Your AI Strategy Is Fine. Your Data Strategy Isn't.

Your AI Strategy Is Fine. Your Data Strategy Isn't.

Here's a pattern I see constantly: a company invests in an impressive AI tool. The demos were convincing. The vendor promised fast time-to-value. Three months later, the results are underwhelming and everyone is looking for someone to blame — usually the technology.

The technology is rarely the problem.

I've yet to encounter a mature LLM or automation platform that simply doesn't work. What I have encountered, repeatedly, are data environments that would make any AI system look incompetent. The real bottleneck in most AI implementations isn't intelligence — artificial or otherwise. It's the mess underneath.

What "Good Enough" Data Actually Looks Like

Let me be clear about something: perfect data is a fantasy, and waiting for it is a strategy for doing nothing. Every organization has data problems. The question isn't whether your data is perfect — it isn't — but whether it's good enough for the specific thing you're trying to do.

Here's my working definition of good enough:

Consistent: The same concept is represented the same way. "Customer" isn't also "client" and "Kunde" and "Klient" in the same database. Fields mean what they say. Formats are predictable.

Complete enough: The fields that actually matter for your use case are filled in. Not every field — the relevant ones. For most applications, 85%+ completeness on the critical fields is workable.

Not embarrassingly stale: Your data reflects the current world, not the world as it existed three years ago. Dead contacts, discontinued products, outdated prices — these aren't just inconvenient. They actively mislead AI systems that treat them as ground truth.

If your data meets these three criteria for your specific use case, you can probably start. If it doesn't, no amount of AI investment will save you.

The Three Ways Bad Data Kills AI Projects

Inconsistency creates pattern blindness

AI systems learn by finding patterns. When the same thing is represented differently across your data, the patterns disappear. A model trying to learn from sales data where "won," "closed," and "✓" all mean the same thing will struggle to learn anything meaningful.

This sounds like a data entry problem. It is — but it's also a governance problem. Someone needs to own naming conventions and enforce them. That's not a technology decision.

Completeness gaps appear exactly where you need them

I've seen this too many times: a company launches an AI-powered recommendation engine, only to discover that the feature their model needs most is missing for 40% of records. The system works perfectly on the other 60% and produces garbage for the rest.

The missing data isn't random. It tends to cluster around information that was inconvenient to collect — which is often exactly the information that would be most predictive.

Stale data teaches AI the wrong lessons

An AI model working with last year's data will make last year's decisions. For slowly-changing domains, this is fine. For anything competitive, customer-facing, or price-sensitive, it's a serious problem.

The compounding issue: AI systems are often confident even when they're wrong. A human looking at an outdated record might pause and double-check. An LLM will generate a beautifully formatted, completely incorrect answer.

The Two Failure Modes I See Equally Often

The rescue fantasy: "Our data is a mess. Maybe AI can organize it." AI can help with data cleaning and normalization — but it can't fix a broken data culture. If bad data is being generated faster than it can be cleaned, you have a process problem, not a technology problem.

The perfectionism trap: Waiting for data to be "ready" before starting any AI work. Data is never ready. If you require perfect data before launching, you'll be in perpetual preparation.

The pragmatic path lies between: identify a specific use case, assess whether the data that matters for that use case is good enough, and start there. Let real-world results tell you where the gaps are.

A Quick Data Readiness Check

Before committing to any AI project, I run through five questions:

1. Where does the data live? How many systems, how many formats, how many owners?

2. How consistent is it? Pick a key concept — "customer," "product," "project" — and search for variants. How many do you find?

3. How complete is it? Sample 50 records. What percentage have the fields you'll actually need filled in?

4. How current is it? When was the average record last updated? How many haven't been touched in over two years?

5. Who owns data quality? If the answer is "everyone" or "nobody," you have your answer.

This isn't a comprehensive audit. It's a sanity check. If any of these questions produces an alarming answer, address it before you start building AI on top.

Where AI Actually Works With Imperfect Data

Some use cases tolerate data imperfection much better than others:

Text-based applications — writing assistance, document summarization, email drafting — depend on the quality of the input text at the time of use, not historical database records. A well-written brief produces good output regardless of what's in your CRM.

Document processing — extracting structured data from PDFs, invoices, contracts — works on the document in front of it. Your broader database quality is mostly irrelevant.

Internal knowledge retrieval (RAG systems) works well if your source documents are reasonably well-organized, even if other data is a mess. If your handbooks and process docs are consistent, you can build something useful immediately.

Pick the use case where your data is already solid. Build there first. Use that success to earn the investment needed to clean up the rest.

Can AI Actually Fix Your Data Problems?

Yes — and this is one of the most underrated use cases for AI, one that delivers value before your main AI project even starts.

The caveat from earlier still stands: AI can't fix a broken data culture. But it can significantly improve the state of your existing data — faster and more thoroughly than any manual cleanup project.

Deduplication and entity resolution. Probably the highest-leverage application. LLMs can merge records that humans would instantly recognize as identical but that rule-based deduplication algorithms miss: "Acme Corp," "ACME Corporation," and "Acme Corp. Ltd." are trivially the same to a language model. Tools like OpenRefine with AI extensions, or direct API calls against database exports, can process thousands of records in hours instead of weeks.

Normalizing inconsistent formats. Phone numbers in ten different formats. Country codes as abbreviations and full names. Dates in DD/MM/YYYY and MM-DD-YY mixed throughout the same column. AI standardizes these batch by batch, reliably. A language model understands that "Feb 29, 24," "02/29/2024," and "29-02-24" are the same date and consistently outputs the same format.

Gap-filling through contextual inference. If a CRM record has a complete company address but no industry classification, AI can often infer a plausible industry from the company name and domain. This isn't guaranteed to be correct, but it's a reasonable starting point that can be spot-checked. Better than an empty field blocking your model.

Enrichment from external sources. Tools like Clay, Apollo, or Clearbit combine AI with external databases to automatically fill missing fields — company size, LinkedIn profiles, revenue ranges, tech stack. This isn't a build-it-yourself project, but for sales CRMs it's often the fastest path to usable data.

Anomaly detection. AI finds records that look "wrong" — a German postal code in a US address field, a birthdate from 1800, an email without an @ symbol. Standard validation rules catch many of these, but AI also spots semantic anomalies: a record flagged as "new customer" with contract dates from 2015 is probably a data error, not a time traveler.

The practical workflow: export your most problematic data table → AI-assisted cleanup → manual review of a sample set → import corrected data. This cycle can be completed in one to two weeks and creates a significantly better foundation for the actual AI project.

The catch remains: it's a one-time cleanup. If the input processes don't change, the data will be just as messy in six months.

The Bottom Line

Data quality is strategy — not a technical detail, not a data team problem, strategy. The organizations that extract the most value from AI are the ones making data governance decisions today, while competitors are still debating which model to use.

The technology will keep improving regardless. Your competitive advantage isn't access to the best model — everyone has access to the same models. It's the quality of the proprietary data you feed them.

Not sure where your actual gaps are? Our free Automations Check takes 30 minutes and will tell you honestly where AI can add value in your business right now — and where you need to fix the foundations first.