Dirty Data Breaks AI the Same Way It Broke CRM

I wrote a colleague today to reinforce how important data cleanliness is.

Says I, “Data Cleanliness is next to Data Godliness.”

Says She, “Then I will be in Data Heaven!”

We chuckled, but it’s no small feat. I have spent a career watching companies treat that line as a cleanup they will get to later. They do not get to it later. The bad record is already in the system that is about to make the next decision.

It Always Started with Data

On every digital transformation I worked, the project started with data. Not with the software we were about to install, and not with the process map on the wall. The first weeks were spent finding out what the company actually knew about its customers, and how much of that knowledge was wrong.

The pattern did not change when the label on the project changed. It showed up on the implementation of web services. It showed up when we stood up a CRM. It showed up in marketing systems, then in customer data platforms, and it is here now in AI. Same problem. Repeated.

One Failure, Passed from System to System

Web services were supposed to let one system talk to another. They did. They also carried a bad phone number, a merged account, and a status that nobody owned from one database into five more. An open endpoint with no checks was not a technical footnote. It was an entry point that had never been locked down. Once a record left through that door, every system on the other side treated it as fact.

A CRM made the mess visible because sales lived in it every day. Duplicate accounts. A contact attached to the wrong company. An opportunity marked closed in one place and still open in another. Reps stopped trusting the screen and kept their own spreadsheets. The next load was dirtier than the one before it.

Marketing systems inherited that file and used it. The wrong offer went to the wrong person. A customer who had already bought was treated as a prospect. Someone who had asked us to stop was mailed again. Those are client problems. They start as a field that was never required.

A customer data platform (CDP) was supposed to end the argument. One profile. One history. It only holds together when every source that feeds it is disciplined. When the sources disagree, the platform does not referee them. It publishes the disagreement as a single, confident profile. The CDP did not create the error. It gave the error a louder voice.

What Incorrect Data Does to AI and MCPs

AI does something those earlier systems did not do. It decides.

A CRM can hold a bad phone number and wait for a person to notice. A model does not wait. It reads the record, treats that record as the situation, and recommends the next offer, the next answer, or the next action. When the source data is incorrect, the decision is incorrect, and it arrives in the voice of something that has already checked its work.

That is the disaster companies are walking into with AI-enabled systems. The model is not inventing the failure out of nothing. It is acting on a customer record, a product catalog, a policy, or a ticket history that was already wrong. The result is a refund that should not have been issued, a promise the company cannot keep, a forecast a leadership team believes, or an assistant telling a customer something that was never true.

A Model Context Protocol (MCP) makes the failure sharper, not softer. An MCP hands an assistant a tool and a source: the order system, the knowledge base, the CRM, the catalog. The connection can be built carefully. The tool can return exactly what it was asked. If the source is wrong, the answer is still wrong. The assistant did not guess. It retrieved a bad fact from the company’s own systems and acted on it. People trust that answer more than a guess, because it came back looking official.

The earlier generations already cost relationships, campaigns, and quarters. A bad merge lost an account. A bad segment wasted a quarter’s media spend. An AI system with tools does the same things faster, across more customers, without the pause where a person used to say that the record did not look right.

Everything Downstream

When a company is not disciplined about data quality, the pain does not stay with a data team. It shows up as inefficiency, then as client issues, marketing issues, and sales issues. Everything downstream of the bad record inherits it.

Service cannot find the right account, so the customer repeats the story. Marketing spends against people who are not in the market, or against people who have already left. Sales walks into a meeting with a history that contradicts what the buyer remembers. Finance forecasts from a pipeline that was never real. Give that same history to an assistant, and the assistant will defend it.

Most of this is ordinary. A field left optional. A form that accepted anything. An import that matched on a last name. A sync that ran one direction and never reconciled what a person had already corrected.

Where the Errors Get In

On those projects, the holes were rarely exotic. We found the same three, over and over.

• Entry points that were not locked down. A public form, a spreadsheet load, a partner feed, a screen where a rep could save a record with the important fields empty. If the front door accepts garbage, every system behind it stores that garbage and then treats it as the truth.
• Processes that introduced the error themselves. A merge rule that joined two households because the last name matched. A nightly job that wrote an old address back over the one a person had just fixed. A field everyone called temporary that became the field of record because a report depended on it.
• The ordinary drift of a business that kept changing. New products, new regions, new brands, new tools. A status that meant one thing a decade ago now means three things. One customer sits in commerce, in support, and in marketing under three identifiers. Nobody was careless on the day those systems were purchased. The business moved, and the rules for the data did not move with it.

Lock the doors, and the nightly job can still put the error back. Fix the job, and the next acquisition can reopen the argument. The work is all three, kept current, not a one-time cleanup before a launch.

Discipline Before the Next System

The remedy was never the next platform. The next platform arrives hungry, and it eats whatever you feed it. Web services did. The CRM did. Marketing systems did. The CDP did. AI, connected to those systems through MCPs, will do it with more confidence than any of them.

Data cleanliness is still next to data godliness. The chuckle was fair. The work is no longer something a team gets to after go-live, because the system reading the record is no longer waiting for a person to notice.