Proponents are hailing them as transformative. Critics are warning of massive risks, including privacy breaches and loss of human control. Could "agentic AIs" redefine how we work and live, or are they a dangerous step toward unchecked automation? — Photo: Zacharie Scheurer/dpa
BERLIN: Developers say they can do nearly any task a human can at a computer. Critics say they are incredibly dangerous.
After a host of major tech companies announced a new level of artificial intelligence that can perform complex tasks independently and act on behalf of users across various websites and software, 2026 is shaping up to be the year of AI agents.
Unlike the AI chatbots and voice assistants that are widespread today, so-called AI agents can independently plan several steps, access various digital services and make decisions on behalf of a user.
Developers say they can find a recipe for you online and order all the ingredients, can take over most of the HR and IT tasks required with on-boarding a new employee and can master many complex software tasks currently handled by humans sitting at computers.
"AI agents will change how we all work and live," Amazon head Andy Jassy said in June. "Agentic AI is a new labour model," said Marc Benioff, chief executive of customer relations giant Salesforce in February.
Jensen Huang, chief executive of NVIDIA even went so far as to say as that whole HR departments will disappear into IT departments, while customer support software company Zendesk now says that AI agents are already able to handle around 80% of support problems.
However, sceptics say that allowing AI agents to autonomously plan, act and pursue goals poses no less than "catastrophic" risks, with concerns ranging from a major loss of privacy all the way to a public loss of control under superintelligent agentic AI.
'This is a huge threat'
The problem with agents, says Meredith Whittaker, head of the privacy-focused chat app Signal, is that they "demand huge amounts of access, nearly unlimited permissions and... provide vectors into key applications and services, whether it's your calendar or your Gmail or your Signal."
Speaking to dpa, Whittaker gives a common example of an AI agent that makes a restaurant reservation. The agent finds a time on a calendar when it works for you and your three friends. It then messages your three friends telling them the reservation is made.
To do this, it needs access to your credit card information, potentially to your browser and to everything that's in your calendar, Whittaker says.
"It can do all of this without asking for your permissions, in a way that pools a huge amount of sensitive data that is ripe for attack," she says.
"So we think this is incredibly dangerous. This is a huge threat."
Margaret Mitchell, a former AI ethics leader at Google, is among the AI researchers to have published papers arguing for an end to the development of agentic AI. "The more control a user cedes to an AI agent, the more risks to people arise," Mitchell and her co-authors write.
"Despite how useful these systems might be, unchecked AI agency poses significant risks to public safety and security, ranging from misuse by malicious actors to a potentially irreversible loss of human control," writes deep learning pioneer Yoshua Bengio with co-authors in another paper.
Whittaker says that to ensure that apps with private user content can effectively defend against surveillance, the manufacturers of operating systems for computers and smartphones must take responsibility and implement developer-level options to oppose AI agents.
Major tech companies have shown little sign of slowing down, however, and Google, Microsoft, Amazon and ChatGPT developer OpenAI, as well as many smaller rivals, are all working on their own AI agents. – dpa
What AI agents already exist?
- Google has developed AI agents based on Gemini 2.0 with "Project Mariner". Just like humans, these can click, type and scroll in a web browser and perform tasks independently, such as ordering various products online for a project named by the user.
- Salesforce is switching to AI agents in customer service. The company demonstrated how its software can make telephone calls independently, for example when exchanging a jumper that is too small. Salesforce chief executive Marc Benioff expects customers to create 1 billion such AI agents for customer service by the end of 2026.
- DeepL, a German AI-supported translation start-up, has unveiled an AI agent that can already autonomously work in editorial systems. The "DeepL Agent" can handle "nearly any task a human can," notably pointing and clicking and using web browsers, the company says.
- Microsoft has demonstrated its "Factory Operations Agent", an AI-supported assistant for optimising processes in factories.
- Amazon is developing Alexa into a personalised AI chatbot that knows your contacts, appointments and habits, making it more personal and useful than other programmes. Amazon sees the future in AI agents that can independently perform multiple tasks on behalf of users.
- Parloa, a Berlin start-up, specialises in AI agents for customer meetings. Its AI agents can even actively call customers, for example to offer upgrades for air travel.
