Enterprises may not be really ready for Agentic AI or autonomous agents. For one there is confusion and Agentic washing on what is an AI agent and chat bots are often referred to as autonomous agents. Then there is an underestimation of how disruptive autonomous agents can become inside an enterprise. And most importantly autonomous agents need governance with identity and cost control. There are challenges ahead but also gains of positive disruption.
During 2025 we saw the adoption of Artificial Intelligence accelerate and the number of use cases surge forward. Across 2026, hybrid AI and distributed AI will continue to be the standard way of doing AI and the GPU world will continue on as before. There also promises to be lots of interesting dialogues around AI business models.
“While Agentic was number one on the list last year, this year, the word of the year is going to be governance,” says John Roese, Global Chief Technology Officer and Chief AI Officer, Dell Technologies.
As enterprises get through the learning curve of understanding AI technology and move towards production to get the best returns from AI or if they want to reduce the risks of using AI, they need rules.
The rules for governing AI are both external and internal. At present the state of external governance is challenging with more than a thousand governmental jurisdictions that have been created independently of each other and none of them are coordinated with each other.
The rules for governing AI are both external and internal.
“So as a practitioner using this technology, it is difficult for me to know how to navigate that. And so, I choose to avoid many highly regulated use cases,” says Rosse.
It is therefore essential and a call for action by governments around the world, and the industries in each country, to get together and spend time and energy to produce a more harmonised and rational governance from a regulatory perspective.
In 2026, if the external regulatory environment does not get better, it can become a significant drag on the ability of enterprises to put AI into production. It may be a drag to move use cases forward even when they are rational, simply because the enterprise may not be able to understand the regulatory frameworks that they are operating within.
The other half of governance is internal governance, which is the governance and the rules in which an enterprise defines how they will manage AI.
We do not know the full effect of AI in the near term, other than it will be disruptive.
According to Roese, Dell figured out early, that the right answer for them is to implement AI as a top-down, highly governed approach. This would help to move strategic projects into production without rest of the noise.
By comparison, Rosse finds that many enterprises have not yet achieved a strong level of internal governance. 
“They do not know how to create prioritisation. They do not know how to say no to things. They do not know how to associate ROI with a use case so that they can prioritise it,” point out Rosse.
In 2026, regulatory frameworks will either get better or worse, but are going to have a bigger impact on the enterprise. Inside the enterprise, there will be an increase of maturity as enterprises realise that internal governance is necessary for them to execute their AI strategy and move forward.
As AI technologies continues to evolve and new capabilities emerge, governance inside an enterprise will become a priority.
Arrival of autonomous agents
In 2025, one of the big mistakes made by enterprises was to call everything in AI an agent, which is not accurate.
Autonomous agents are different from chatbots, as an example. A true autonomous agent is a software system that has multiple components to it. An autonomous agent uses LLM to create world knowledge, to have language capabilities, and even to do reasoning.
In addition, an autonomous agent also includes specialised knowledge in the form of things like knowledge graphs, where you can add information that is not trained into the LLM in a way that is very specific to a human skill or a team skill.
Autonomous agents also store long and short-term memory so that they can keep track of things. They can learn skills. They can enhance their knowledge graph as they do work.
If you create a productive organisation, the work that can be done by machines will be done by machines.
Autonomous agents, unlike chatbots, are able to perceive the world around them and act on it. To do this, the protocol that has emerged is something called Model Context Protocol or MCP. It is a new protocol.
Model Context Protocol allows autonomous agents to read information from other datasets, to read external information, to access events, and to use tools.
However, one of the most important characteristics is that autonomous agents can talk to each other and collaborate. Another protocol facilitates not just agent to agent talk, but agent to agent talk across the organisation and with other organisations.
“When I use the term agent, I am referring to a fully autonomous system that has specialised knowledge, memory, can use tools, can perceive the world around it, and can work as a team inside the organisation and across countries,” says Roese.
Initially when organisation bring autonomous agents into the enterprise they use them as tools. However once organisations start using them they behave more like digital workers and because of their capabilities they change the way work is done around them.
Overtime they become good at managing complex activity and are elevated to the position of a coordinator or continuity manager for the complete team of humans and other autonomous agents. And even project managers.
And going from Roese’s own user experience inside Dell, “No matter what you think they are going to do for you, and no matter how you think they are going to integrate into your organisation, you are underestimating the change that is coming.”
Benefits of autonomous agents
Over time, organisations can become competent enough to clone their experts through an autonomous agent.
An expert human can share their knowledge and skills with an autonomous agent, and then use that skilled agent to work with other teams. That is the equivalent of cloning that individual human expert and then propagating that agent across the enterprise, doing far more than the individual expert could do on their own.
Another benefit is to go back and visit tasks that could not be completed due to excessive costs emerging from the time required to complete that task or the number of humans required to complete that task.
“You can go back and revisit projects that you would never have done because the cost and complexity of throwing people at a problem was too high,” explains Roese.
If an autonomous agent can do that task in the background, then that is a cost-effective way of completing that task.
In other words, as organisations start using autonomous agents they will change the way organisations operate, in ways organisations may not be able to anticipate. Autonomous agents improve the functioning of other agents since they talk to each other, but they also improve the functioning of human teams.
“We think there will be a lot of surprises at how people use agents and what it does to the organisational structure and the team dynamics in a positive way that was maybe not expected,” points out Roese.
Training and control of agents
For some use cases and some industries, it is vitally important to have full control of the agent as well as the infrastructure behind the agent. For example, if a robot is being used for national defence or for the military.
