Cleaning Marketing Data Before AI Touches It
Dirty CRM and analytics data is why AI marketing outputs go wrong. The five-part audit and the cleaning order I use, none of it needing a data team.
Every AI marketing workflow inherits the data underneath it, and most of that data is worse than anyone admits. In Salesforce’s State of Data and Analytics research - 3,800 analytics and IT leaders plus 3,852 business leaders across 18 countries - respondents estimated 26% of their organizational data is untrustworthy, 84% said their data strategies need a complete overhaul before their AI ambitions can succeed, and among those already running AI in production, 89% have experienced inaccurate or misleading AI outputs. That last number is the one I want on the wall: the AI failures teams blame on models are, a large share of the time, data failures wearing a model’s name.
This post is the cleaning workflow I use before pointing AI at a marketing stack - segmentation, personalization, scoring, reporting, all of it. It is written for a marketer, not a data engineer, because the highest-value cleaning in a marketing stack is judgment work: knowing which record is right, which field matters and which campaign name means what. It belongs to the measurement series because clean data is the unglamorous floor under every number in it.
Why AI raises the price of dirty data
Dirty data used to fail visibly and locally: a bounced email, a mangled first name in one send. AI fails with it invisibly and at scale, and that is the change I would ask any team to internalize before their first AI project. A model segmenting on a CRM where a third of the job titles are stale builds confident segments of people who changed jobs. Personalization built on duplicate records greets the same customer as two different strangers. And a reporting model summarizing inconsistently named campaigns aggregates things that should never share a row. The marketing-side symptom shows up in Salesforce’s State of Marketing as the personalization wall: 98% of marketers hit barriers to personalization, and no model fixes a barrier made of data the model cannot trust. Garbage in, confident garbage out - the confidence is what makes it expensive.
The audit: five questions before any cleaning
I audit before cleaning, because cleaning without a map fixes whatever is emotionally annoying instead of whatever is load-bearing. Five questions, each answerable in under an hour with exports and a spreadsheet.
Duplicates: how many people exist twice? Export contacts, sort by email, then by name plus company. The duplicate rate on your top thousand records is your headline number, and the merge decisions are pure marketing judgment - which record is current, which activity history is real.
Staleness: when was this last true? Sample a hundred records and check last-activity and last-updated dates. Contacts silent for two years are not audience, they are archaeology, and they distort every rate and segment they sit in.
Naming: do the labels mean one thing? Pull the distinct values of campaign names, UTM sources and lifecycle stages. If “paid-social,” “PaidSocial” and “meta_ads” coexist, every report and every AI summary over them is quietly wrong. This is usually the most broken and cheapest to fix.
Completeness: are the fields the work needs actually filled? Not all fields - the ones your segmentation and personalization plans depend on. A field 40% filled cannot carry a segment, and knowing that before the AI project starts is the difference between a plan and a surprise.
Sync integrity: do the systems agree? Pick twenty customers and trace them across CRM, email platform and analytics. Disagreements - suppressed here, active there - are the objection-rights problem from the GDPR guide and a modeling problem at once.
The cleaning order that pays first
The audit yields a fix list, and my order is by leverage, not by ease. First naming, because a standard applied today makes every future export legible and costs a document plus discipline. Second deduplication of the records that matter - customers and active pipeline before the long tail, merged by a rule you write down so the next thousand merges are consistent. Third an archive policy: stale records leave the working set (archived, not deleted, with retention rules honored) so every rate, segment and model runs on people who exist. Fourth, targeted field completion only where the audit tied a field to planned work - enrichment for its own sake is how CRMs got dirty in the first place. The sync repairs come last and may be a real systems task: that one goes to whoever owns the integrations as a specific, evidenced request (“these twenty records disagree in these fields between these two systems”), which is a one-day fix framed that way and a quarter-long project framed as “our data is a mess.”
The integration gap is common enough to plan around: in Supermetrics’ survey of 435 marketing leaders, 36% said they lack the systems integration to activate their data and the same share said connecting marketing data needs improvement. If that is you, the honest sequencing is to clean within the systems first - naming, duplicates, staleness pay off inside each tool regardless - and treat unification as its own later decision.
Keeping it clean: ownership beats heroics
A cleaning project without an ownership change is a photograph - accurate the day it was taken. The maintenance layer is small and boring: the naming standard enforced at creation (AI helps here, generating compliant names from a brief description rather than correcting bad ones later, a trick from the marketing ops playbook), a monthly half-hour re-run of the audit’s five questions on new records, and one named owner per system for merge and archive decisions. My rule of thumb for AI projects specifically: the data feeding any live AI workflow gets its audit refreshed quarterly, because the 89% inaccurate-outputs figure is what unowned drift eventually produces.
FAQ
How clean does the data need to be before starting with AI? Clean enough for the specific workflow, which is why the audit maps fields to planned work. Reporting automation needs naming consistency and little else. Personalization needs deduplication and the segmentation fields filled. Do not gate all AI work on total cleanliness - 84% of leaders saying their strategy needs an overhaul is also a description of paralysis, and scoping the cleaning to the workflow is how you avoid joining them.
Can AI clean the data itself? It can accelerate the mechanical parts: proposing duplicate matches, normalizing naming variants, flagging anomalies. The decisions stay human - which record survives a merge, what the standard is, what gets archived - because those encode business knowledge no model has. Same division of labor as everywhere else on this site: the model proposes, you decide.
We have no data team. Is this really doable? The audit and the first three fixes are spreadsheet work plus judgment, and marketing judgment is the scarce ingredient, not SQL. The genuinely technical piece is sync repair, and the audit converts that from a vague complaint into a specific request a developer can schedule.
Does this apply to analytics data too, or only CRM? Same questions, different exports: UTM consistency is the naming audit, untagged campaigns are the completeness audit, and the GA4 setup guide covers the collection side that determines what there is to clean. The channel-level version of staleness is a report full of properties and views nobody has looked at since the migration.
This post is part of my measurement series. It is the floor under the reporting automation workflow and the KPI framework, and the collection side lives in the GA4 + AI setup guide.