Making customer data and knowledge sources suitable for AI costs, on average, more time and money than the AI implementation itself. The preparation, also known as data readiness or data onboarding, involves cleaning, structuring, unlocking, and validating information that AI needs to work reliably. Costs vary significantly per organization, depending on the state of your current systems, the number of data sources, and the desired AI application area. In this article, we answer the most frequently asked questions about what this preparation exactly entails, what it costs, and how to approach it smartly.
What types of customer data and knowledge sources does AI need?
AI for customer contact needs two main categories: customer data (who are your customers, what have they asked, what interactions have taken place) and knowledge sources (what does your organization know, which answers are correct, which procedures apply). Without both categories, an AI system cannot provide reliable, relevant answers.
Customer data includes, among other things, conversation history, email threads, chat conversations, CRM records, and complaint registrations. Knowledge sources are the internal documents, FAQ pages, manuals, product information, policies, and process descriptions that employees use daily to help customers.
For an AI in a contact center, the most valuable sources are:
- Structured customer data from CRM systems and ticketing software
- Historical call recordings and transcriptions
- Internal knowledge bases and manuals
- Frequently asked questions with corresponding answers
- Product catalogs, rate overviews, and contract terms
- Escalation and routing rules
The quality of these sources directly determines the quality of the AI. Outdated, contradictory, or poorly structured information leads to incorrect or unreliable AI answers, regardless of how advanced the model is.
What are the biggest cost drivers when preparing data for AI?
The biggest cost drivers when preparing customer data and knowledge sources for AI are cleaning existing data, unlocking fragmented systems, and manually validating and structuring knowledge documents. These activities are labor-intensive and are often underestimated in project planning.
Specifically, you can divide the costs into four categories:
- Data audit and inventory: Before you begin, you must know what you have. Which systems contain relevant customer data? Are those systems linked or separate? This costs time from both IT and operations.
- Data quality and cleaning: Duplicate records, outdated contact details, inconsistent categorization, and missing fields must be corrected. With older systems, this is often the heaviest task.
- Integration and unlocking: AI needs access to the right sources. If data is in multiple standalone systems, API links or data pipelines are required. This demands technical capacity and brings integration costs.
- Knowledge structuring: Documents that are readable for humans are not automatically suitable for AI. Texts must be split, labeled, classified, and sometimes rewritten to be consistent and unambiguous.
Avoid the pitfall of underestimating these costs. In practice, data preparation often accounts for 40 to 60 percent of the total lead time of an AI project.
How long does it take to make customer data and knowledge sources AI-ready?
The lead time to make customer data and knowledge sources AI-ready varies from a few weeks to several months, depending on the size of your data landscape, the state of your current systems, and the type of AI application you want to deploy. A targeted focus on one channel is faster than an organization-wide implementation.
A realistic indication per phase:
- Inventory and analysis: 2 to 4 weeks
- Data quality improvement: 4 to 10 weeks, depending on the current state
- Integrations and unlocking: 3 to 8 weeks per system
- Knowledge structuring and validation: 4 to 12 weeks, depending on the volume of documents
In organizations with highly fragmented customer contact, multiple suppliers, and outdated systems, the total preparation time increases more quickly. A phased approach, starting with the most valuable or most structured data sources, helps to achieve results sooner without having to wait for the entire process to complete.
What must be in order before AI can be meaningfully deployed?
Before AI can be meaningfully deployed in customer contact, three things must be in order: a reliable data foundation, an up-to-date and unambiguous knowledge base, and clear ownership of both the data and the AI outcomes. Without this foundation, AI produces answers that are unreliable, outdated, or contradictory.
Specifically, this means:
- One source of truth for customer information: If customer status, contract details, or interaction history are in multiple places and are not synchronized, the AI does not know which version is correct.
- Consistent knowledge documentation: Information must be unambiguous, up-to-date, and well-structured. Contradictory answers in different documents lead directly to incorrect AI output.
- Clear process definition: AI must know when it is allowed to handle a question independently and when it must transfer to an employee. Those boundaries must be established in advance.
- Privacy compliance: Customer data processed by AI must comply with the GDPR. Ensure you know what data you use, where it is located, and how it is secured.
A thorough business analysis helps you systematically map out these basic requirements before you invest in AI technology.
When is outsourcing data preparation smarter than doing it yourself?
Outsourcing data preparation is smarter than doing it yourself when your organization lacks the internal capacity, technical expertise, or time to tackle it thoroughly. Doing it yourself sounds cheaper, but incomplete or incorrect data preparation leads to an AI that does not work well and ultimately turns out to be more expensive.
Outsourcing is wise in the following situations:
- You have no internal data engineer or knowledge manager available
- Your systems are highly fragmented and integrations require specialist knowledge
- You want to achieve results quickly without a long internal process
- You want to combine data preparation with the AI implementation itself, so there is one point of contact for the entire process
Doing it yourself has advantages if your organization already has a well-equipped data team, clear processes, and modern systems. In that case, remote guidance or a hybrid approach is often more efficient than full outsourcing.
