">
Faciliter AI
Try MyFAQ.ai
AI Trends & Research

Agentic AI: From Promise to First Real Business Impact

JN
Julien Nadaud
| | 7 min read | English

My agent prediction was partly right, but the reality was uneven. The strongest agentic shift happened in software development, moving far faster than expected. This created an agentic divide, as enterprise automation hit a wall due to fragmented processes.

Agentic AI: From Promise to First Real Business Impact
Last year, I wrote several posts saying that 2025 would be the year of AI agents. My view was simple: we were moving beyond pure LLM usage and entering a new phase built around agent orchestration, tool use, and more autonomous execution. The idea was that AI would not only generate content, but start doing real tasks for businesses by combining structured outputs, tool calling, code generation and execution, vision, browser interaction, and data search.

Now, in March 2026, I think this prediction was partly right. We did see a major shift from generative AI to more agentic systems. But the reality is more uneven than many people expected. Some areas moved very fast. Others are still much more limited than the hype suggested. That is why, for me, the real story is not that agents are suddenly everywhere. It is that agents became real first where the environment was already ready for them.

The coding catalyst: where the value became real

The biggest breakthrough did not come first from general enterprise automation. It came from coding. Tools like Cursor, Claude Code, OpenAI Codex, and GitHub Copilot agent mode pushed agentic AI in software development much faster than I expected. In this domain, the value became concrete: writing code, debugging, refactoring, testing, navigating repositories, and handling more complex workflows with much less manual effort.

Honestly, this progress was above my expectations for 2025. We have clearly moved from simple autocomplete to something closer to delegation-first workflows, where the engineer increasingly becomes a supervisor, reviewer, and decision-maker rather than the person doing every step manually. Cursor now openly presents agents as a way to hand off implementation while you focus on decisions. OpenAI describes Codex as a software engineering agent that can work on many tasks in parallel. GitHub describes Copilot agent mode as an autonomous peer programmer able to read files, propose edits, run commands, and iterate on its own output.This was the real surprise for me: the strongest agentic shift happened first in software development, and it happened faster than I expected.

The agentic divide: coding vs. enterprise

What is interesting is that this acceleration in coding has not yet translated at the same scale to other business areas, especially inside large enterprises. This is probably one of the biggest lessons of the last 12 months.

Coding was a good environment for agents because the work is already digital, the tools are programmable, the feedback loops are faster, and success can often be measured clearly: tests pass, builds run, code compiles, diffs can be reviewed. In enterprise operations, things are different. Processes are fragmented, systems are disconnected, responsibilities are unclear, and governance creates friction everywhere. So while agentic coding moved very fast, business automation hit a wall much earlier.

A lot of first-generation frameworks and platforms were presented as if they would become magic tools for process automation. In reality, most of them were useful building blocks, but not complete solutions. Real business workflows are rarely linear. They involve loops, exceptions, approvals, state, poor data quality, and many hidden dependencies.

This is exactly why frameworks like LangGraph, CrewAI, and Microsoft Copilot Studio agent flows are more interesting than the earlier “prompt chain” mindset. They focus much more on stateful orchestration, long-running execution, and human-in-the-loop control.

A new generation of agents is emerging

Another development I find especially interesting is the move from reactive assistants to more persistent and proactive systems.

Instead of waiting for one prompt after another, these agents are increasingly designed to observe, act, monitor, and continue working across tools and environments with less direct supervision. That is visible in coding, but also in newer operational agent projects like OpenClaw, which presents itself as a personal AI assistant and autonomous agent able to interact with files, the browser, messaging tools, and system-level actions.

I am still careful with many of the claims around this new wave, because this space is full of buzz and exaggeration. But the broader direction is real: we are starting to see systems designed not just to answer, but to operate. That is a meaningful change.

Why first-generation frameworks were not enough

My take is that many first-generation agentic frameworks were overestimated. Tools like early LangChain, browser automation layers, or no-code agent platforms from large vendors all brought something useful. But they did not become a universal answer for business process automation.

Why? Because the hard part was never only connecting a model to a tool. The hard part is making the system reliable inside the reality of a company: permissions, compliance, bad data, unclear ownership, broken workflows, change management, and the fact that most organizations are not designed to work with autonomous systems.

That is why the next step is not just better models. It is better architectures, guardrails, memory, evaluation, orchestration, and organizational readiness. The growing attention around resources like the OWASP Top 10 for Agentic Applications for 2026 shows that the conversation is also becoming more operational and more serious.

2026 should be the year of specialized business agents

This is why I believe 2026 will be the real beginning of the next phase. Not because all companies will suddenly become autonomous, and not because entire businesses will run on agents overnight. But because we are reaching the point where specialized agents can start doing real work in specific parts of companies.This will probably happen first in narrow, high-value use cases:

  • customer support
  • sales operations
  • internal research
  • legal support
  • translation
  • marketing production
  • consulting support tasks
  • reporting and internal coordination
We already see early movement in some of these areas. In customer support, vendors like Zendesk are openly positioning AI agents as the next generation of service automation. In regulated work, Thomson Reuters CoCounsel is pushing agentic and generative capabilities into legal, tax, audit, and compliance workflows, with a strong emphasis on trusted content, validation, and security.

For now, these are still early shocks, not a full transformation. But I expect the pace of change to accelerate this year. What happened in coding will not stay limited to coding.

The SaaS pressure is only the beginning

The shift in coding is already strong enough to put pressure on parts of the software industry. And this same pressure is starting to appear in the broader SaaS ecosystem. A lot of software categories were built around the idea that humans would manually perform repetitive digital work inside structured interfaces. But if agents become able to perform more of that work directly, then the value of some software layers starts to change.

This does not mean SaaS disappears. But it does mean that many software products will need to be redesigned around a world where the main user is not only a human, but also an agent. Even investment and operator discussions are now shifting toward questions of API control, access to systems of record, and what happens when AI products depend on data that incumbents can restrict. A good example is this a16z piece on the new API battleground.

Security, control, and business reality

Of course, as agents move from chatting to acting, the risks also increase. Security, governance, permissions, and monitoring become much more important when systems can browse, execute, access data, send messages, or trigger workflows. This is another reason why enterprise adoption has been slower outside coding. The technical capability is one thing. Trusting it in production is something else.

This is also why I think the winners in this space will not only be the companies with the smartest models, but the ones that can create reliable, auditable, controllable agentic systems.

My own experience

What I find most striking is that this shift is not only technological. It is also personal and organizational.

In my own case, I am still amazed by the possibilities, and I still feel I have not reached the limit yet. But I also see that this is not just about using better tools. It is a new paradigm for running a business. And that requires a much deeper transformation.

The hardest part is often not the technology itself. It is the readiness to rethink roles, workflows, decisions, and responsibility. At some point, you stop asking how AI can help one person work faster, and you start asking how a company should be designed if agents can take over an increasing part of execution.That is a very different question. For me, this shift is opening the possibility to build an entirely new business based on AI and agents. Not as a side feature, but as a core operating model.We are still early. But the direction is becoming hard to ignore.

The real challenge is no longer only what AI can do, it is whether we are ready to let it act.

Share this article

Related Articles

Back to Blog