AI agents are transforming into digital middlemen that promise unmatched convenience while quietly compromising user privacy, data security, and personal autonomy for corporate profit.

AI agents are transforming into digital middlemen that promise unmatched convenience while quietly compromising user privacy, data security, and personal autonomy for corporate profit.

AI agents are transforming into digital middlemen that promise unmatched convenience while quietly compromising user privacy, data security, and personal autonomy for corporate profit.

“I can’t wait to pay a broker” – has anyone ever heard someone saying this? Brokers are tolerated only because they stand between people and something they need—real estate, a flight or a loan. They find ways, usually unknown to the public, to fetch something normal and charge for it. A ray of hope twinkled when the internet arrived, raising hopes of their removal. Delete the travel agent, the bank clerk, the shop counter – go direct! However, the middlemen have now evolved into a new avatar, and they live inside users’ phones, for a change, at the users’ own invitation.

An agent will handle it all: buy a garment, book a trip, answer an email. It seems so attractive. But look at what has actually been built. It not only provides advice but also sits between the user and everything they do online, deciding what they see, skip or buy. This is the oldest definition of a middleman, but with a difference: it never sleeps and never sends a bill that is easily decipherable.

Agentic matters now reach courtrooms. Perplexity was sued by Amazon in November 2025 after its Comet browser agent started placing orders on Amazon’s site. Comet disguised itself as a Google Chrome browser and refused to identify itself as an AI agent while shopping, Amazon claimed. Perplexity was blocked by a federal judge in March 2026. However, a US appeals court overturned the ruling on August 4, 2026, stating that it was not the company but the users who accessed Amazon. Yet both sides are playing the middleman game. Amazon wants to stay as the layer between the user and everything else. Perplexity wants to be the layer between the user and the store. Amazon’s lucrative advertising business, which earns money from product placement on its pages, is at risk due to shopping agents. Perplexity’s point is that agents have no eyes to see all that advertising. Funny enough, but notice what it accepts: the agent’s whole value is that it hides the marketplace’s noise from the user, yet who decides what gets hidden next? And no one is fighting for the customer. Two toll collectors are arguing over who gets the check post on the bridge.

An astonishing twenty-fold increase in revenue has been reported by Rezolve AI in the first half of 2026, reaching $130 million, compared to $6.32 million in the first half of 2025. However, the emphasis is not on profit as such. Rezolve saw its first profitable month only at the end of 2025. Yet money is flowing in tonnes. CB Insights has valued Cursor, a coding agent founded in 2022, at about $500 million in revenue. Premium valuations are being given to businesses rapidly replacing human support teams with customer service agents. The gamble is on agents replacing support workers rather than assisting them. Cash may be burnt on compute, but several such firms are chasing scale. But this is the way toll roads are created. Early traffic is subsidised until the route is captured, and then the actual toll starts.

And the price of convenience? It is users’ data. Only when the agent knows a user’s bank login or other details can it book a tour or pay for cable. It needs the passwords to the user’s whole life to remain useful. Yet the locks are weak enough. In November 2025, security firm SquareX discovered a major loophole in Comet, allowing an embedded server to reach local files on a user’s machine. OpenAI’s CISO, Dane Stuckey, has confirmed that this is a frontier, unsolved security problem. Indirect prompt injection, where hidden instructions on a webpage hijack the agent, is a systemic vulnerability across AI-powered browsers, as flagged by Brave’s researchers. As users’ login details are available to agents, they can easily access their emails, company systems and even banks.

A research study by researchers at UC Davis, UCL and Mediterranea University of Reggio Calabria, published at the USENIX Security Symposium, highlights an acute privacy crisis within the generative AI browser extension ecosystem, revealing that AI browser assistants send full webpage content, including form data such as financial or medical details, to remote servers without adequate transparency or opt-outs. No one reads the micro-sized terms and conditions. Middlemen used to take commissions. This creates a copy of the user’s life.

And this vicious circle of modern slavery only becomes more entrenched. The boring stuff is delegated to agents first and subsequently, the very choices themselves. And one fine morning, the user is shocked to find strange flights, unheard-of restaurants, unmatched curtains – all selected by agents. A system designed for revenue-maximisation quietly captures the user’s individuality. Clicking ‘approve’ is still needed, yet it is only notional, and applies to the catalogue created by someone else, effectively turning the user into a signature at the bottom of a machine’s discretion. And this dependency has high-growth potential. The more the agent learns, the harder it is to leave, as starting over means retraining a stranger from scratch.

Humanity has been wired since time immemorial to believe that it cannot survive without middlemen, whether real or artificial. An agent that really works for the user, is present on one’s device, answers to the user alone and takes no cut from the seller may be a genuine gift. Consider it as the opposite of a middleman: an employee. However, this may not be the business model on offer. It may be a company that stands between the user and the world, learns the user’s details, and gets paid from both sides. This has a familiar name which has been hated for centuries.

There may be bigger issues cropping up in the future, such as a common agent working for a married couple. What if they get divorced, and who would own the agent then? What about the ownership of common jobs done by it? Also, a Mint report shows IT firms trying to police agents as there are increasing instances of them going rogue.

Yet some innocuous things may hold the key, such as ensuring that sensitive data stays on the user’s device unless there is clear consent to use it somewhere else, and that such consent must not be hidden in settings or obtained automatically through a so-called opt-out option. Loyalty: an agent should owe a duty to its user, the way a lawyer or a doctor does. If there is a fee or a favour from a company or partner, it must be explicit. A real exit: the history must be exportable with no traces left, so that the user may walk away any fine morning.

Until then, every agent should be treated in a similar manner to how one would treat a stranger who offers to hold one’s wallet just for convenience without disclosing the charge.

When the service is swift, smooth and free, try to discover who is collecting the toll. If one cannot see the booth, it is probably where one stands.

(The author is a public policy expert based in New Delhi. The opinions expressed in this article are those of the author and do not purport to reflect the opinions or views of THE WEEK.)