What if software AI Agents could move around enterprise applications, execute actions and transfer data between apps without hard coding automation scripts?
The Browser is turning into an interface where AI can read screens, decide and take action for the user. With knowledge workers spending over 56% of their workday in the browser, according to Omdia research reported by Google Cloud, this shift could have far-reaching consequences.
This article examines how computer-use systems work, how they could change enterprise application design, where businesses may see practical value, and why security, reliability, and operating costs will determine how far the AI browser runtime advances by 2028.
The Shift from Document Viewer to AI Sandbox
For many years, automation has required an analysis of the underlying web-page structure. Old methods inspect the Document Object Model (DOM), find elements using selectors and take specific actions. It works efficiently if the interface is unchanged; however, if the button is redesigned or some unexpected popup appears, it will cause issues for automation and somebody should fix the script. With the AI browser runtime, there is a new method of operation that allows the system to perceive and interact with the web-page depending on its current state.
We need three core skills to do this: perception, reasoning and action. First, the VLM looks at the screenshot and makes sense of what it sees (buttons, forms, tables, and anything else on the interface). Then the model makes a decision and picks the next thing to do. Next, we perform the action on the browser (click, scroll down, type something into a field, etc.) and check if it worked or not. By using this, automation is less reliant on unreliable HTML selectors but is less powerful visually.
Progress in computer-use models offers a clearer picture of what is becoming possible. According to Anthropic, Claude Sonnet 4.6 attained an 72.5% first try success rate for OSWorld-Verified, calculated across five runs, in its test of real-world computing tasks. This is certainly a step forward, but it does not necessarily promise the same success rate when deployed in enterprise applications and business flows.
What follows relies not only on capabilities. Reduced costs for inference, faster reaction times, and greater integration within browsers may make it possible to engage in multiple interactions. Once all these elements come together, the runtime for AI within the browser could evolve into a functional interface layer between AI and enterprise applications.
Also Read: AI Browsers vs. Traditional RPA: Where Does Computer Use Fit?
Core Changes to Enterprise Application Design
The rise of the AI browser runtime will put pressure on enterprise software teams to rethink how applications are built, connected, and secured. Interfaces designed around human navigation may still function, but they could become harder for agents to operate reliably. Three changes are likely to shape the next generation of enterprise applications.
1. Designing for Agents as Well as People
Complicated navigation, poorly designed labels, and controls requiring visual recognition can lead to unreliable automation. Designers should pay more attention to well-designed labels, consistent layout design, semantic HTML, and accessibility guidelines like WAI-ARIA. These techniques will help assistive devices understand how to navigate interface elements and will give useful information for AI agents too.
This doesn’t imply that hierarchical menus will cease to exist at all. Applications of an enterprise level frequently require complex navigation due to complicated business processes. The more radical change will be the effort to make it easy to recognize actions and verify their results. AI browser runtime will work more effectively if it can differentiate between similar controls, recognize forms, and check if the operation was performed successfully.
2. The Rise of Bring Your Own Agent
Traditionally, enterprise software used the use of API, native connectors, and custom integration to facilitate the exchange of data between the systems. This is still a relevant approach, particularly for transaction volumes, but the cost of connecting each application is substantial.
Alternatively, the Bring Your Own Agent (BYOA) method could be considered. The agent of the user’s AI browser would be able to transfer the data from the CRM system to the ERP platform and the supplier’s portal by communicating with their interface. This could be a useful option for automation of processes within applications that lack appropriate APIs or for which integration efforts cannot be justified.
From the perspective of SaaS providers, this introduces a design problem. They would have to provide the applications that would allow interaction with external agents without compromising data safety and integrity.
3. Rethinking Identity and Session Security
An agent that can navigate a browser also needs carefully managed access. It may encounter sign-in screens, multi-factor authentication, CAPTCHAs, expiring sessions, and permissions that were designed around human users. Giving an agent unrestricted access to a logged-in browser would create obvious security risks.
OpenAI’s addition of computer use to its Agents API illustrates one approach. Tasks are performed by agents on the OpenAI-hosted browser, whereas web-access permissions and log-ins take place through the application. This ensures a separation between an agent’s capacity to do work and its permissions to enter a website.
Similar controls will be required for enterprise teams in terms of identity, minimum privilege access, session time, and sensitive task permission. The AI browser engine can fit into the enterprise architecture only when it is possible to know what the agent did, what it changed, and what tasks it performed.
Industry Use Cases and Realizing ROI
The most compelling reason to use AI as a browser runtime has to do with workflows with repetitive screen-based activities, disconnected applications, and outdated APIs.
Purchasing and supply chain personnel can use agents to browse through supplier websites, compare prices, ensure availability, and generate purchase orders. Less time will be spent gathering information from different sites and more on reviewing the supplier terms and purchasing decisions.
Data migration and legacy systems offer another opportunity. For the older software lacking APIs, agents can use the existing interface to collect record details and then move the same to newer applications. This might lead to reduced cost and effort for the company; however, validation and audit trail are necessary to avoid mistakes.
The customer support agents can explore the account dashboards to identify billing problems, change customer details, and perform multi-step processes. This might help them save time while allowing them to deal with complicated situations.
Microsoft’s computer-use ability allows agents to interact with web pages and desktop programs via screen interaction, making them perform actions including data entry, invoicing, and data collection in case APIs are not available.
ROI will rely on how accurate the process is, exception handling ability, and the cost involved in the supervision by humans. The companies must start from repetitive and measurable actions.
The Challenges and Roadblocks Ahead
The AI browser runtime isn’t ready for prime time at scale either. Latency is one challenge. Visual models need to read a screen, figure out an action and verify the response, iterating this process. Slower response times, and the possible need to run more pages or form data, will lead to greater compute costs. Fast models and improved session management can reduce these costs, but costs will need to be balanced with savings.
The user can inadvertently click on a wrong button, enter incorrect data, or accidentally select ‘Delete’ in place of ‘Save.’ Although such human controls are vital to create beforehand in preparation for wider deployment, they become even more essential at the point of adoption especially when a lot of money is involved or any other business critical data.
Session setup also affects efficiency. AWS reports that reusable browser profiles can reduce setup time from minutes to tens of seconds for enterprise customers running large volumes of automated browser sessions. By preserving authentication state between sessions, these profiles reduce repeated setup work.
In conclusion, however, wider acceptance will only be achieved through demonstrable reliability, restricted accessibility, and reasonable operational cost. Before intelligent agents can successfully manage exceptions, their first application should be within bounded processes where errors can be detected and rectified.
Preparing Enterprise Software for the AI Browser Runtime
In 2028, AI browsers may transform how enterprise applications are designed, built, and operated Applications will need to be developed in a manner that allows agents to move between interfaces, execute workflows, and communicate across systems and with humans.
Technology leaders should start by auditing their existing roster of web applications. Make sure the site has semantic HTML, accessible attributes, explicit interactions, and secure workflows for agents. Set up permissions, approval and audit points before handing over important work to the agents. Enterprises will win those who can create their software agent-friendly while retaining safety and security.
