Do you need clean data before AI? Usually not a data lake

Why "clean up the data first" sounds right

Ask a room of owners why they haven't started with AI, and someone will say it: our data is a mess, so we need to clean it up first. It sounds responsible. Nobody wants to pour money into a tool that runs on bad information.

The problem is what happens next. "Clean up the data" turns into a project with no finish line. Someone suggests a data warehouse. Then a data lake. Then a consultant to map every table in the ERP. Six months later, the order desk is still retyping POs, and the AI project hasn't started.

So it's worth asking a narrower question: clean which data, for which AI?

What the failure numbers actually say

AI projects do fail a lot. That part is real.

• In a survey of more than 1,000 people in North America and Europe, S&P Global Market Intelligence found the share of companies scrapping most of their AI initiatives jumped to 42%, up from 17% the year before, as reported by CIO Dive in March 2025. The average company scrapped 46% of its AI proofs of concept before they reached production.
• RAND, in a 2024 report based on interviews with 65 data scientists and engineers, notes that by some estimates more than 80% of AI projects fail, about twice the rate of IT projects without AI. That 80% is an estimate RAND cites, not its own count.
• Gartner predicts that through 2026, companies will abandon 60% of AI projects that aren't supported by AI-ready data. In the same piece, Gartner says 63% of organizations either don't have, or aren't sure they have, the right data practices for AI. That's a prediction and a survey, not a measured result.

Now read the fine print. S&P's respondents named cost, data privacy and security risks as the top obstacles, not messy data. RAND lists five root causes, and data is only one of them. The first one on its list is that people misunderstand, or miscommunicate, what problem the AI is supposed to solve. Another is chasing the latest technology instead of a real problem for the people who'll use it.

Data does matter. In Informatica's 2025 survey of 600 chief data officers, data tied with technology as the top obstacle to moving AI from pilot to production. But look at how close the rest are.

Top obstacles to moving generative AI from pilot to production

Informatica CDO Insights 2025, survey of 600 chief data and analytics officers. Percent naming each obstacle.

1. Data43%
2. Technology43%
3. People35%
4. Process35%
5. Regulations34%

Source: Informatica, CDO Insights 2025.

Data is one problem among five, and not far ahead of people and process. If messy data were the whole story, it would stand far above the rest. It doesn't.

What "AI-ready data" really means

Here's the part that surprises people. Gartner, the same firm warning about the 60%, doesn't describe AI-ready data as spotless data. It says the data has to be representative of the use case, including the patterns, errors and outliers the AI will meet in real work. It also says traditional data management is too slow, too structured and too rigid for AI teams, and that getting data ready is "not one and done."

In plain English: get the data right for the one job you're doing, including the messy cases, and keep fixing it as you go. That's the opposite of a two-year cleanup before you start.

Clean the data the job touches. Not every table in the building.

Two kinds of AI projects, two kinds of data needs

A lot of confusion comes from mixing up two very different kinds of projects.

The first kind trains its own model on your history. Think demand forecasting, predicting which machine will fail, or scoring which quotes will close. These learn from years of past records. RAND's second root cause fits here: many projects fail because the company lacks the data needed to train an effective model. If your sales history has five years of wrong ship dates, a forecast trained on it will be wrong too. For this kind of project, a data cleanup may really need to come first.

The second kind uses a model that already knows how to read. This is most of what a 40-400 person office would automate first: reading customer POs, vendor invoices, order-status emails and packing lists, then drafting the entry in your ERP. These tools don't learn from your history. They read the document in front of them. Microsoft says its invoice model handles invoices of various formats and quality, including phone photos, scans and digital PDFs. Amazon says Textract can pull data from almost any invoice or receipt without templates or setup. And Microsoft's prebuilt models let you add document reading without having to train and build your own models.

If you're a manufacturer or distributor, the messy part of your world is usually the incoming paper: 40 customers, 40 PO layouts, half of them scanned. That isn't a data problem to fix first. That's the work the AI is there to do. We cover the reading side in what intelligent document processing is.

What order entry actually needs

Take a common first project: reading customer POs from email and drafting sales orders. Here's the data it really depends on.

• Your customer list, so the PO can be matched to the right account and ship-to.
• Your item list, with active items and their units of measure.
• A cross-reference of customer part numbers to your part numbers. This is often the gap. It may live in a rep's head or a spreadsheet on someone's desktop.
• Prices or contracts, so the draft can flag a PO price that doesn't match.

That's it. You don't need your 10 years of sales history cleaned. You don't need a warehouse that joins your ERP to your CRM. You need four lists that are right for the customers in the pilot. And your ERP already has three of them.

When the cross-reference is missing, a good setup builds it as it goes. The first time a customer writes their own part number, a person picks the matching item. The system remembers that pairing. Over time, more and more lines match on their own. That's data cleanup too, but it happens inside the work instead of before it.

Where "clean it first" has a point

To be fair, the skeptics aren't wrong about everything.

• Duplicate customers and items will bite. If one customer exists under three account numbers, any automation will pick one, and maybe the wrong one. Fix the duplicates for the customers in your pilot before you start.
• Prices you can't trust make price checks useless. If half your contract prices are out of date, the system will flag everything, and your team will start ignoring flags.
• Some data really is missing. RAND's point about not having enough data stands for prediction projects. If you want a forecast, you need the history.
• Systems matter too. RAND's fourth root cause is not having the setup to manage data and run finished models. If your ERP can't accept an order from outside, that's worth knowing early.

The fair answer isn't "data doesn't matter." It's "the data for this job matters, and you'll find out which data that is faster by starting small than by auditing everything."

Where the returns show up

There's one more reason to start with a reading-and-drafting project instead of a big data platform. MIT's 2025 study of generative AI, as reported by Fortune, found the biggest ROI in back-office automation, even though more than half of generative AI budgets went to sales and marketing tools. The same report found that buying from specialized vendors and building partnerships worked about 67% of the time, while internal builds succeeded about a third as often. The lesson for a smaller company: the money is in the boring work, and you don't have to build it alone.

A 30-day test instead of a data project

If you're stuck at "our data isn't ready," try this instead:

1. Pick one process. Customer POs, vendor invoices or order-status emails. One, not three. Our guide on what to automate first can help you choose.
2. Pull 50 real documents from the last month, including the ugly ones.
3. List the reference data each one touches: which customer, which items, which prices.
4. Run them through and keep a person on approval. Every time something fails, write down why: bad scan, unknown part number, duplicate customer, wrong price.
5. Fix only what broke. At the end of the month, you'll have a short, specific list of data to clean, ranked by how often it got in the way.

That list is your data project. For most offices it's a short, specific job for someone who knows your customers, not a six-month platform build. And the order desk gets help while it happens.