Before the question “Which model?” can be answered, another question arises: Is there even enough data to begin with? Most teams answer this intuitively, saying, “Actually, yes.” This is precisely where the most expensive part of many AI projects begins: the work of cleaning up data that no one checked before the project started.
Five criteria provide a reliable answer to the initial question without having to wait weeks. This guide explains what each criterion means, how to determine whether it is valid, and how to conduct the assessment in your own company.
What role does data preparation play in the success of an AI project?
Experience from past projects shows that 70 to 80 percent of project time is spent on data preparation, not on the model itself. Industry reports go even further, estimating that 70 to 85 percent of all AI projects fail to meet their goals, with the data set being the most common cause. Gartner estimates that poor data quality costs an average of $12.9 million per year per company.
These figures are not a reason to postpone an AI project. They are a reason to plan for the effort in advance rather than discovering it during the project. Those who are familiar with the five key areas before launch will avoid shifting the data work to the most expensive phase of a project: troubleshooting during live operations.
What happens if you skip the exam?
The difference usually only becomes apparent after the project has started, when making corrections is significantly more expensive than conducting a review beforehand:
Which five areas determine AI readiness?
The following five areas cover all the aspects that should be clarified before embarking on an AI project. Each area can be evaluated independently of the others; together, they provide a complete picture.
Data quality: The foundation
Data quality checks whether the same metric yields the same value across all systems, whether values are regularly missing, and whether a number remains plausible from a business perspective. A revenue metric that is calculated differently in Sales than in Controlling will produce contradictory results in any AI application, regardless of how good the model is.
Domain knowledge & semantics: The invisible context
In most companies, domain knowledge resides in people’s minds, not in documents. Here’s the simplest test: Would a new colleague understand the meaning of the data without having to ask? If not, an AI lacks that same information and fills the gap with a guess rather than with knowledge.
Data access & governance: Authorization before access
Before an AI system automatically accesses data, it must be clarified who authorizes this use and how data protection is ensured. Addressing these issues retroactively, after a system is already in production, is the more time-consuming and riskier approach.
Architecture: The technical approach
An open architecture does not mean that all systems must be replaced. It means that data can be retrieved automatically via a defined interface, rather than requiring a manual export for every new request. Recurring manual exports are the clearest sign of an architecture that is too closed.
Use Cases: The reason behind everything else
The use case determines which of the previous four fields are actually relevant. A use case involving a high, easily measurable amount of manual effort and manageable data requirements is a better first step than a technically elegant but data-intensive project.
Which of the five areas is your biggest gap?
We’ll figure that out together in a free conversation.
How do you evaluate these five areas in your company?
The following five steps can be completed in one to two days without the need for an external tool. They do not provide a perfect picture, but they do offer a reliable initial assessment.
Use Cases: List all recurring processes that involve a high degree of manual effort and select the one with the best balance between effort and data availability.
Data quality: For the 2 to 3 most important metrics of the planned use case, check on a random basis whether two systems return the same value. If the values differ, the cause is often a difference in calculation logic, not a technical error.
Domain Knowledge: Ask a colleague from another department to explain a key metric. If they have difficulty explaining it, the knowledge is not sufficiently documented.
Governance: In a brief discussion with the business department, clarify who currently decides which data may be used for what purposes, and whether this decision is documented in writing.
Architecture: Trace back to the most recent report created and check whether it required a manual, separate export instead of an automated retrieval.
Why BI2run
BI2run uses this five-field model in every project as a common language between the business unit, IT, and management. It replaces vague assessments with a transparent structure that can be explained in a conversation, not just in an internal report. And if the review shows that a use case isn’t ready to go yet, we’ll say so – rather than starting a project that will later stall.
- 70–80% of project time typically goes into data preparation, not the model.
- 70–85% of AI projects miss their goals according to industry reports, mostly because of the data foundation.
- Poor data quality costs a company $12.9 million per year on average (Gartner).
- 5 areas are enough to assess AI readiness in a structured way.
- Every area that remains unclear marks a work item before the AI launch, not a reason to abort.
Glossary
Frequently Asked Questions
Do all five fields have to be green before an AI project can begin?
No. It is sufficient if the fields relevant to the specific use case are functional. Other areas for improvement can be addressed in parallel or later.
Which field is most often underestimated?
Domain knowledge and semantics. Domain knowledge is often stored in people’s minds, not in documents, and is therefore overlooked, even though it is just as important as technical data quality.
Is checking the five fields once sufficient?
For the first use case, yes. For each new use case, it’s worth taking another quick look, since requirements and data sources may change.
In practice, how long does it take to examine the five fields?
A preliminary, rough assessment can be completed in one to two days. A complete, documented assessment takes several weeks, depending on the system environment.
What if multiple fields are red at the same time?
In that case, it’s better to prioritize rather than try to solve everything at once. Generally speaking, not every red box hinders the planned use case to the same extent.
