AI agents in 2026: 80% of companies already generate measurable ROI. The obstacle isn't the technology.

There's a tipping point that changes the conversation: when the data goes from "potential" to "measurable return". For AI agents, that point has already arrived. A recent study of more than 500 technical leaders at companies like Thomson Reuters, Doctolib, L'Oréal and eSentire confirms it: 80% report measurable economic returns on their AI investments.

And the most relevant thing isn't that figure. It's what comes next.

The three main obstacles to scaling AI agents aren't technical. They are integration with existing systems (46%), data access and quality (42%) and change management (39%). In other words: the brake isn't the language model. It's how teams work.

This connects directly with something we've been observing at onext in every implementation: the technology works. What's missing is the way of working to sustain it.

The real state of AI agents in 2026

Before diving into the analysis, it's worth putting the scale of adoption in context. The numbers from the "State of AI Agents 2026" study by Anthropic and Material leave little room for interpretation:

This is no longer an experimentation phase. It's adoption in production. And the savings figures reinforce it: 59% report savings in code generation, another 59% in documentation, 59% in code review and testing, and 58% in planning and ideation.

The pattern is clear: agents are absorbing the repetitive, routine work so teams can spend more time on decisions, architecture and strategy.

Four cases that explain where the real value is

The aggregate data tells one story. The individual cases tell a more concrete one. Four companies from the study illustrate patterns we see repeated in our own implementations:

Thomson Reuters: from hours to minutes in legal research

Thomson Reuters deployed CoCounsel, a legal agent that provides access to 150 years of case law. What used to require hours of manual searching is now resolved in minutes.

The key wasn't just the model. It was the integration with the existing data corpus. Without structured access to 150 years of documents, the agent is useless. The competitive advantage isn't in the LLM. It's in the context you provide it.

eSentire: cybersecurity analysis from 5 hours to 7 minutes

eSentire built a threat-analysis agent that cuts investigation time from 5 hours to 7 minutes, with 95% alignment with the conclusions of senior analysts.

Critical figure: 95% alignment with senior experts. It's not a generic chatbot. It's a system that has internalized the judgment of the team's best analysts. That's only possible if someone designed, structured and validated the context the agent consumes. Exactly what at onext we call context engineering.

Doctolib: replacing legacy infrastructure in hours, not weeks

Doctolib used code agents to replace legacy testing infrastructure. What would normally take weeks was completed in hours. The result: 40% faster feature delivery.

This case is particularly relevant because it shows agents aren't just for writing new code. They're for modernizing existing code, which is where most teams have the real bottleneck.

L'Oréal: conversational analytics with 99.9% accuracy

L'Oréal deployed a conversational analytics system that reaches 99.9% accuracy and is used by 44,000 monthly users.

Here the pattern is different: it's not software development, it's democratizing access to data. 44,000 people who previously depended on an analytics team now query data directly. That frees up the technical team for strategic work.

The three obstacles that aren't technical

If 80% report economic return, the logical question is: and the other 20%, and those who still don't scale, what's holding them back?

The study is fairly precise about it:

Notice: none of the three is "the model isn't good enough" or "the technology isn't mature". They're problems of work organization, knowledge and change management.

This confirms something we've been articulating in previous insights: AI success in development teams is organizational. Data from 500+ companies validates it at scale.

Where the real savings come from: not where you think

When people talk about AI and productivity, the first association is "generate code faster". But the study data paints a more nuanced picture:

What stands out is the uniform distribution. There's no single area that dominates the rest. The savings come from the entire cycle: planning, code, review, testing, documentation. Agents are acting as multipliers of the whole flow, not just one phase.

And there's a figure that often goes unnoticed: 60% use agents for data analysis and reporting, and 48% for internal process automation. That is, agents aren't just in the code. They're in operations.

"9 out of 10 leaders report that agents are changing how their teams work, spending more time on strategic activities, relationships and skill development instead of routine execution."
-- State of AI Agents 2026, Anthropic + Material

This quote sums up the fundamental change: it's not that teams do more. It's that they do different things. More strategy, less mechanical execution. More decisions, fewer repetitive tasks.

Software development as the epicenter of adoption

One figure from the study deserves its own section: 90% of organizations use AI for software development. Not for marketing, not for customer support. For writing, reviewing and deploying code.

And within that, 86% deploy agents specifically for code in production. Not proofs of concept. Not demos. Production.

This has direct implications for any CTO:

• If your team isn't using AI agents for development, you're competing against 90% of companies that are.
• If you're using them but without an adapted way of working, you're probably in the group that doesn't see real return.
• If you're using them with structured context, clear specifications and redesigned processes, you're in the 80% group that reports measurable return.

The difference between the second and third group isn't the tool. It's the system.

Related reading: In Code quality in the AI coding era we analyze how to maintain standards when 86% of code passes through an agent. The challenge isn't that AI writes code. It's that it writes the right code.

The leap ahead: from workflows to cross-functional processes

The study data shows an evolution in three phases that matches what we observe in real implementations:

Phase 3 is where organizational complexity spikes. It's no longer about a developer using an agent. It's about an agent operating across teams, with data from multiple sources and decisions that affect several departments.

And it's exactly where the study's three obstacles (integration, data, change management) become critical.

What the study validates about our thesis

At onext we've been implementing AI agents in development teams with a specific premise: technology is a necessary but not sufficient condition. The way of working is what determines the result.

The study data validates three pillars of our approach:

1. Context engineering as the foundation. The Thomson Reuters case (150 years of structured case law), the eSentire case (95% alignment with senior experts) and the L'Oréal case (99.9% accuracy) all demonstrate the same thing: the result depends on the context you provide the agent. Not on the agent itself. At onext, this is context engineering: the practice of designing and managing context as an engineering asset.

2. Structured specifications over ad hoc prompts. When 86% deploy agents in production, you can no longer depend on each developer writing their own prompt. You need specifications that guarantee consistency, traceability and quality. This is Spec-Driven Development (SDD).

3. Centers of Excellence as a transformation mechanism. The change-management obstacle (39%) isn't solved with one-off training. It's solved with a dedicated team that redesigns the way of working, measures impact and scales what works. Exactly what an AI Center of Excellence does.

What SDD solves: The study's three main obstacles -- integration (46%), data (42%) and change management (39%) -- are exactly what an AI Center of Excellence with SDD addresses. The project "constitution" structures the context. The specification templates guarantee the quality of the data the agent consumes. And the prompt playbook and the code review flow are the change-management mechanism: they change how the team works, not just which tool it uses.

Self-diagnosis: is your team ready for AI agents?

Based on the study data and our implementations, these are the signs that separate the teams that are ready from the ones that will run into the three obstacles:

The 80% ahead

The study also looks forward. 81% of organizations plan to tackle more complex use cases during 2026. 56% plan to implement agents for research and reporting. 29% are developing cross-functional agents.

This means the gap between companies that master AI agents and those that don't is going to widen. Not because of the technology, which will be available to everyone, but because of the organizational capacity to integrate it into the way of working.

The teams that reach phase 3 (cross-functional processes) first won't be the ones with the best access to models. They'll be the ones that have:

• Structured context as an engineering asset.
• Specifications that guarantee consistency at scale.
• Redesigned processes, not patched ones.
• Technical leadership that treats AI adoption as a system change, not a tool deployment.

80% already generate return. The obstacle for the rest isn't the technology. It's the way of working. And that can be changed.