AI Data Readiness: How to Assess and Prepare Company Data Before an AI Project

Many companies start an AI project by asking which model to use. A few weeks later, the team realises the real problem is not the model. Procedure documents are scattered across personal folders in conflicting versions. Customer data in the CRM does not match the finance system. Important transaction records still sit in a spreadsheet that only one person understands. Even the most advanced AI will give wrong answers when it is fed the wrong material.

Data readiness is the factor that most often decides whether an AI project makes it into daily operations or stops as an impressive demo. This article explains what data readiness means in practice, how to assess it quickly, the preparation work with the most impact, and how to start an AI project without waiting for all of the company's data to be perfect.

Summary

• Data readiness means the relevant data is available, accurate enough, securely accessible, and clearly owned
• Data needs differ by type of AI project: document based assistants, process automation, analytics, or AI agents
• Assess data readiness per use case, not for the whole company at once
• The most impactful work is usually settling sources of truth, cleaning out outdated documents, and defining access rights
• Do not wait for perfect data: choose one use case with good enough data, then improve the data as you go

Why data determines AI results

Modern AI models are very good at processing language and patterns, but they know nothing about your company except what you give them. When an AI assistant answers an employee's question about leave policy, it can only be as good as the policy documents available. When AI prepares a draft quotation, it depends on correct price lists and customer data. When a model predicts demand, it learns from the recorded transaction history.

The problem is that data errors in AI systems are often invisible. A report with wrong numbers may be questioned immediately by a finance manager. An AI answer delivered in fluent, confident language is easier to accept at face value, even when it comes from an outdated document. That is why data quality becomes more important, not less, in the age of AI.

Data needs differ by project type

Before assessing readiness, understand what data the planned type of project needs.

Document based assistants

An internal chatbot or customer service assistant that answers from company documents, usually with a RAG approach, needs documents that are up to date, not contradictory, and have a clear status. Old policies that have been replaced must be archived or marked, not left side by side with the new version.

Process automation and AI agents

Projects such as an AI agent that prepares orders or processes invoices need structured transaction data, consistent master data, and API access to systems such as the ERP and CRM. Business rules, such as approval limits or payment terms, must also be clearly documented, because AI cannot guess rules that exist only in senior employees' heads.

Analytics and prediction

Predictive models, such as demand forecasting or customer risk scoring, need a long enough data history, recorded consistently over time, that can be linked across systems. Definitions that changed midway, such as a change last year in how active customers are counted, need to be documented so they do not confuse the model.

Six dimensions of data readiness

Data readiness can be assessed with six simple questions. Answer them for each planned use case, not for the whole company.

1. 01Availability: does the needed data actually exist in digital form, or is it still on paper, in personal email, and in employees' memories?
2. 02Accuracy: how often is the data wrong, duplicated, or incomplete, and is there a process to correct it?
3. 03Consistency: is the same data recorded the same way across systems, with uniform codes and definitions?
4. 04Timeliness: how quickly is data updated, and can old versions be told apart from the latest?
5. 05Accessibility: can other systems retrieve the data through a reliable API or export, not just view it on screen?
6. 06Governance: is it clear who owns the data, who may access it, and which rules apply to personal or confidential data?

The assessment does not have to be complicated. A one or two hour discussion with process owners and the IT team, while opening real data samples, is often enough to know whether a use case is ready or needs preparation first.

The preparation work with the most impact

Define sources of truth

For each important type of data, decide which single system is the official reference. Customer data may come from the CRM, prices from the ERP, and employee data from the HRIS. AI and other integrations should pull data from that source, not from a copy in a spreadsheet. This principle is covered in more detail in our article on ERP master data.

Clean up documents and give them a status

For document based projects, the most valuable step is often not technical at all: gathering the relevant documents, removing duplicates, archiving old versions, and giving every document an owner and a review date. A well organised document management system makes this work far easier to sustain.

Digitise data that is still manual

If important data is still on paper or in scattered spreadsheets, consider moving it into a structured system. Our articles on moving from Excel to an integrated business system and automating data entry from documents explain how to do it step by step.

Design access rights from the start

AI connected to company data must respect the same access rights as the original system. An employee who may not see colleagues' salaries in the HRIS must not get them through an AI assistant either. Map access rights before connecting data, following the principles of designing user access rights. For customer and employee personal data, make sure processing complies with applicable personal data protection rules and check the latest guidance from the relevant authority.

Prepare secure access paths

Data that can only be viewed on an application screen is hard for AI to use. Prepare reliable APIs or scheduled exports, and consider standards such as the Model Context Protocol if several AI applications will use the same data sources. For analytics, bringing data together in a data warehouse is often a useful foundation for many needs at once.

Choosing the first use case

A common mistake is choosing the most commercially exciting use case without looking at the state of its data. A safer approach is to map candidate use cases on two axes: how much value they bring to the business, and how ready their data is. Use cases with high value and reasonably ready data are the best candidates for a first project.

• High value, data ready: do this first as a pilot project
• High value, data not ready: make it a data improvement project with clear targets, then run AI afterwards
• Low value, data ready: can serve as team practice, but should not be the yardstick for AI success
• Low value, data not ready: postpone

Improving data as you go

No company has perfect data, and waiting for perfection means never starting. AI projects can actually become a tool for improving data. A document based assistant will quickly reveal which documents contradict each other, because users will report strange answers. A predictive model will expose periods of data that make no sense. What matters is giving users a way to report problems and assigning who fixes them at the original source.

Track simple data quality metrics too, such as the number of duplicate customer records, the percentage of items without a category, or the number of documents past their review date. These metrics make data improvement visible and accountable, rather than a side task that keeps being postponed.

Signs your data is not ready

Some everyday symptoms can be an early hint that company data needs tidying before AI uses it. If several of these feel familiar, treat them as a to do list, not a reason to cancel your AI plans.

• Meetings are often spent arguing over which figures are right, because each department brings reports from a different source
• New employees have to ask senior colleagues to find the latest version of a procedure
• The same customer appears several times with different spellings of the name or different phone numbers
• Monthly reports are assembled by copying and pasting from several systems into one spreadsheet
• Nobody can say for certain who has access to salary data or customer data

Each of these symptoms carries a cost even without AI: wasted time, delayed decisions, and the risk of mistakes. That is why tidying data pays off twice. It clears the way for AI while also improving the operations already running, as discussed in our article on the hidden cost of data silos.

The role of AI governance

Data readiness cannot be separated from an AI usage policy and AI governance. That policy defines which data may be entered into third party AI services, how AI output is reviewed, and who is responsible when mistakes happen. Without it, employees tend to use AI tools on their own by copying company data into services the company does not control.

A short checklist before you start

• The use case and its success measures are clearly written down
• The required data has been identified along with its source systems
• Real data samples have been reviewed with the process owner
• Outdated documents or data have been separated from current versions
• Access rights and the handling of personal data have been mapped
• Technical access paths, such as APIs or exports, are available or planned
• There is an owner responsible for fixing data when problems are found

Conclusion

Successful AI projects rarely start with the model; they start with data that is well understood. Assess data readiness per use case, do the preparation work with the most impact, and choose a first project whose data is good enough to deliver real results. That way AI produces not just an impressive demo but a system that is trusted and used every day, while pushing the company to tidy up data that has long been neglected.