I’m a big fan of connecting tools like Claude or ChatGPT directly to Salesforce or HubSpot. For most teams, that’s where AI starts becoming genuinely useful. Instead of asking someone on the RevOps team to build another report, an end user can simply ask questions like: Where did our best leads come from last quarter? How many opportunities do we have in manufacturing? Which campaigns generated the most pipeline this year? Giving people the ability to query and report on CRM data on demand is a huge improvement.

But there’s a catch. Imagine your CRM has three different fields called Lead Source — one on Leads and two on Contacts. They were probably created for perfectly reasonable reasons over the years. Different teams needed different things, someone added a field, someone else built an integration, and another field became the one Marketing actually uses. Humans who have worked in the CRM for years often know these quirks. AI doesn’t.

Ask three people the same question using their AI tools and they may get three different answers, all based on seemingly legitimate CRM data. That’s why I don’t think the first step in AI-enabling a CRM is actually connecting the AI. The first step is simplifying the CRM.

Before we build the AI connection, we still build reports and dashboards directly in Salesforce or HubSpot. These aren’t necessarily the dashboards executives are going to look at every Monday morning. They’re closer to audit and data-quality reports. We want to know which fields are actually being used, which contain duplicate or conflicting information, whether important fields are populated consistently, and whether Sales, Marketing and RevOps agree on what those fields mean.

One of the most important tests is simple: can we build a report in the CRM today that everyone agrees is correct? If five people can’t agree on the numbers in a Salesforce report, connecting ChatGPT to Salesforce isn’t going to solve the problem. It may just make the disagreement happen faster.

The goal isn’t to make the CRM perfect. That’s usually unrealistic. The goal is to establish a reasonably small set of fields, definitions and reports that people trust. Lead Source means this. Pipeline means this. An active customer is defined this way. This field is the authoritative field for Industry.

Once the team generally agrees on those definitions, AI becomes much more interesting. Now we can connect the AI tools to the CRM and start building shared prompts, reports and other AI artifacts on top of a trusted foundation. When someone asks, “How much pipeline did we generate from events this quarter?” there’s a much better chance that everyone gets the same answer.

There’s a tendency to make AI readiness sound like a big technology initiative. For CRM teams, I think it’s often much simpler. Clean up the important fields, agree on definitions, and build reports that prove the data works. Then connect the AI.

The AI integration may be the exciting part, but the boring CRM cleanup is what makes it trustworthy.