I recently joined Fern Halper from TDWI and Josh Caplan from Microsoft for a discussion on why enterprises aren’t ready for AI, a topic TDWI also recently covered in a report. The conversation reinforced something I see every day: the AI readiness gap is really a data readiness gap. 

I once asked a CDO at a major networking company, “Where do you rank on data maturity on a scale of 1 to 10?” He said, “We’re probably a 3.5.” I would have expected 8 or 9 from a company that provides critical infrastructure for data and AI.

Data maturity is a fundamental prerequisite for AI maturity. Without data maturity, AI maturity is impossible.

Look at the root causes

According to both TDWI and my own observations, the following recurring issues hold companies back:

• Architecture is too complex and brittle
• Data is siloed across the org, with everyone operating off different versions of the truth
• Delays and project time overruns keep piling up, sprint by sprint

The engineering demands require highly specialized teams to get anything done.

• Infrastructure reliability is a persistent problem

Many data teams struggle with bespoke pipelines, inconsistent schemas, and governance issues that kill trust. The higher you go in an organization, the more honest people are about the actual state of their data program, like the CDO I spoke with.

Data matters more than algorithms or models

The true value of any AI project comes from combining it with your unique enterprise data. You can build the perfect AI application for healthcare clinical decision support or CPG consumer insights with clever prompt engineering, instructions, examples, and guardrails, along with the latest foundation models and frameworks. But, without specific, focused, quality data from your enterprise to augment it, that AI application fails.

Take healthcare clinical decision support. You need not only general knowledge, but also data and analysis of patient outcomes, treatment effectiveness, and medical error rates. You need patterns in medication recommendations and readmission risks. That requires combining internal data from systems like Epic or Cerner with external sources like ClinicalTrials.gov and NIH.

The maintenance trap that nobody talks about

TDWI found that 67% of organizations spend over 25% of engineering time on pipeline maintenance. Only 19% prioritize modernizing those pipelines.

Here’s what those numbers miss: as enterprises adopt an expanding spectrum of applications, tools, and databases, the backlog never ends. For each new source, your team must design, build, test, deploy, document, automate, maintain, and refactor. Then rinse and repeat for the next source.

The data integration process — extraction, loading, and transformation — poses serious challenges at every turn, including:

• Source interaction: sync strategy, data filtering, delete capture, schema definition, state management
• Pipeline platform processing: orchestration, connector resyncs, infrastructure for custom code, scaling, log management
• Destination delivery: retries, idempotency, schema inference, failure recovery, reliability, security

Moving data from one tool or platform to another is deceptively complex.

Building a healthy data foundation

The key implication of TDWI’s findings is that most organizations lack a solid data foundation for analytics and AI. To build a healthy data foundation, organizations need the ability to reliably and automatically integrate data, as well as an ecosystem of tools for storing, governing, and otherwise handling all forms of data.

How to choose a good data integration platform

For data integration, your team will need the following key capabilities. With an automation-first infrastructure, enterprises can reduce long-term costs, enhance agility, and ensure their AI initiatives are fueled by reliable, high-quality data. In short, look for the following capabilities in a data integration platform:

• Automated data integration: Data integration is deceptively complex, involving a wide range of engineering and design considerations. Look for a turnkey solution that requires minimal engineering time so that your team can focus on higher-value work more directly related to optimizing internal operations or improving customer experiences.
• Enterprise-ready performance, reliability, and scalability: The stringency of your needs will grow with time. Make sure your team is set up for future success with capabilities that can meet the needs of not only simple use cases but also large-scale, complex ones.
• Interoperability and compatibility: Your data integration solution should be able to work with as many of your data sources as possible, as well as support any cloud environment and destination your team may need.
• Security, regulatory compliance, and governance: Ensure you can know, protect, and control access to all data handled by your data integration solution. Make sure your data integration platform can support whatever residency and deployment options your enterprise might need.

Fully managed data integration platforms like Fivetran offer these capabilities out of the box, enabling your teams to work on projects that directly produce value for your organization.

How to choose the right cloud data platform

As platforms have become more mature and bundled with a broader range of capabilities, I’ve come to prefer centralized platforms over piecing together architectures from best-of-breed components. Centralized platforms offering governance and AI capabilities are not only more convenient but also more trustworthy.

