Connecting Your Legacy CRM to AI: A Practical Guide

Here's a situation we run into constantly. A client has been on the same CRM for five, eight, twelve years. Their team is trained on it, their pipeline data lives in it, their custom fields and workflows encode a decade of institutional knowledge. And it doesn't have an AI integration, because it's not one of the five platforms that the big AI vendors built native connectors for.

The client has two options, as they see it. Option one: migrate to a modern CRM that has AI built in. Option two: live without AI in the CRM and accept the productivity gap. Both options are terrible.

There's a third option that nobody tells them about: connect the existing CRM to AI by writing a small custom integration layer. Here's what that actually looks like in practice.

This is the step everyone skips, and it's the most important one. 'Connect AI to the CRM' is not a goal. 'Let our sales team ask Claude to summarize a prospect's history before a call' is a goal. 'Have AI draft follow-up emails based on the latest meeting note' is a goal. 'Flag deals that have gone quiet for more than two weeks' is a goal.

Write down the three to five real workflows you want AI to support. If you can't come up with any, your team probably isn't using AI enough yet and you should start there first.

Every CRM has some way to get data in and out. In rough order of preference: a modern REST API, an older SOAP API, a data export / import mechanism, a direct database connection, or (worst case) a CLI or scraped UI. Figure out which category yours falls into. The good news: almost every CRM released in the last 20 years has at least one of these.

Also figure out what authentication looks like. API keys? OAuth? Basic auth with a service account? This will determine how you isolate AI access from regular user access, which matters a lot when you get to step 5.

Here's where the actual work happens. You write a small service — we usually use Python or Node.js — that exposes a handful of capabilities to AI via Model Context Protocol. The capabilities are the ones you identified in step 1, translated into specific operations: 'search contacts by name', 'get recent activity for contact ID', 'list deals in stage X', etc.

Keep it small. The temptation is to expose everything the CRM can do. Don't. Expose the minimum set of operations that enables your real workflows. More operations means more attack surface and more confusion for the AI about which tool to use when.

If your team uses Claude Desktop or ChatGPT Enterprise, this step is a config file. If you're building something more custom, it's an API call away. Either way, this is the fastest step — usually an afternoon.

This is where most amateur integrations break. The AI agent can now read (and potentially write to) your CRM. You need: audit logging so you can see every action the AI took, credential isolation so the AI is operating with a service account that can't do anything you don't want it to, read-only by default for most operations, with write operations requiring an explicit confirmation step.

Get this right and AI becomes a safe, useful member of the sales team. Get it wrong and you're one typo away from an AI deleting last quarter's pipeline.

A first-pass integration for a single workflow is typically a 2-4 week engagement. Ongoing maintenance is minimal if you scope the initial build well. Compare this to the alternative: a full CRM migration, which is a 6-12 month project with a very real chance of failure and an enormous productivity hit during the transition.

If you're sitting on a CRM that works fine and you're being told you need to rip it out to 'get AI', the right answer is almost always to connect what you have instead. We do this for a living — reach out if you want to talk about whether it's a fit for your business.