Everywhere you look, there’s a new AI headline telling you what should be happening.
But the reality inside product teams is usually very different from the stories on LinkedIn.
The latest survey on AI builders gives us a clearer picture of what teams are actually doing. Not the hype, not the wishful thinking, but the real workflows, the common tradeoffs, and the messy experiments happening behind the scenes.
AI development is maturing quickly. Teams are experimenting, operationalizing, evaluating, fine-tuning, and connecting AI directly into the heart of their systems.
Everyone is past the honeymoon phase and rushing to execution.
Companies that build solid foundations today such as strong evals, model strategies, and reliable agent tooling, will be the ones shipping the most compelling AI products in the years ahead.
Below are the nine themes that stood out most. Some of them match the public narrative. Most of them don’t.
1. Teams Are in the “Try Everything” Phase
If there’s one theme that shows up everywhere, it’s this:
Teams are testing every approach they can get their hands on.
Agents? Yep.
RAG? Of course.
Evals? Several types.
Synthetic data? More common than you’d expect.
Tools that launched four months ago and are still half-baked? Also yes.
Even the Model Context Protocol (MCP) is already showing up in real workflows. Adoption is still low, but the fact it’s showing up at all tells you where things are heading.
This is the exploration phase.
Messy, experimental, and absolutely necessary.
2. Open Source Is Winning Where It Matters
One of the clearest signals: open-source models dominate real-world usage.
Here’s how teams describe their model strategies:
- 21.8% → exclusively open source
- 44.6% → mostly open source
- 66.4% total → open source is the default
Closed-source APIs still have strong adoption, but they’ve become “precision tools” reserved for when latency or accuracy requirements justify the cost. When teams want control, affordability, and customizability, they choose open source.
3. The Biggest AI Impact Isn’t Customer-Facing
There’s this narrative that AI innovation is all about customer-facing experiences, but teams report something different. Internal tools are actually leading the pack:
- 65.6% → using AI to build internal tools
- 61.3% → using AI to improve existing products
- 57.1% → building internal AI experiences
This makes sense: internal workflows have fewer constraints, less regulatory pressure, and usually a clearer ROI. So while everyone talks about the shiny customer-facing features, the real productivity boost is happening behind the scenes.
4. Agents Are Getting Real System Access (Not Just Demo Mode)
A lot of agent frameworks still feel experimental. But teams are clearly giving them meaningful access:
- 72.4% → database access
- 59.1% → web search
- 55% → memory systems and file systems
- 45.8% → code interpreter capabilities
This is a strong sign that agents aren’t just being used for weekend hackathon prototypes. They’re being plugged into real systems, with real responsibilities, and real risks.
It also means agent safety and oversight are about to get a lot more important.
5. Evaluations Are Everywhere (And Getting More Rigorous)
If AI systems are going to be trusted, they need evals. According to the survey, over 99% of respondents are running evaluation processes.
Types of evals:
- AI-based evals: 66.6%
- User surveys: 60%
- A/B testing: 47.7%
- Human judgment: 41.2%
And for automated checks:
- Verification: 52.5%
- Error analysis: 50.8%
- Regression tests: 49.4%
- Unit tests: 40.2%
This isn’t just a QA step anymore. Evaluations are being used throughout the entire lifecycle, from early specs to post-deployment monitoring.
It’s one of the clearest signs that AI engineering is maturing into its own discipline.
6. RLFT Is Showing Wildly Strong Results
Reinforcement learning from human feedback (RLFT) is still early… but the numbers are hard to ignore:
- 74.6% → saw >16% performance lift
- 30.5% → saw >30% lift
- A handful → reported >45% lift
In any other field, a 30% performance improvement would be considered a breakthrough.
Here, it’s becoming normal.
This is probably the clearest “this will be standard practice soon” signal in the entire survey.
7. Fine-Tuning Is Mainstream But Startups Are Behind
Fine-tuning is widely used overall (80% adoption) but there’s a major divide:
- 52.4% of startups → not fine-tuning
- Only 17% of enterprises → not fine-tuning
The simplest explanation:
Startups prefer base models because they’re cheaper, faster to implement, and “good enough” early on.
Enterprises, on the other hand, have the data, the regulatory pressure, and the scale to make fine-tuning worth the investment.
As models continue improving, this gap may close. But for now, it’s a clear signal of where organizational maturity plays a role.
8. MCP Is Still Early But Adoption Is Rising Fast
Model Context Protocol (MCP) is still in its early adopter phase, but the numbers already look promising:
- 16.7% → using MCP servers
- 32.9% → using MCP via LLM clients
Use cases include:
- connecting to customer data
- hooking into data warehouses
- bridging internal tools
- powering more capable agents
This feels similar to the early days of APIs-as-products: a small group experimenting, and everyone else watching closely.
Expect adoption to accelerate quickly.
9. Synthetic Data Is Quietly Becoming Standard
Synthetic data has moved from “interesting idea” to everyday tool:
- 63% → using it for evaluations
- 22.3% → using it for fine-tuning
Teams aren’t waiting for perfect datasets anymore.
They’re generating what they need, stress-testing agents, and bootstrapping new model behaviors before real data arrives.
A new market of synthetic data marketplaces and evaluation sets is likely just around the corner.
