Is your CRM ready for AI? Run this check first.

Most CRMs aren't ready for AI yet. The people selling you the AI have no reason to say so.

Here is what "ready" means in practice. AI reads your CRM the way a brand-new hire would, except it never asks a clarifying question and never gets a bad feeling about a record. It takes the fields at face value. If your deal stages mean nothing and half your properties are empty, the model inherits all of it, and then answers with total confidence anyway.

That's the part the demo skips. AI on top of a messy CRM doesn't fail. It fails faster, in a fluent and plausible voice that is much harder to argue with than a blank cell.

So before you buy anything, run four checks. They cost you an afternoon. Getting it wrong costs a lot more, and it costs you later - once it's wired into everything.

Check 1: Can you trust what's in the fields?

Open ten records at random. Read them the way a stranger would. Do the fields agree with each other, or does one customer look like three different companies depending on which property you believe - one says they're a lead, one says closed-lost, one says active?

Clean data isn't a one-time state, either. HubSpot's own data puts B2B contact decay at about 22.5% a year. So a fifth of what you have is quietly going out of date on its own, every year, with or without AI on top.

A person reading a contradictory record pauses and asks someone. An AI doesn't pause. It picks the reading that fits and moves on.

Check 2: Do your objects and stages mean anything?

Structure is the part AI leans on hardest. If a deal in "Negotiation" sometimes means a signed contract and sometimes means one hopeful email, the stage is decoration. A person learns to read around that. A model treats the label as the truth.

If you can't write a one-line definition of each stage that every rep would agree with, your stages aren't defined - they're just labelled.

Same for your objects. Company, contact, deal, ticket: each should mean one thing, cleanly. When the shape underneath is vague, everything built on top inherits the vagueness. This is why the model matters more than the data itself, a point worth reading in full in our note on the messy HubSpot CRM.

Check 3: Is the process written down anywhere?

Ask three people how a lead becomes a customer. If you get three answers, you don't have a process. You have a folklore, passed down and slightly different each time.

AI can automate a process that's written down. It cannot infer one that lives only in a senior rep's head. What you hand a model is the boring, explicit version: this happens, then this, and here is the rule for the awkward edge case everyone currently just "knows."

In my experience, undocumented process is the most common reason an AI project stalls in the first month. Nobody can tell the model what good looks like, because nobody ever had to write it down before.

Check 4: Do you own the model, or just rent the answer?

This one is quieter, and it compounds. A lot of AI features hand you an output with no way to see the reasoning. It scored the lead a 90. It will not tell you on what.

When the logic lives inside a vendor's black box, you can't audit it, correct it, or take it with you. When you own the model - the rules, the examples, the definition of a good outcome - you can do all three. A year in, that difference is the whole game.

This is the real case for a custom build over a bolt-on. Not that it's cleverer. That you can open it up and fix it when it's wrong, and it will be wrong sometimes. Owning it means doing something about that, instead of filing a support ticket and waiting.

What "ready" looks like when you put it together

Fields you can trust. Objects and stages that each mean one thing. A process written down in plain sentences. A model you can open up and inspect.

Notice that three of the four have nothing to do with AI. They're just a CRM built on purpose instead of by accretion. The AI part is almost the easy bit, once the foundation actually holds weight.

That's the uncomfortable order of operations. Most teams want to buy the AI first, because it's the exciting purchase, and fix the data later, because that's the tedious one. It runs the wrong way round. The foundation is what makes the AI worth anything, and it's the part no vendor is paid to slow down and check. Having someone check it with you before you spend is a fair amount of what our services actually are.

AI is a very fast reader. Most CRMs are handing it a book that contradicts itself every other page - and then acting surprised when the summary comes back wrong.

Common questions