer left tomorrow, you could reconstruct their mental model of the architecture from documentation that already exists.
• [ ] Your team can confidently describe which parts of the codebase are stable enough to build AI features on top of, and which parts would make that dangerous right now.

Section 4: Your Infrastructure Can Handle What AI Actually Asks of It