5.2.1. Data Readiness and the Cost of Data Silos

💡 First Principle: Data readiness asks four questions: does the data exist, is it good enough, can the AI system reach it, and are we allowed to use it for this purpose? Data silos fail the third question and often the second, because data held separately by each department is defined differently, maintained differently, and joined by no one.

Dimensions of data readiness:

• Availability: the data needed for the use case exists and is captured consistently.
• Quality: accuracy, completeness, consistency, timeliness, representativeness (2.1.3).
• Accessibility: the data can be reached by authorized systems and people with reasonable effort, through governed interfaces rather than one-off extracts.
• Discoverability: people know the data exists and what it means (catalogs, documentation, business definitions).
• Permissibility: the organization has the legal and contractual right to use it for AI and for this purpose.

The effects of data silos on AI initiatives:

• Incomplete context: a customer-service AI that sees support tickets but not orders or billing history gives partial answers.
• Inconsistent definitions: "active customer" means different things in sales and finance, so models trained on one conflict with reports from the other.
• Duplicated effort: each project builds its own extracts and cleaning pipelines.
• Slower delivery: negotiating access across departments for each initiative adds months.
• Retrieval gaps: RAG systems can only ground answers in content they can reach, so siloed knowledge produces incomplete or wrong answers.
• Governance blind spots: copies of data proliferate outside controls.

A readiness review for a specific initiative should list the data required, where it lives, who owns it, its quality, how it will be accessed, and whether its use is permitted. Gaps become work items with owners before build begins.

⚠️ Exam Trap: When a scenario describes an AI initiative failing because it cannot combine data across departments, or producing inconsistent results because definitions differ, the answer addresses data silos and accessibility (ownership, sharing agreements, common definitions, governed access), not a more powerful model.

Reflection Question: A telecom company wants an AI agent to resolve billing complaints, but billing, network outage, and customer-contact data are owned by three departments with no shared customer identifier. What data readiness gaps exist, and what must be resolved before the agent can work reliably?