LLMs have already reshaped the way we consume data from the internet. This revolution reached us as users long before it reached organizations, and even though organizations understand the potential of LLMs to empower them internally and externally, they still struggle to translate that potential into concrete use cases, technical requirements, and implementation plans. One thing is sure: the train has already left. LLMs will redefine the way organizations interact with their data, because we all already understand what we could gain from them. It is not a question of “if,” but of “when” and “how.”

At Shift Happens 2025, I presented an analysis based on a survey we conducted with Georgian, HFN and SNC of more than 300 technical executives (C-level, VPs & directors of Data, R&D, Engineering, etc.) from global enterprise and growth-stage organizations. We set out to answer a simple question: what is the current stage of AI adoption in organizations when it comes to interacting with their data? The survey was complemented with dozens of in-depth interviews with industry leaders.

The adoption gap is clear. According to the survey, 78% of organizations are investing in AI for data access, analysis and insights, but only 40% manage to get these capabilities into production. When we asked what motivates these investments, 62% said their main goal was to stay competitive, while only 38% pointed to internal efficiency.

Read the full report

What worked for organizations that reached production?

In order to assist more organizations cross the chasm and implement AI over their organizational data, we dug into the data to explore what worked for those companies that successfully reached the production phase.

Here are the four patterns we identified that consistently separated organizations that did reach production:

1. Data, Data, Data

In every interview, data leaders repeated the same cliché: “garbage in, garbage out.” For them, data quality, a clear understanding of what data is stored where, and the ability to keep it accurate and up-to-date over time were the foundation for any successful production deployment.

90% of organizations that reached production said data quality strongly impacts outcomes; 47% built new processes and tools to ensure data accuracy, and 59% added orchestration to keep data reliable over time. Beyond hygiene, they built a living business-context metadata layer in natural language that explains what each asset means for the organization. This context layer became a single source of truth for analysts, internal tools and LLMs alike.

Download the full report

2. Build > Buy

About 80% of organizations that reached production built internal tools for core pieces of the data stack – 55% built primarily in-house; 21% wrapped narrow products with internal tools; and only 24% bought end-to-end products.
Leaders described this as a feature, not a bug. Existing products were often too immature for their needs or did not justify long procurement and rollout cycles, while owning the stack gave them control, a better fit and the option to move faster.

The old build vs. buy dilemma is relevant today more than ever for organizations: as the infrastructure space keeps shifting, it is hard to bet on the “right” piece of the stack. Moreover, AI makes it easier than ever for technical teams to build their own solutions with the existing AI code assistants. On the other hand, we still believe that infra products that are siloed and have a hard-to-build technological moat will always be easier to buy and plug into any tech stack.

3. Talent is The Key for Successful Implementation

Over 65% of organizations that reached production credited their human resource as the decisive factor. Every data leader we interviewed chose a small group of trusted engineers and data experts who deeply understood the organization’s systems, data and business, and gave them ownership of AI-for-data initiatives. These teams became both the technical pathfinders and the internal champions of the change.

4. Ongoing Learning and Improvement Process

Organizations that reached production started by asking where in the organization it made the most sense to begin: which users would get the highest value in the first phase, which users were detail-oriented enough to provide precise feedback that would improve the data behind the answers, and which users, if empowered with these capabilities, would amplify impact across the organization.

These questions guided their choice of initial data sources, which they then connected to LLMs. They released minimal versions into production, collected data on which information users actually cared about, and used that signal to see which datasets needed refinement to become AI-ready. According to the survey, 60% of organizations that reached production added LLM monitoring, vs 48% of those that did not.

Watch the full report

Insights and Guidelines for Organizations and Founders

1. Data is the Most Critical Asset in the AI Era

Organizations – ask yourselves: If you have not started thinking about it yet – why not, and what is holding you back? Is your data truly AI-ready, telling a story that both data teams and LLMs can understand? Is there a business unit that needs this more than any other and should be your starting point? And looking ahead, what do you want your data strategy to look like in this new world?

Founders – we are still in the early stages of AI adoption, and the stack around it is only beginning to evolve. The key question is what opportunities will appear as AI meets data at scale: what happens when organizations truly democratize access to data, when BI and visualization tools must serve many more users and when infrastructure needs to handle an exponential rise in queries.

Imagine a world where not only tables and queries, but also transformations, automation and migration are described in natural language, and think about what new data and AI stack will be built on top of that. Your challenge as entrepreneurs is twofold: develop a strong opinion on how this world will look in 4-5 years, while also delivering clear, practical value to organizations in the next 1-2 years as they figure out what this shift means for them.

2. Organizations Prefer Building Internal Infrastructure Rather than Buy Off-The-Shelf Products

This insight puts the spotlight back on organizations: where do you stand in this dilemma? Do you actually have the talent and time to build? Would you hire to support that choice or wait for off-the-shelf products to mature? Will the same answer still hold in 1-2 years or across other layers of your stack, such as DevOps?

For founders, the toughest competitor is often the internal build. Technical buyers will ask whether your product truly justifies buying something new: does it solve a problem that is painful enough, and is your solution genuinely hard to build in-house? They need to believe that nothing in their existing stack will deliver this soon, that it is worth allocating budget and going through procurement, and that you will support their requests and customization over time. This is where AI can work in your favor: can you use current AI capabilities to reduce their concerns by making your product adapt to their environment with minimal integration and configuration?

3. Success in the AI Era Requires Top Talent with Data Skills

Data is the fuel of this shift, and many organizations now see their proprietary data as the moat that protects them from competition. Any startup that wants to build AI-based products must be able to answer a basic question: what is your data strategy, and how is both the content and structure of your data unique to you. The message to both organizations and founders is simple: talent is the foundation. If you already have people with strong data skills inside the company or in your close network, do everything you can to keep and empower them; and if you don’t have any, hire them tomorrow.

Authors:

Or Git
Investor
[email protected]