AI Amplifies Your Data Quality

In The CRM Data-Debt Spiral I described how revenue stacks decay: one rational shortcut at a time, until reports disagree and nobody trusts the numbers. In The Missing Role I argued that this decay is an ownership problem, not a competence problem.

This essay is about what changed in the last two years — why a twenty-year-old problem suddenly got much more expensive. The short version: AI arrived, and AI amplifies your data quality. In both directions.

Three weeks

A while back I watched a company roll out AI-generated account summaries for their sales team. The details here are changed, but the story is real, and I have seen versions of it more times than I can count since.

The demo was genuinely impressive. Point the model at an account, and out comes a crisp briefing: relationship history, open deals, recent activity, suggested talking points. The sales leadership was enthusiastic. The rollout was fast. Week one, people used it before every call.

Week two, the first screenshot appeared in the sales team's group chat. A summary confidently described their best customer as churned — a renewal that had been closed in the wrong pipeline years earlier and never corrected. Everyone on the account team knew the customer was thriving. The database did not.

More screenshots followed. A contact who had left the company three years ago listed as the key stakeholder. Two duplicate records merged into one fictional person with a hybrid job title. Deal values that disagreed with what finance had invoiced, because "closed" meant different things in different pipelines.

Week three, nobody opened the tool. Not because anyone decided to stop. It just quietly became something you double-checked, and a briefing you have to double-check is slower than no briefing at all.

Here is the uncomfortable part: the AI did nothing wrong. It did exactly what it was told. It read the database and summarized it fluently. The summaries were faithful representations of a CRM nobody had trusted for years — delivered with a confidence the CRM itself never had.

The rule, and why it is mechanical

AI amplifies your data quality. In both directions.

This is not a metaphor. It is a description of how the systems work. A language model reading your CRM inherits everything in it: every duplicate, every stale record, every undocumented workaround, every field whose meaning drifted three admins ago. It has no way to know that the "temporary" deal type from 2023 should be ignored, or that everyone mentally excludes the migration-era records from pipeline counts. The tribal knowledge that lets your team navigate a messy CRM — the shared understanding of which numbers to trust and which to wave away — is precisely the thing the model does not have.

So the output is not garbage in the familiar sense. It is worse: it is fluent garbage. A human reading a messy CRM sees the mess and calibrates. A model reading the same CRM produces polished prose with no visible seams. The errors arrive wearing the same confident tone as the truths, which means the reader can no longer use tone as a signal — and trust collapses on the whole output, not just the wrong parts.

Flip the foundation and the same mechanics work for you. On a clean data model with agreed definitions and maintained records, the amplification runs the other way: the model surfaces patterns humans miss, briefings get trusted, usage compounds, and the team gets genuinely faster. The technology is identical. The foundation decides the sign.

The cost nobody budgets for

The wasted license fee is the smallest part of what that three-week rollout cost.

The real cost was political. A sales team that watched AI confidently invent facts about their own customers did not conclude "our data model needs work." They concluded "AI doesn't work here." That conclusion is sticky. The next initiative — even a well-founded one, even on a repaired foundation — starts in a hole, pitched to people holding screenshots.

I think of this as burning the political capital for the next attempt, and I have come to believe it is the largest hidden cost of premature AI rollouts. Budgets recover in a quarter. Credibility with a burned sales team takes years. And in most organizations, the people deciding to ship the AI feature are not the people who will spend those years rebuilding the trust.

If you take one thing from this essay: a failed AI rollout does not leave you where you started. It leaves you behind where you started.

How common is the shaky foundation?

Recently I asked a webinar audience to place their own CRM in the data-debt spiral. These were people engaged enough to attend a session on the topic — a self-selected, above-average crowd. The result:

Seventeen percent in stage 1, where the shortcuts are still fresh. Half the audience in stage 2 — integrations working, but every new connection routing around old shortcuts. A third in stage 3, the fog, where reports disagree and meetings open with a debate about whose number is right. Nobody in stage 4.

Eighty-three percent, in other words, with measurable data debt in their revenue stack. And the stage-2 half is the group I worry about most, because stage 2 is the stage that does not hurt yet. The sync errors are silent. The reports still roughly agree. It is exactly the moment when an AI feature looks like a safe, exciting add-on — and exactly the foundation that will make it fail the way the account summaries failed.

If a room full of people interested in data debt looks like this, the broader picture is not better.

The inversion

All of which leads to a conclusion that still sounds strange when said out loud: the boring work is now the highest-ROI AI investment most companies can make.

Not the model choice. Not the vendor's AI add-on. Not the prompt library. The data model. The definitions — one agreed meaning of customer, of active deal, of closed. The deduplication. The map of what connects to what, and the ownership that keeps all of it from decaying again.

Two years ago this work was worthy and ignorable: it improved reporting, saved some duplicate effort, made audits easier. The cost of skipping it was diffuse. AI concentrated that cost. Every messy record is now a potential confident hallucination in front of your sales team; every undocumented workaround is a wrong fact delivered fluently to someone who trusted the tool. The foundation used to determine how efficient you were. Now it determines whether your AI initiatives survive contact with their users.

The good news is the same as in the previous two essays: this is not a two-year program, and it does not start with buying anything. It starts with ownership — someone with the mandate to hold the whole picture. Then a map. Then a gate. The spiral essay covers the how; the missing role covers the who.

Before your next AI initiative, ask one question — the same one I now ask every company that shows me an impressive demo: would you trust a new employee to learn your business by reading your CRM?

If the answer is no, the AI will not do any better. It will just fail more confidently.