You paid for the CRM. You paid for the implementation. You paid, quietly and for years, for the seats.
And when you open it, what you find is 6,400 contacts, 4,100 of which have no last-activity date in the past eighteen months, 900 of which are duplicates, and a lead status field where the most common value is blank. The pipeline shows $2.3 million in open deals, and you know for a fact that at least half of those closed or died last year.
The natural next thought — the one every vendor in the market is counting on — is: we should put AI on top of this.
Please don't. Not yet.
Why "dirty data" is the wrong diagnosis
"Dirty data" implies a hygiene problem, which implies a cleanup project, which is why so many companies spend a miserable quarter deduping records and then watch the exact same decay happen again over the following twelve months.
The data isn't dirty. The data is a perfectly accurate reflection of a system that nobody has a reason to use.
Think about it from the agent's side. She closes a deal. To log it properly she has to open the CRM, find the contact, update four fields, attach the note, move the stage, and set a follow-up task. That takes six minutes. What does she get for those six minutes? Nothing. She already knows what happened. The CRM gives her back exactly zero — no reminder she needed, no insight she didn't have, no commission she wouldn't otherwise get.
So she doesn't do it. Not out of laziness. Out of a completely rational assessment that the input costs more than the output is worth.
Every empty field in your CRM is a small, honest vote about the value of that field.
What AI does to a system nobody feeds
It amplifies the emptiness.
If you build a lead scoring model on top of a database where 70% of the activity history is missing, you'll get scores. They'll look confident. They'll be nonsense, and worse, they'll be nonsense with a number attached, which is far more dangerous than an obvious blank.
If you build a "which past clients should I call this month" agent on top of a contact list where half the records haven't been touched since 2022, it will recommend that your top producer call people who moved out of state, bought elsewhere, or died. She'll do it twice, get embarrassed once, and never open it again — and you will have permanently spent the goodwill you needed for the next thing you build.
This is the mechanism by which good companies conclude that AI doesn't work for them. It isn't the model. It's that the model was pointed at a system of record that was never a record of anything.
Fix the incentive, not the database
The fix is not a cleanup sprint. The fix is to make the CRM give something back, so that using it stops being an act of charity.
Here's the reframe that works: stop asking humans to feed the system, and start having the system feed itself.
The call happened — capture it and log it, automatically, without asking anyone to type. The email thread with the buyer's lender exists — parse it and attach it to the deal. The showing was scheduled in the calendar — that's an activity, record it. The text message went out from the company line — that's a touch, log it. Almost every meaningful interaction in a modern brokerage already exists as data somewhere. It's just not in the box you paid for.
That's an integration problem and a parsing problem, and both are extremely solvable now in a way they weren't five years ago. The result is a CRM that populates from reality instead of from willpower.
Then make it pay the user back. If the system knows the buyer went quiet for eleven days, it should tell her that on Monday morning, in a list, with a suggested next step and a draft message. Now the six minutes she never spent are being repaid with a deal she would have lost. Now the CRM is useful. Now she opens it.
Usage follows value. It has never once followed a policy.
What to do this month
Don't buy anything. Do this instead.
Pick one field that actually matters — the one that, if it were reliably populated, would change a decision you make. Not "industry." Something like next follow-up date or deal stage or source. One field.
Ask why it's blank. Sit with the person who's supposed to fill it and ask what it would take. The answer will almost never be "more training." It will be "I have no idea what to put there," or "it duplicates something I already track elsewhere," or "nobody has ever looked at it."
Then either automate the capture of it or delete the field. Both are progress. A field nobody fills is worse than no field, because it teaches everyone that the system is optional.
The bigger point
There is a version of this business where your team spends its energy on relationships and the systems quietly keep up with them. And there's the version most brokerages are living in, where people spend their energy feeding software that gives them nothing, and eventually stop.
AI is very good at the first version. It is worse than useless in the second, because it puts a confident face on an empty room.
Get the system feeding itself. Then, and only then, put intelligence on top of it — and it'll be intelligence built on what actually happened, which is the only kind worth having.