the final quarter of 2025, it’s time to step back and examine the trends that will shape data and AI in 2026.

While the headlines might focus on the latest model releases and benchmark wars, they’re far from the most transformative developments on the ground. The real change is playing out in the trenches — where data scientists, data + AI engineers, and AI/ML teams are activating these complex systems and technologies for production. And unsurprisingly, the push toward production AI—and its subsequent headwinds in —are steering the ship.

Here are the ten trends defining this evolution, and what they mean heading into the final quarter of 2025.

1. “Data + AI leaders” are on the rise

If you’ve been on LinkedIn at all recently, you might have noticed a suspicious rise in the number of data + AI titles in your newsfeed—even amongst your own team members.

No, there wasn’t a restructuring you didn’t know about.

While this is largely a voluntary change among those traditionally categorized as data *or *AI/ML professionals, this shift in titles reflects a reality on the ground that Monte Carlo has been discussing for almost a year now—data and AI are no longer two separate disciplines.

From the resources and skills they require to the problems they solve, data and AI are two sides of a coin. And that reality is having a demonstrable impact on the way both teams and technologies have been evolving in 2025 (as you’ll soon see).

2. Conversational BI is hot—but it needs a temperature check

Data democratization has been trending in one form or another for nearly a decade now, and Conversational BI is the latest chapter in that story.

The difference between conversational BI and every other BI tool is the speed and elegance with which it promises to deliver on that utopian vision—even the most non-technical domain users.

The premise is simple: if you can ask for it, you can access it. It’s a win-win for owners and users alike…in theory. The challenge (as with all democratization efforts) isn’t the tool itself—it’s the reliability of the thing you’re democratizing.

The only thing worse than bad insights is bad insights delivered quickly. Connect a chat interface to an ungoverned database, and you won’t just accelerate access—you’ll accelerate the consequences.

3. Context engineering is becoming a core discipline

Input costs for AI models are roughly 300-400x larger than the outputs. If your context data is shackled with problems like incomplete metadata, unstripped HTML, or empty vector arrays, your team is going to face massive cost overruns while processing at scale. What’s more, confused or incomplete context is also a major AI reliability issue, with ambiguous product names and poor chunking confusing retrievers while small changes to prompts or models can lead to dramatically different outputs.

Which makes it no surprise that context engineering has become the buzziest buzz word for data + AI teams in mid-year 2025. Context engineering is the systematic process of preparing, optimizing, and maintaining context data for AI models. Teams that master upstream context monitoring—ensuring a reliable corpus and embeddings before they hit expensive processing jobs—will see much better outcomes from their AI models. But it won’t work in a silo.

The reality is that visibility into the context data alone can’t address AI quality—and neither can AI observability solutions like evaluations. Teams need a comprehensive approach that provides visibility into the entire system in production—from the context data to the model and its outputs. An socio-technical approach that combines** data + AI **together is the only path to reliable AI at scale.

4. The AI enthusiasm gap widens

The latest MIT report said it all. AI has a value problem. And the blame rests – at least in part – with the executive team.

“We still have a lot of folks who believe that AI is Magic and will do whatever you want it to do with no thought.”

That’s a real quote, and it echoes a common story for data + AI teams

• An executive who doesn’t understand the technology sets the priority
• Project fails to provide value
• Pilot is scrapped
• Rinse and repeat

Companies are spending billions on AI pilots with no clear understanding of where or how AI will drive impact—and it’s having a demonstrable impact on not only pilot performance, but AI enthusiasm as a whole.

Getting to value needs to be the first, second, and third priorities. That means empowering the data + AI teams who understand both the technology and the data that’s going to power it with the autonomy to address real business problems—and the resources to make those use-cases reliable.

5. Cracking the code on agents vs. agentic workflows

While agentic aspirations have been fueling the hype machine over the last 18 months, the semantic debate between “agentic AI” an “agents” was finally held on the hallowed ground of LinkedIn’s comments section this summer.

At the heart of the issue is a material difference between the performance and cost of these two seemingly identical but surprisingly divergent tactics.

• Single-purpose agents are workhorses for specific, well-defined tasks where the scope is clear and results are predictable. Deploy them for focused, repetitive work.
• Agentic workflows tackle messy, multi-step processes by breaking them into manageable components. The trick is breaking big problems into discrete tasks that smaller models can handle, then using larger models to validate and aggregate results. Image: Monte Carlo’s Observability Agents

For example, Monte Carlo’s Troubleshooting Agent uses an agentic workflow to orchestrate hundreds of sub-agents to investigate the root causes of data + AI quality issues.

6. Embedding quality is in the spotlight—and monitoring is right behind it

Unlike the data products of old, AI in its various forms isn’t deterministic by nature. What goes in isn’t always what comes out. So, demystifying what good looks like in this context means measuring not just the outputs, but also the systems, code, and inputs that feed them.