“You could not imagine a military buying autonomous drones as weapon systems and not controlling where the brain actually ran. And in many cases, the requirement will be that it has to be on sovereign infrastructure,” says Roese.
For such use cases, it is important that not only is the robot a national asset, but the infrastructure and brains behind the robot is also using sovereign infrastructure. And the same requirements could apply to healthcare, other national infrastructure and other compliant industries.
“When robotics gets deployed, the physical robot is in your country. But the backend systems that train it and optimise it and coordinate it, which are also AI factories, need to be somewhere that you have control over if it is in the national interest,” indicates Roese.
Going forward, the requirement to make autonomous agents into experts may include the need for governments to ratify them. For example, law practice, taxation and accounting, healthcare and so on. These ratifications for the agents extend from where the governments are hosted that is from the country’s sovereign infrastructure.
“Those agents that will now do that work will carry the same burden of having a connection back to the country that has authorised it. Not just logically, but physically, and sovereign infrastructure is a wonderful place to do that,” says Roese.
Humans and agents
“We do not know the full effect of AI in the near term, other than it will be disruptive. But we do know in the long term, if you create a productive organisation that is able to scale, the work that can be done by machines will be done by machines. But the work that is inseparably human, those roles continue to scale because agents allow you to reach more customers, to have better experiences,” says Roese.
Dell is finding is the biggest impact of autonomous agents is twofold. Autonomous agents are very good at cleaning up work that should not have been there for humans in the first place. For example, connecting and helping IT and data systems to work together more effectively without human effort.
“AI is a direct threat to those jobs and will absolutely displace them,” points out Roese. “On the other hand, when we bring agents into human teams, what we find is the agent actually brings order to some of the chaos that human teams have.”
Human teams become more structured and more efficient when they have an autonomous agent embedded inside their team.
Agents are good when they are given a goal. If the goal is for a team of human and agents to work together successfully to get to an outcome, that agent, especially because it speaks human, can be a very powerful tool to improve the productivity, not just of the agents, but of the people that are in that organisation with them
The presence of an autonomous agent helps humans to do their specific human work, while the agent completes the work that should not have been there for humans in the first place.
The agent is also able to help humans know what to do next, know what they did yesterday, know what other people are doing, and know what processes they are trying to achieve.
The months and years ahead promise to be interesting and disruptive for organisations that take the leap.
Hidden world of Agentic AI by John Roese
Agent washing
A chatbot is not an agent; a chatbot is just a tool. If you call it an agent, that does not make it an agent and it will create confusion. One of the biggest areas is there is a lot of confusion and misrepresentation of what agents actually are and what is an agent.
An autonomous agent is a new team member and not a new tool. When you bring an autonomous agent into your workforce, or a set of them, they will disrupt your organisation and will change the hierarchy.
That is the thing to realise with agents and it is the same as if you hired new people into your organisation. The team dynamic adjusts and the same thing will happen with agents.
Revisit work
Agents allow us to go and revisit work that historically humans have chosen not to do because there was not a good ROI. It was hard work that required a lot of resources to produce minimal value. Throwing people at it made no sense.
We have to go back and revisit them now because agents give us a more cost-effective way to do that.
Irrational costs
These technologies are resource intensive and can be very expensive. The pricing models for agents as a service are expensive and are not rational.
Dell has chosen to deploy many agents internally on its own systems and that allows Dell to control the cost. Dell also uses agents that exist in other providers and with some partners. In many cases, Dell had to spend a tremendous amount of time negotiating a pricing model that made sense.
People are underestimating the fact that it is an immature model and most people do not know how to price it properly. And if you sign up to the wrong price model, this could cost you a lot of money.
These are realistic problems that could materialise. And if you are not careful about things like pricing and economics with agents, especially if you are not building and operating them on your own infrastructure, they can get expensive, very fast.
I have told people if you thought cloud economics were problematic, agentic economics, if you do not manage and control them, could be a much more significant impact to your budget.
Kill switch
Inside Dell, all agents must carry an identity issued by Dell. It is like having an employee badge or a contractor badge. And the reason we are doing that is agents do work. And just like a human being coming in as a contractor to do work in Dell, I issue them the right to do that and I have the ability to revoke it.
That is a very obvious statement but no one is talking about it and most people have not even considered how they will handle broad agent identity from a perspective of control. Those are the decisions people are going to have to make as they go forward.
Governance
All of these are solvable, but they require governance, a set of rules and parameters in which you are going to operate.
It is not hard, but it is new and it is going to be a big piece of the governance that is necessary to avoid some of the potential problems like excess cost, lack of control, transfer of knowledge inadvertently.
All of these things can be mitigated if you think about how to put governance around it upfront.
Key takeaways
- As enterprises get through the learning curve of understanding AI technology and move towards production, they need rules.
- In 2026, regulatory frameworks will either get better or worse, but are going to have a bigger impact on the enterprise.
- In 2026, if the external regulatory environment does not get better, it can become a drag on the ability of enterprises to put AI into production
- Human teams become more structured and more efficient when they have an autonomous agent embedded inside their team.
- Dell figured out that the right answer for them is to implement AI as a top-down, highly governed approach.
- Autonomous agents, unlike chatbots, are able to perceive the world around them and act on it.
- Model Context Protocol allows autonomous agents to read information from other datasets, to read external information, to access events, and to use tools.
- One of the most important characteristics is that autonomous agents can talk to each other and collaborate.
- For some use cases and some industries, it is vitally important to have full control of the agent as well as the infrastructure behind the agent.
- You could not imagine military buying autonomous drones as weapon systems and not controlling where the brain actually ran.
- The work that is inseparably human, those roles continue to scale because agents allow you to reach more customers.
