GenAI has already made an extraordinary impact on enterprise productivity. Marc Benioff has stated Salesforce will keep its software engineering headcount flat due to a 30% increase in productivity thanks to AI. Users leveraging Microsoft Co-pilot create or edit 10% more documents.
But this impact has been evenly distributed. Powerful models are a simple API call away and available to all (as Meta and OpenAI ads make sure to remind us).
The real disruption lies with “data + AI.” In other words, when organizations combine their first-party data with LLMs to unlock unique insights, automate processes, or accelerate specialized workflows.
No one knows exactly when this tidal wave will hit, but based on our conversations with dozens of teams actively working on data + AI applications, it is clear the time is nigh.
Why? Well, this follows a pattern we’ve seen before. Multiple times. Every major technology shift sees initial adoption that is magnified once it reaches enterprise level reliability. We saw this with software and application observability, data and Data Observability, and soon data + AI and data + AI observability.
In this post, we’ll highlight the progress of Enterprise Data + AI initiatives as well as the path many teams are taking to cross the tipping point.
Past is prologue
Data + AI will deliver exponentially more unique value, but it is also exponentially harder.
Most organizations don’t have $500 billion to spare for science fiction-themed initiatives. Enterprise applications need to be economically feasible and reliable.
Looking at past technology advances–namely cloud computing and big data–we can see it typically happens in that order. Infrastructure and capacity breakthroughs create demand and increased reliability levels are required to sustain it.
Before the internet was powering the world’s most impactful SaaS applications with increasingly critical tasks from banking to real-time navigation, it was mainly the domain of cat pictures, AOL chatrooms, and email chain letters. That change only happened once we reached the fabled “5 9s of reliability.” S3, Datadog, and site reliability engineering practices changed the world.
Prior to data powering valuable data products like machine learning models and real-time marketing applications, data warehouses were mainly used to create charts in binders that sat off to the side of board meetings. Snowflake and Databricks changed the economics and capacity of data storage/processing and data observability brought reliability to the modern data stack.
This pattern is repeating with AI. 2023 was the year of GPUs. 2024 was the year of foundational models. 2025 has already seen dramatic increases in capacity with DeepSeek and the initial ripple of agentic applications will become a tidal wave.
Our bet is 2026 will be the year when data + AI changes the world…and, if history is any indicator, it will be no coincidence this revolution will be immediately preceded by advances in observability.
Where data + AI teams are today
Data + AI teams are further along than they were last year. Based on our conversations:
- 40% are in the production stage (30% just got there)
- 40% are in the semi or pre-production stage
- 20% are in the experimentation stage
While you can see the critical mass building, all of them are facing challenges as they attempt to reach full scale. The most common themes:
Data readiness — You can’t have good AI with bad data. On the structured data side of the house, teams are racing to achieve “AI-ready data.” In other words, to create a central source of truth and reduce their data + AI downtime.
On the unstructured side, teams are struggling with conflicting sources and outdated information. One team in particular cited a “fear of an unmanageable knowledge base” as the main impediment to scale.
System sprawl — Currently, there is not what we would call an industry standard architecture, although hints are emerging. The data + AI stack is actually four separate stacks coming together: structured data, unstructured data, AI and oftentimes the SaaS stack.
Each stack on its own is difficult to govern and maintain high reliability levels. Piecing them together is complexity squared. Almost all of the data teams we have talked to are trying to consolidate the chaos where they can, for example, by leveraging large modern data cloud platforms for many of the core components rather than purpose-built vector databases.
Feedback loops — One of the most common challenges inherent in data + AI applications is that evaluating the output is often subjective. Common approaches include:
- Letting human annotators score outputs
- Tracking user behavior (such as thumbs up/down or accepting a suggestion) as an indirect measure of quality
- Using models (LLMs, SLMs and others) to score outputs on various criteria
- Comparing outputs with some known ground truth
All approaches have challenges, and creating correlations between system changes and output results is near impossible.
Cost & latency — The progress of model capacity and cost is breathtaking. During a recent presentation, Thomas Tunguz, a leading venture capitalist in the AI space, shared this graph showing how smaller (less expensive model) performance is reaching similar performance levels as larger models.
But we aren’t quite at commodity infrastructure prices just yet. Most teams we spoke with had concerns around the financial impact of AI adoption. If there was any monitoring taking place, it was more often than not on tokens and cost rather than outcome reliability.
The next frontier: Data + AI observability
Data + AI is an evolving space with unique challenges, but the principles of building reliable technology systems have remained consistent for decades.
One of those core principles is this: you cannot just sporadically check the product at the end of the assembly line or even at certain points throughout the assembly line. Instead, you need full visibility into the assembly line itself. For complex systems, it is the only way to identify issues early and trace them back to the root cause.
But you need to observe the whole system. End-to-end. It doesn’t work any other way.
To achieve data + AI reliability, teams will not be successful by observing models in a vacuum. For data + AI observability, that means integrations across the core system components. In other words, the four ways data + AI products break: in the data, system, code, or model.
Detecting, triaging and resolving issues will require visibility into structured/unstructured data, orchestration/agent systems, prompts, contexts and model responses. (Stay tuned for an upcoming deep dive on exactly what this means and how each component beaks).
Data + AI are no longer two separate technologies; they are a single system. By next year, let’s hope we’re treating it like one.
Change happens slowly, then all at once
We’re at that precipice with data + AI.
No organization will be surprised by the what or the how. Every member of the boardroom, the C-suite, and the breakroom has seen how past platform shifts have created Blockbusters and Netflixes.
The surprise will be in the when and the where. Every organization is racing, but they don’t know when to pivot when to break into a sprint, or even where to run.
Standing still is not an option, but no one wants to use rapidly evolving infrastructure to build bespoke AI applications that will quickly become commoditized. No one wants their picture accompanying the next AI hallucination headline.
It’s clear achieving reliability at scale will be the tipping point that crowns new industry titans. Our recommendation is that as the data + AI space matures, make sure you are prepared to pivot.
Because if the past has shown us anything, it’s that the organizations with the right foundations for building reliable systems with high levels of data readiness will be the ones crossing the finish line.
