Most business leaders have heard the phrase "AI agents" by now. Far fewer can say what an agent is doing inside their company today, measured in hours saved or errors avoided.
That gap is the real story of 2026. Not whether agents work, but where they work, and why so many projects never leave the pilot stage.
I run ThalerTech Group, where several of our companies build and deploy agents for businesses and the public sector. This is not a news roundup. It is what I believe, based on that work, about what separates the useful agents from the expensive demos.
The adoption gap is real
The public data tells a consistent story: lots of experimentation, much less scaling.
According to McKinsey's "The State of AI in 2025" survey, 62 percent of respondents said their organisations were at least experimenting with AI agents, but only 23 percent said they were scaling them somewhere in the enterprise. In any single business function, no more than 10 percent reported scaling agents.
Gartner was even more direct. In June 2025, it predicted that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls. Gartner also warned about "agent washing", where existing chatbots or automation tools get relabelled as agents.
I do not read these numbers as a verdict against agents. I read them as a verdict against vague projects. When the goal is "do something with AI", the project dies. When the goal is "cut the time to process a supplier invoice in half", it has a chance.
What actually works
In our projects we often see the same patterns among the agents that survive past the pilot.
1. Narrow, repetitive, text-heavy tasks
The best first agents do one job that a person currently does many times a week: triaging incoming emails, drafting first replies, extracting data from documents, preparing a daily summary, updating a CRM after a call. These tasks are boring, which is exactly why they are a good fit. Success is easy to define and easy to measure.
2. Agents that live where people already work
An agent that requires a new dashboard, a new login and a new habit will struggle. An agent that sits in the messaging app the team already uses gets adopted much faster. This is why we work a lot with OpenClaw, an open-source platform for personal AI agents that live in Telegram or WhatsApp, with memory, tools and scheduled tasks. Through Claws, we deploy it for French companies that want agents inside their existing workflows rather than another piece of software.
3. A human in the loop, at the right place
The agents that work rarely act fully alone on high-stakes decisions. They prepare, draft, suggest and flag. A person approves the payment, sends the sensitive email, signs off the contract. Over time, as trust builds, the approval step can move. Starting with full autonomy is how projects end up in Gartner's 40 percent.
4. Clear access to data and tools
An agent is only as good as what it can see and do. In our projects we often see that when a pilot stalls, the model is rarely the problem. The usual cause is that the agent had no reliable access to the right documents, the right system, or the right permissions. Integration work is unglamorous, and it is most of the job.
5. Trained teams, not just installed software
This one is underrated. A team that understands what the agent can and cannot do will use it well, give it better instructions and catch its mistakes. A team that was simply told "we have AI now" will either ignore it or over-trust it. That is why at ThalerTech we pair custom agents with team training. The training is not an add-on; it is part of what makes the deployment work.
What usually does not work
Some patterns come up again and again in stalled projects.
- The "general assistant for everyone" project. Broad scope, no owner, no metric. It impresses in a demo and fades within weeks.
- Automating a broken process. If the underlying workflow is unclear, an agent will execute the confusion faster.
- Ignoring security. Agents that read emails, browse the web and call tools create a new attack surface. Testing them like any other exposed system is not optional. It is one of the reasons we built Cybeau, which focuses on AI offensive security testing.
- Forgetting language and context. Many agents work beautifully in English and poorly elsewhere. For companies operating in Arabic-speaking markets, this is a serious blind spot, which I discuss in Arabic AI: the market's blind spot.
How I would start in 2026
If you run a small or mid-sized business and want to begin, here is the approach I would recommend.
- Pick one task. Something done often, mostly with text, with a clear before and after.
- Measure the baseline. How long does it take today? How often does it go wrong? Without this, you will never know if the agent helped.
- Put the agent where your team already is. Email, Telegram, WhatsApp, your existing tools.
- Keep a human approval step on anything that touches money, customers or legal commitments.
- Train the people, not just the system.
- Review after a few weeks, then decide: expand, adjust or stop. Stopping is a valid outcome.
For individuals who simply want to feel what a personal agent is like, a consumer product is the easiest entry point. That was the idea behind Nino, and I shared what we learned building it in Building Nino: lessons from launching a consumer AI assistant.
AI agents are not magic and they are not a fad. They are a new kind of software that is very good at certain jobs and still unreliable at others. The companies getting value in 2026 are not the ones with the most ambitious vision. They are the ones that picked a narrow problem, measured it, and executed patiently.
Ideas, no matter how brilliant, are mere potentialities. That applies to AI strategies too.
Key takeaways
- Surveys show broad experimentation with agents but limited scaling, according to McKinsey.
- Gartner expects over 40 percent of agentic AI projects to be canceled by the end of 2027, mostly due to cost and unclear value.
- Start with one narrow, repetitive, measurable task.
- Deploy agents inside tools your team already uses, with a human approval step.
- Treat integration, security and training as core work, not extras.