Embeddings are one such system.

When embeddings fail to represent the semantic meaning of the source data, AI will receive the wrong context regardless of vector database or model performance. Which is precisely why embedding quality is becoming a mission-critical priority in 2025.

The most frequent embedding breaks are basic data issues: empty arrays, wrong dimensionality, corrupted vector values, etc. The problem is that most teams will only discover these problems when a response is obviously inaccurate.

One Monte Carlo customer captured the problem perfectly: “We don’t have any insight into how embeddings are being generated, what the new data is, and how it affects the training process. We are scared of switching embedding models because we don’t know how retraining will affect it. Do we have to retrain our models that use this stuff? Do we have to completely start over?”

As key dimensions of quality and performance come into focus, teams are beginning to define new monitoring strategies that can support embeddings in production; including factors like dimensionality, consistency, and vector completeness, among others.

7. Vector databases need a reality check

Vector databases aren’t new for 2025. What IS new is that data + AI teams are beginning to realize those vector databases they’ve been relying on might not be as reliable as they thought.

Over the last 24 months, vector databases (which store data as high-dimensional vectors that capture semantic meaning) have become the de facto infrastructure for RAG applications.** **And in recent months, they’ve also become a source of consternation for data + AI teams. ** **

Embeddings drift. Chunking strategies shift. Embedding models get updated. All this change creates silent performance degradation that’s often misdiagnosed as hallucinations — and sending teams down expensive rabbit holes to resolve them.

The challenge is that, unlike traditional databases with built-in monitoring, most teams lack the requisite visibility into vector search, embeddings, and agent behavior to catch vector problems before impact. This is likely to lead to a rise in vector database monitoring implementation, as well as other observability solutions to improve response accuracy.

8. Leading model architectures prioritize simplicity over performance

The AI model hosting landscape is consolidating around two clear winners: Databricks and AWS Bedrock. Both platforms are succeeding by embedding AI capabilities directly into existing data infrastructure rather than requiring teams to learn entirely new systems.

Databricks wins with tight integration between model training, deployment, and data processing. Teams can fine-tune models on the same platform where their data lives, eliminating the complexity of moving data between systems. Meanwhile, AWS Bedrock succeeds through breadth and enterprise-grade security, offering access to multiple foundation models from Anthropic, Meta, and others while maintaining strict data governance and compliance standards.

What’s causing others to fall behind? Fragmentation and complexity. Platforms that require extensive custom integration work or force teams to adopt entirely new toolchains are losing to solutions that fit into existing workflows.

Teams are choosing AI platforms based on operational simplicity and data integration capabilities rather than raw model performance. The winners understand that the best model is useless if it’s too complicated to deploy and maintain reliably.

9. Model Context Protocol (MCP) is the MVP

Model Context Protocol (MCP) has emerged as the game-changing “USB-C for AI”—a universal standard that lets AI applications connect to any data source without custom integrations.

Instead of building separate connectors for every database, CRM, or API, teams can use one protocol to give LLMs access to everything at the same time. And when models can pull from multiple data sources seamlessly, they deliver faster, more accurate responses.

Early adopters are already reporting major reductions in integration complexity and maintenance work by focusing on a single MCP implementation that works across their entire data ecosystem.

As a bonus, MCP also standardizes governance and logging — requirements that matter for enterprise deployment.

But don’t expect MCP to stay static. Many data and AI leaders expect an Agent Context Protocol (ACP) to emerge within the next year, handling even more complex context-sharing scenarios. Teams adopting MCP now will be ready for these advances as the standard evolves.

10. Unstructured data is the new gold (but is it fool’s gold?)

Most AI applications rely on unstructured data — like emails, documents, images, audio files, and support tickets — to provide the rich context that makes AI responses useful.

But while teams can monitor structured data with established tools, unstructured data has long operated in a blind spot. Traditional data quality monitoring can’t handle text files, images, or documents in the same way it tracks database tables.

Solutions like Monte Carlo’s unstructured data monitoring are addressing this gap for users by bringing automated quality checks to text and image fields across Snowflake, Databricks, and BigQuery.

Looking ahead, unstructured data monitoring will become as standard as traditional data quality checks. Organizations will implement comprehensive quality frameworks that treat all data — structured and unstructured — as critical assets requiring active monitoring and governance.

Image: Monte Carlo

Looking forward to 2026

If 2025 has taught us anything so far, it’s that the teams winning with AI aren’t the ones with the biggest budgets or the flashiest demos. The teams winning the AI race are the teams who’ve figured out how to deliver reliable, scalable, and trustworthy AI in production.

Winners aren’t made in a testing environment. They’re made in the hands of real users. Deliver adoptable AI solutions, and you’ll deliver demonstrable AI value. It’s that simple.