The most important criterion is not the short-term cost, but the quality of the end result. Poorly prepared data results in an AI that disappoints employees and customers, regardless of how good the AI model itself is.
How do you keep knowledge sources up to date after AI is deployed?
Keeping knowledge sources up to date after AI implementation requires a structural management process with clear ownership, periodic reviews, and a way to quickly signal outdated or incorrect information. AI is only as current as the knowledge base it runs on, and that knowledge base becomes outdated without active management.
A working management process consists of three elements:
Ownership and editorial process
Assign a responsible person per knowledge domain who translates changes in products, services, policies, or procedures into updates in the knowledge base. Without an explicit owner, updates will be neglected until the AI gives demonstrably wrong answers.
Feedback loop from practice
Employees and customers signal when AI answers are incorrect. Ensure those signals have a clear route to the knowledge base manager. Monitoring AI interactions helps to recognize patterns: which questions are regularly answered incorrectly, which topics are missing, which answers lead to escalation?
By continuously monitoring the data quality of your AI, you prevent the knowledge base from slowly diluting. Quarterly reviews per knowledge domain, combined with an accessible reporting process for employees, are an effective combination in practice.
How Pegamento helps you with data preparation for AI in customer contact
We guide organizations from start to finish in making customer data and knowledge sources AI-ready, so that your AI implementation immediately stands on a reliable foundation. No loose advice, but a total package from analysis to implementation and ongoing management, all under one roof.
What we do for you:
- Business analysis and data inventory: We map out your current data landscape, identify bottlenecks, and determine which sources have priority.
- Data quality improvement and integrations: We unlock fragmented systems, correct inconsistencies, and ensure AI has access to the correct, up-to-date information.
- Knowledge base structuring: We help you convert internal knowledge documents into an AI-suitable structure, without costly customization but with proven modules that fit your organization.
- AI implementation and management: From contact center AI and Agentic AI to process automation, we implement and manage the solution and continue to optimize after go-live.
- Privacy and security: We work in accordance with ISO 27001, ISO 9001, and ISO 26000, so that your customer data is processed safely and compliantly.
Do you want to know what your organization needs to deploy AI meaningfully? Contact us and we will look together at the best approach for your situation.
Frequently Asked Questions
How do I know if my organization is already ready to start with data preparation for AI?
A good starting point is an internal scan of your current data landscape: how many systems contain customer data, are those systems linked, and does a central knowledge base already exist? If you cannot answer those questions directly, a formal data inventory is the first step. Organizations that already have a CRM system, a structured knowledge base, and clear process owners generally have a shorter preparation time than organizations that still need to build those up.
What are the most common mistakes when preparing data for AI?
The most common mistake is starting data preparation too late, causing the AI project to be delayed or the implementation to take place on an unreliable data foundation. A second common mistake is underestimating the amount of outdated or contradictory information in existing knowledge documents. Finally, we often see that organizations do not assign an owner to the knowledge base, causing the quality to decline rapidly after go-live.
Can I already start with AI while data preparation is still ongoing?
That is possible through a phased approach: you start the AI implementation on the best available, most structured data sources, and then expand step by step as more sources are made AI-ready. This gives you results sooner and makes it possible to learn in practice what the AI needs. Keep in mind that the AI will have more limited coverage in this phase and that you must manage expectations internally.
How do I ensure that my employees maintain the AI knowledge base correctly and consistently?
The key is a simple, accessible editorial process with clear owners per knowledge domain. Make reporting outdated or incorrect information as easy as possible, for example via an internal form or a direct feedback button in the AI interface. Combine this with fixed quarterly reviews per domain and automatic monitoring of AI interactions, so that you proactively signal which knowledge items need attention instead of correcting reactively.
What is the difference between data preparation for generative AI and for traditional chatbots?
Traditional chatbots work based on fixed decision trees and keyword recognition, where data requirements are relatively limited and predictable. Generative AI, on the other hand, derives meaning from large amounts of unstructured text and is therefore more sensitive to contradictory, outdated, or poorly formulated information in the knowledge base. This means that knowledge structuring for generative AI must be more thorough and consistent, but also that the AI handles variations in customer questions more flexibly once the foundation is correct.
How much budget should I reserve for data preparation as a percentage of the total AI project?
As a rule of thumb, you reserve 40 to 60 percent of the total project budget for data preparation, including inventory, cleaning, integrations, and knowledge structuring. This percentage shifts to the lower end if you already have modern, well-linked systems and an up-to-date knowledge base. With fragmented or outdated data landscapes, the share of data preparation can rise to 70 percent of the total investment. Calculate this in advance based on a data audit to avoid surprises in the project budget.
Which GDPR points of attention are specifically relevant when preparing customer data for AI?
When deploying customer data for AI, you must, among other things, determine which legal basis you use for processing, whether the AI provider acts as a processor and whether a processing agreement is required, and whether personal data is used for training the model. Additionally, it is important to check whether customer data is stored outside the EU and which retention rules apply to call recordings and chat history. A data protection impact assessment (DPIA) is recommended before you actively integrate customer data into an AI system.
