Your AI Won't Fix Data You Haven't Fixed First

Your CRM has one count of your customers. Your accounting tool has another. Your inbox has a third. None of them agree, and none of them are wrong exactly — they're just measuring different moments, different definitions, different people's decisions about what to record. Feed any of those three sources into an AI tool and you get a confident answer built on an argument your data is having with itself.

The connector argument is worth taking seriously

The obvious objection is that modern AI tools handle this. You plug in your CRM, connect your accounting software, and the tool ingests everything through a pipeline. No manual migration required. This argument is not stupid. It's the default assumption of most operators who have already paid for structured tools with export capability, and it saves real time.

The problem is that connectors route data. They don't repair it. A survey on data quality dimensions for machine learning identified what researchers call "data cascades" — quality deficits at early stages don't stay contained in the source system. They propagate through pipelines and compound in downstream applications. A connector pulling from a CRM where company names appear in four different formats, where a third of contact records have no industry field, and where duplicate entries exist under slightly different spellings delivers all of that, intact, into the AI's input layer. Faster access to fragmented data is not an improvement.

What the research on model failure actually shows

A study testing nineteen machine learning algorithms across classification, regression, and clustering tasks found that performance dropped sharply as data quality degraded along the dimensions of accuracy, completeness, and consistency. The drop held across all nineteen models. No algorithm was robust enough to compensate for dirty inputs. This matters because the instinct when AI outputs look wrong is to try a different model. The research suggests you're solving the wrong problem.

RAND's interviews with AI project stakeholders identified absent or inadequate data as one of the five leading causes of project failure. Not model selection. Not algorithm design. The data that goes in.

What a governed spreadsheet actually does

Migrating your most critical records into one structured spreadsheet forces decisions you've been deferring. What counts as a customer? Is "ABC Corp" the same as "ABC Corporation"? Which revenue figure is the one you trust? A spreadsheet doesn't answer those questions automatically, but it makes them impossible to avoid. You have to pick a format for company names and apply it to every row. You have to decide which fields are required and which are optional. You have to assign someone ownership of the file.

That process — tedious, unglamorous, done in Google Sheets or Excel at no cost — is what produces complete, consistent, accurate records. Those are the exact three dimensions the nineteen-algorithm study identifies as the primary drivers of model failure when they degrade.

Start with the records your AI use case depends on most directly. If you're trying to forecast revenue, that's your customer list and your transaction history. If you're trying to score leads, that's your contact records and whatever behavioral data you track. Pull from each source, deduplicate, standardize the fields, fill the gaps you can fill, and flag the ones you can't. One file, one owner, one definition of each field.

The spreadsheet is not the destination

The same research that supports this migration explicitly warns against treating a spreadsheet as a long-term system of record. Auditability breaks down at scale. Version control is manual. One person with edit access and a bad afternoon can corrupt months of work. The governed spreadsheet is a transitional structure, not an architecture. It's where you do the quality remediation work that connectors skip, and it's where you build enough trust in your data to know what a more robust system needs to contain.

An operator who skips this step and connects their fragmented sources directly to an AI tool isn't being efficient. They're paying for a model to process an argument their data is still having.