Here are the ideal characteristics of a cloud data platform:

• AI-ready data platform architecture: Compute and storage can be combined on a single platform, or decoupled using data lakes and open table formats like Delta and Iceberg, which provide more flexibility based on your specific compute needs.
• Support for all data types: You will need to be able to store and process structured, semi-structured, and unstructured data. Your AI projects may eventually encompass text, images, audio, and video.
• Technical catalog: You need complete oversight, management, and lineage tracking over your data assets.
• Access to leading LLMs: You want to be able to plug in the latest Anthropic Claude, OpenAI GPT, Meta Llama, DeepSeek, and Gemini models without setting them up with individual accounts and API keys. Different use cases require different models. If possible, you also want that functionality to be natively accessible via SQL, using functions like COMPLETE, CLASSIFICATION, EXTRACT, etc.
• Agent framework: To get started with agentic AI, you will want the ability to declaratively configure agents with low or no code or build agents programmatically through an API.
• UI and data app frameworks: Integrations with app frameworks ensure you don’t have to build applications from scratch. Options include Streamlit, Flask, Gradio, Node.js, and others.
• Security and governance handled by the platform vendor: Your team shouldn’t worry about security and governance. Platforms should apply governance policies consistently across all datasets and AI workloads.

Managing platform risk without stalling progress

The main risk posed by a centralized platform is vendor lock-in. There are several ways to manage this risk:

1. Try different combinations of platforms and cloud environments: Try taking three different platforms over three two-week sprints each and getting a good sense of what’s best for you, at least for now. The platforms are quick and easy to configure and get up and running. From a risk management standpoint, prototyping across multiple platforms is a cost-effective approach, and in many cases, you can do it with free services during each two-week sprint.
2. Accept that you might use multiple clouds: Although a unified data architecture is the aspiration, many organizations are, in practice, multi-cloud due to the specific advantages and capabilities each cloud service provider offers.
3. In the worst-case scenario, migrations are possible: If you ever need to move or consolidate clouds, migrations are also possible

One caveat: if you’re stuck on-premises, this becomes difficult or nearly impossible. Get to the cloud first. Doing this on-prem with a pure build is a non-starter.

Start simple with AI

Without making a start on AI, you risk letting any competitive advantage slip away to faster-moving competitors. The good news is that a solid data foundation can be readily assembled out of interoperable tools and platforms, and micro Gen AI applications and agentic workflows don’t require massive datasets or months (or years) of transformation refactoring.

Start small, but solve a meaningful business problem. Find a solid use case that is limited in scope and solvable with a limited number of data sources. Make sure you understand the workflow, business process, potential impact, and datasets involved. Simple examples include:

• Internal knowledge assistant: Generative AI is unparalleled for its ability to retrieve and synthesize information. An automated, interactive help desk means members of your team don’t need to annoy each other with simple questions. Start with small examples like HR and company policy; expand into product and technical assistance. Extract from sources such as
  
  • HR apps
  • Company knowledge base
  • Product documentation
• Customer support reference bot: Similarly, your customers could benefit from an automated point of contact for fielding relatively simple questions. You will need data from the following kinds of apps:
  
  • Ticketing
  • CRM
• Sales enablement copilot: Your sales team needs all manner of collateral on hand to answer more technical or otherwise complicated questions from prospects. Use data from:
  
  • Sales and marketing collateral
  • Product documentation
  • Battlecards

Keep your eye on the ultimate goal. Unlike more traditional machine learning, generative AI is broadly accessible because it uses a natural language interface. In what capacity can people in your organization benefit from radically improved information retrieval and synthesis?

Then, set up a workflow to automatically integrate this data. Your system should be set up and run with minimal intervention from an engineering or data team.

The real milestone

Gen AI and LLMs have been accessible for three years now. If you aren’t taking advantage of AI daily, you’re leaving money on the table and lighting it on fire. Individual productivity increases are a given, but we now also have the potential for specific, measurable business value and impact, not just technical value, such as data engineering time saved.

Healthcare organizations can deliver improved patient outcomes and higher quality guest experiences. Ranches can change livestock health trajectories. Telecom companies can improve churn prevention for customers and offer them more value. Financial services can enable better product recommendations. Universities can retain students at higher rates. The list of very specific, focused applications that deliver value on quality, trusted, usable data goes on.

You should ask yourself:

• How do I design and engineer innovative new products?
• How do I more effectively market and sell my products or services?
• How do I deliver the best overall experience for prospects, customers, and partners?
• Where can I gain operational efficiency?

Most organizations can’t achieve these outcomes because they’re stuck in the infrastructure maintenance cycle that TDWI documented. Sixty-seven percent of engineering time is spent keeping pipelines running instead of building business-changing applications.

Break that cycle. Get automated data movement that works, like Fivetran, and start delivering high-quality, trusted, usable data for any data workload. You’ll understand why I recommended Fivetran 100+ times as a data consultant in my prior life. Let your teams focus on business value, not data infrastructure.

[CTA_MODULE]