Five essential practices for AI-ready data modernization

It’s no secret that a large proportion of AI projects stall at the proof of concept (POC) stage. Estimates of the AI abandonment rate vary, with some research indicating that almost 90% of AI pilots fail to make it into production. What’s more, the problem appears to be getting worse, not better: research by S&P Global found that the percentage of organizations shelving the majority of their AI initiatives before production jumped from 17% in 2024 to 42% in 2025.

The pilot purgatory trap emerges for a variety of reasons: customer and employee resistance, unreliable outputs, unclear ROI and a scattershot approach to use case selection. But one common factor is the lack of reliability in the underlying data and data foundation.  

In a recent Thoughtworks Perspectives issue, we explored how leaders can build confidence in their data through value alignment, strong governance and change management. To put these foundations into practice, data practitioners must evolve. Traditional practices provide an essential foundation of engineering discipline and data governance. They build and maintain the data pipelines, information and insights that are vital for turning raw data into effective actions on the ground. 

In an agentic world, fundamental data engineering practices are more important than ever. But new practices are needed, too. When AI agents are making decisions and taking independent actions, the data they use will have a major impact on whether they make the right calls.

Here are five modern data practices that help ensure high-quality, well-governed data is available, discoverable and consumable by agents and humans – so they can make better decisions and take more effective actions. 

1. Knowledge curation makes AI outputs more reliable

Large organizations operate through a complex network of relationships between vast numbers of entities including products, employees, customers, suppliers, business units and sites. Business analysts might have a reasonable knowledge of many of these relationships, allowing them to apply organizational context to data. But without this organizational knowledge, large language models (LLMs) take the data at face value, which can lead them to draw incorrect conclusions and generate unreliable outputs.

By taking an intentional approach to knowledge management, organizations can build a curated information environment for AI, allowing LLMs to understand and use data more effectively by capturing meaning and relationships in a consumable format. This might include creating ontologies of entities and their relationships, curating a semantic layer, building data dictionaries and providing data lineage and quality scores. All these assets help LLMs understand organizational knowledge and context, increasing the reliability of AI outputs.

2. Computational data governance makes it nearly impossible to do the wrong thing

All data use comes with some element of risk. Managing this risk at scale is crucial for enabling LLMs and agents to operate effectively. But it’s extremely hard to achieve by relying on slow, manual, centralized data management.

Computational governance is one of the core principles of the data mesh architectural paradigm. Instead of manual audits and approval gates, it automatically enforces data governance rules by baking compliance into the data ecosystem. 

With automated controls for data quality standards, access management, data sharing and privacy, it becomes very difficult, if not impossible, for AI or humans to do the wrong thing with data. When computational governance is built into AI tools and agents, the risk/reward calculation looks very different, allowing organizations to confidently move from small experiments to enterprise-wide production systems. 

3. AI-ready data products make data discoverable and usable for agents

Another pillar of data mesh is the concept of data as a product (DaaP). The idea is to treat data as a valuable asset and apply the principles of product thinking to how it’s managed and used. Data products transform an organization’s ability to generate value from data by addressing specific business needs and solving real-world problems.

Where in the past data products were designed for domain teams to find, share and reuse, they must now take into account a new AI audience. That means ensuring data products are discoverable and usable by both humans and agents, helping the entire organization operate more efficiently and effectively. 

4. Evaluations and observability provide transparency and auditability

Without trust, the potential risk of deploying agentic systems into production will always outweigh the potential rewards. The probabilistic nature of generative and agentic AI can make it hard to trust that outputs will be reliable, accurate and safe. Ask an LLM the same question five times, and you might get five different answers. 

As AI systems become more autonomous, that unpredictability can carry greater consequences. Organizations need confidence not only in the outputs agents produce, but also in the data they access, the decisions they make and the actions they take.

For AI to become a trusted part of the organization, you must know that agents are only accessing the data and systems they’re supposed to and that they’re making decisions and taking actions within clearly defined constraints. 

Key to this is combining evaluations (or evals) and observability practices to continuously assess and monitor non-deterministic workloads and track what agents see, decide and do. These quality checks paired with increased visibility provide the reliability, transparency and auditability to validate the trust organizations place in agentic systems.

Visibility into agent behavior must also extend to the people who deployed and prompted the agent in the first place and who must retain accountability for the outcomes it delivers. 

With intentional validations, full visibility, auditability and accountability, organizations can reinforce trust by confirming that agents are doing what they’re supposed to do and aren’t straying beyond the boundaries they’ve been set. 

5. Model Context Protocol integration helps scale agentic ecosystems

Connecting AI to all the structured and unstructured data sources it needs could be an integration nightmare. To overcome this, many teams are now using the Model Context Protocol (MCP) to act as an API wrapper that standardizes how agents interact with various parts of the data ecosystem. 

Integrating agents with an MCP server can be easier than building integrations across multiple APIs, enabling faster development and broader access to context. However, it’s important to note that MCP shouldn’t necessarily be the default integration option. Alternative approaches may be more appropriate, depending on the specific use case.

Modernize data engineering for the agentic future

Together, these five practices bolster data teams, enabling organizations to get agents and other AI-powered applications into production with speed and confidence. 

Applied correctly, these practices also create a strong but flexible foundation for future innovation. Organizations gain the freedom to create agentic systems and AI-based products; ultimately getting more use cases and capabilities out of POC and into production.