Data quality for AI decides whether an AI project produces reliable results, more than the choice of model or tool. This article explains why poor data is a business problem, and sets out WeAreBrain’s method for assessing data before scoping any AI use case: a three-category sorting exercise and four questions for every data source.
Key takeaways
- Data quality for AI matters more than the model: unreliable AI outputs usually trace back to incomplete, outdated, inconsistent or irrelevant data.
- Sort available data into useful, partial and unusable before scoping any use case; the exercise takes about an afternoon.
- Ask four questions of every data source: is it up to date, reliable, structured and easily accessible?
- Operations managers and sales administrators are often the most useful people in a data audit, because they know exactly where the gaps are.
The conversation about AI in business tends to start with tools. Which platform, which model, which subscription to approve. In our experience, that’s the second conversation worth having. The first, and the one that most determines whether a project succeeds, is about data: what you actually have, and whether any of it is ready to be useful.
When AI projects disappoint, the post-mortem tends to focus on the model, the vendor, or the implementation timeline. Yet the evidence is consistent, and the pattern in our own client work confirms it: the gap between AI ambition and AI results is almost always a data gap.
A 2025 Gartner report found that 63% of organisations either do not have, or are unsure whether they have, the right data management practices for AI. Gartner also predicts that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data. These are not edge cases or early-adopter mistakes. They represent the dominant experience of organisations that moved straight to tools before understanding what their data could actually support.
Why data quality is a business problem, not a technical one
The phrase “quality in, quality out” tends to get filed under technical concerns, something for the data team to sort out before the real work begins. In practice, it’s a business and decision-making challenge long before it becomes a technical one.
When a model produces unreliable outputs, the instinct is to question the model. The more likely explanation is that the data feeding it was incomplete, outdated, inconsistent, or simply not relevant to the task. According to the IBM Institute for Business Value, only 29% of technology leaders are confident their enterprise data meets the quality, accessibility and security standards needed to scale generative AI. A separate IBM study of 1,700 Chief Data Officers found that only 26% believe their data can support new AI-enabled revenue streams, despite the majority having data strategies formally integrated with their technology roadmaps.
The Informatica CDO Insights 2025 survey puts data quality and readiness as the number-one obstacle to AI success, cited by 43% of organisations, ahead of technical maturity and skills gaps. This tells us something important: the organisations closest to enterprise data, the people whose job it is to manage it, identify readiness as the primary constraint. Managers investing in AI tools deserve the same clarity before they begin.
The three-category sorting exercise
The practical starting point we use with clients comes from our AI in Practice workbook, and it’s deliberately low-tech. Before any scoping, tool evaluation, or use case definition, we ask teams to sort their available data into three categories.
- Useful data is complete, current, consistently formatted, and accessible without significant manual effort. It can be fed into a process or model with reasonable confidence in the output.
- Partial data exists and has value, but requires work before it’s usable. This might mean incomplete fields, inconsistent formats across systems, or data that’s accessible in theory but siloed in practice.
- Unusable data is outdated beyond relevance, so unstructured that processing it is impractical for the use case at hand, or simply not connected to the problem being solved.
The sorting exercise takes an afternoon with the right people in the room. Its value is proportional to how honestly the team engages with it, particularly with the partial and unusable categories, which tend to be more revealing than the useful column. The full exercise, with worked prompts for each category, is in Chapter 3 of the AI in Practice workbook.
The four questions we ask about every data source
Once the sorting exercise is complete, we go deeper on each data source that lands in the useful or partial categories. These are the four questions that shape every data conversation we have at the start of a new engagement.
| Question | What to look for | Red flag |
| Is it up to date? | Data reflects current reality and is refreshed at a cadence that matches the use case | Data is months or years old and hasn’t been maintained |
| Is it reliable? | Consistent input processes, clear ownership, low error rate | Multiple teams entering data differently, no single source of truth |
| Is it structured or free-form? | Organised in fields or formats a system can process without heavy manual handling | Dense free-text, PDFs, inconsistent naming conventions |
| Is it easily accessible? | Available without significant extraction effort or dependency on another team | Locked in legacy systems, requires manual export, or lives in someone’s spreadsheet |
All these questions require honesty and access to the people who actually work with the data day to day. In our experience, the most useful person in the room during a data audit is often an operations manager or a sales administrator, not an analyst, because they know exactly where the gaps are.
A note on sensitive data
Every data audit surfaces some data that requires a different kind of attention: personal records, commercially sensitive information, or data subject to regulatory requirements such as GDPR. The practical reason to identify this early is straightforward. If a proposed use case depends on data you cannot legally or ethically feed into an AI model, that’s a scoping constraint you need to know about before the project is designed around it.
This is worth a focused conversation during the sorting exercise, separate from the quality questions. Not as a legal review, but as a realistic assessment of what’s actually available to work with.
Where to go from here
A data audit is a starting point, and the goal is clarity rather than perfection. Most organisations will find a mix of useful data they can act on now, partial data worth investing in, and gaps worth acknowledging before they become surprises mid-project. That picture is far more useful than a tool evaluation.
The full sorting exercise and data quality prompts are in our AI in Practice workbook, available to download free. It’s designed for managers who want to move into AI with a clear view of what they’re actually working with. The first question we ask a new client is what data they’re sitting on and whether it’s ready to do anything useful.
If you’d like to work through that assessment with your team, our AI strategists can help you move from a data audit to a scoped, ready-to-run use case. Get in touch to find out how we work.
Frequently asked questions
What is AI-ready data?
AI-ready data is data that is current, reliable, consistently structured and accessible enough to feed an AI process or model with reasonable confidence. AI-ready data also needs clear ownership and must be usable legally and ethically. Gartner found 63% of organisations lack, or are unsure they have, the right data management practices for AI.
How long does a data quality assessment for AI take?
WeAreBrain’s three-category sorting exercise takes about an afternoon for a first pass, sorting data into useful, partial and unusable. Applying the four questions to each useful or partial source takes longer, depending on how many systems hold the data, but the first pass already shows where the biggest gaps sit.
Can AI work with unstructured data?
AI can work with unstructured data such as documents, emails and free text, but dense PDFs, inconsistent naming and free-form fields usually need preparation first. Data that a system cannot process without heavy manual work belongs in the partial category until it is cleaned, structured or indexed.
Can you use personal data to train or run AI?
Personal data can only feed an AI system when its use is lawful under GDPR and ethically acceptable for that purpose. Data audits often uncover personal, commercially sensitive or regulated data, so a use case that depends on such data should be identified and assessed before the project is designed.
How do you improve data quality for AI?
Improve data quality for AI by starting with the data your first use case depends on. Assign clear ownership, establish a single source of truth, standardise formats and set a refresh cadence that matches the use case. Investing in partial data that is valuable but incomplete usually gives the fastest return.
