You have surely seen them: these AI agents able to browse websites, click, fill in forms, book tickets, all of it… fully autonomously.
Right now, solutions are exploding: Twin (France), BrowserBase, H Company (France), Operator from OpenAI, Claude Computer Use, etc.
But let's be honest, for a long time I wondered whether it was just a gimmick:
- Why control a browser when a simple API or well-done scraping can be enough?
Example: if I want to know the flights available on Google Flights for a destination, I could ask Claude, which has a Google Flights connection (via MCP).
- Why pay more for an action that a script or an HTTP request can do 10 times faster?
Example: if I want to retrieve information from a site that has no API, I can fetch the site's HTML content so an AI model extracts the information I am looking for.
Yet, after digging into the use cases and testing some platforms, there are concrete cases where AI web agents open doors that neither APIs nor scripts can open. I am sharing my feedback with you.
An explosion of web agent tools
As I dug into the topic, I realized it was not just a fad. The market is exploding, with players who are not in the same league: some, like Twin, move forward on very targeted use cases such as retrieving invoices for Qonto.
Others, like H Company, aim for a full automation OS, having raised 200M€ to get there (despite the CEO being shown the door).
The ecosystem is organizing around a real promise: making AI able to act on the web like a human.
Why this boom now?
Three main reasons, in my view:
- The models have gained in reasoning ability, which lets them follow complex, multi-step instructions.
- Multimodal models like GPT-4V can now 'read' a visual interface like a human.
- And above all, inference costs have plummeted, making these uses technically accessible, where they were reserved for demos just a year ago.
Add to that infrastructures like BrowserBase that make it all scalable, and you understand why the explosion is as brutal as it is logical.
A quick glimpse of what BrowserBase makes possible:
Concrete business use cases
This is where I really started to understand the value. AI web agents are not just there to 'click for us'. They unlock cases you simply could not automate before, or only in a very shaky way.
For example, in the case of Twin: their agent can automatically log in to user accounts to retrieve invoices online:
- Either by using session cookies (to access authenticated interfaces without a password)
- or by displaying an intermediate interface where the user enters their credentials before the agent takes over.
The result: no more need to develop an API integration for each provider, it works on any existing web interface, just as a human would.
But the use cases do not stop there.
- Need to update a product catalog on a marketplace that changes its design every two weeks? The agent adapts.
- Do you want to book a transport slot or a medical appointment as soon as it becomes available, on a public site with no API? The agent can watch and act automatically.
- In QA, some agents like Tester H from H Company can automatically check the display, the payment funnel or the language of a site after each deployment, without maintaining dozens of fragile Selenium scripts.
- Here, form filling on Skyvern:
The current limits
For some cases like lead enrichment, I remain very skeptical. An agent can browse a site or a LinkedIn profile, read all the content, and generate an enriched contextual summary. But in practice, tools like Clay, Apollo or PhantomBuster already do this job far faster, more reliably and more scalably, thanks to optimized APIs.
The web agent adds a degree of intelligence (it can analyze the page as a whole, spot weak signals, cross-reference data) but at a price that does not compete with the existing solutions. The latency and cost-per-task constraints are far too high. It is a bit like using a bazooka to kill a fly.
Beyond cost, the technical limits are still many.
- Maintaining context across multiple pages remains fragile → moving from one step to the next in a sequential navigation without losing the thread is a real challenge.
- Error handling (slow-loading pages, Captchas, unexpected modals…) is not always reliable.
- And even if some agents are multimodal, visual understanding remains imperfect: a moved button or an ambiguous icon can derail the action.
Powerful tools, but still unstable, costly, and often very (too?) fragile for complex or critical workflows.
Why (and when) it is relevant
After exploring these tools (and deploying browseruse on 1 specific case), I come away with a conviction: AI web agents can be very useful on the consumer side (it unblocks very specific situations), but as soon as you have a large volume of tasks in a company, it is very rarely relevant.
Here are the cases where it potentially can be:
Case 1: No API available.
This is probably the most obvious case. And it was my use in an insurance company to retrieve certificates from an interface and let the AI handle tickets in that category autonomously. No API, no webhook, just a client extranet. A web agent can log in and download the document in 1 step (that is 0.10€ per resolution). Without it, this was a recurring manual task for an already overloaded support team.
Case 2: An unstable or frequently changing interface.
Here, agents clearly do better than classic RPA scripts. I am thinking in particular of data entry in an internal application-management tool (an ATS type). Until now, this went through rigid scripts (like UiPath), but the slightest change to the interface made them crash. A web agent can detect the fields visually, adapt its behavior and carry on without us having to rewrite the whole scenario. It is far more robust and flexible.
Case 3: Specific action chains, hard to script.
A concrete example: an HR team wanted to pre-fill files on various partner sites from an internal CRM. No standard, each site has its own form, formats and steps. The agent can read the record in the CRM, understand which info goes where, and fill in the fields in the right place on the external site. We are no longer in a 'data-in / data-out' logic like with an API, but in intelligent orchestration across multiple contexts.
But we must stay clear-headed: these uses have a cost. The pricing varies.
Some, like Skyvern, advertise $0.10 per task, others like Operator cost $200/month with no API and only through an interface, or BrowserBase charges $0.03 per step (sometimes less depending on the models).
Not to mention that you have to run tests and work on the prompting: you do not launch an AI agent like a webhook. You have to describe what it must do, handle error cases, test.
In short, it is not magic. But when you have no API, when interfaces change, or when the need involves business logic scattered across the web, that is where these agents become truly interesting.
Toward a new automation layer
What is interesting about web agents is that we are at the dawn of a new software layer, a universal interface between AIs and the web. It is a major paradigm shift: it is no longer humans who navigate interfaces, but their agents.
At their core, these agents push back today's limits: they can already navigate, fill in forms, click, interpret complex interfaces, and carry out multi-step tasks.
Tomorrow, they will be able to buy, book, compare, negotiate, and even interface with one another in agent-to-agent logics.
But everything remains to be built:
The business model: today, we bill per task or per hour. But if an agent starts running continuous processes, will we have to pay per action? per success?
The platform war: as agents multiply, some sites may block or filter their access. We have already seen it with certain AI crawlers. And tomorrow, what will Google, Amazon or Meta do when agents come to interact massively with their interfaces without an official API? The balance of power is still unclear.
The ultimate step = autonomous payment. The day an agent can, on its own, complete an online order or approve a transfer in a management tool, we will tip into a new era where AI no longer just assists the human, but truly acts in their place (banned in Europe for now).
It is still imperfect, often costly, sometimes unstable. But to quote a16z:
Agents are not just an evolution of the interface… they are economic actors in the making.
And this is only the beginning.
