2025 was supposed to be the year of the AI agent. Companies invested heavily in autonomous systems that could handle tasks without human intervention, only for agents to end up simulating the people they were replacing. The promise was efficiency at scale, but the infrastructure just wasn’t ready.
An agent booking a hotel today opens a browser, types search queries, clicks through menus, and fills out forms. This is browser automation, and it has become the inefficient workaround that makes agents functional in a world of legacy systems. The technical challenges were foreseeable but didn’t make it into the pitch decks, and in the meantime, companies are left waiting on a return on investment.
We are in a bridge period where agents can technically perform actions, but the systems they interact with were never built to support them. Understanding why this gap exists, and what will force it to close, is important for any company making infrastructure decisions now.
THE OLD MODEL CAN’T LAST
An agent navigating a hotel website doesn’t make a direct request for availability and pricing. It burns compute power to simulate keystrokes and mouse clicks, while processing visual layouts designed for human eyes and executing workflows built for someone physically clicking through pages.
Not only is this approach inefficient, but browser automation in general also has reliability problems. WebArena, a benchmark for testing AI web agents, reports that leading models complete just 35.8% of real-world web tasks successfully. At scale, this combination becomes expensive. Economic pressure will eventually force better solutions, as we saw with how systems request data.
Most APIs used to operate like ordering off a fixed menu—you got all the data whether you needed it or not. If an advertising company queried Facebook for a specific user identifier, the API returned everything at that endpoint, even if they only needed a small subset. That created unnecessary internet traffic, database queries, and infrastructure costs.
So the industry created GraphQL, an entirely new language for API requests where you could specify exactly what data you needed and got nothing else. The efficiency gains were significant enough that GraphQL became widely adopted, and I expect a similar trajectory with AI agents interacting with businesses.
AGENTS STEP INTO THE BREACH
Hospitality offers a clear case study for how agent-native infrastructure will need to evolve. In the future, hotel booking systems will have APIs designed specifically for agent interaction, offering direct data access through standardized protocols.
Instead of opening a browser and simulating clicking through pages, the agent will make a structured request to a webhook or API endpoint, receive availability and pricing data in a format it can immediately process, and then complete the transaction.
This shift requires solving a fundamental authentication problem on both ends. Hotels need to distinguish legitimate booking requests from bots scraping their data. And users need confidence that their agents won’t complete transactions with sketchy providers or pay inflated rates. Until that trust is established, there will probably be a little bit of fear on both sides.
The solution will involve associating agent requests with verified humans. Instead of a CAPTCHA designed to block bots, there will need to be some form of AI-native authentication that confirms an actual person authorized the interaction. The technical infrastructure will include public documentation where businesses essentially say: Point your agent to this URL.
Beyond authentication, though, rebuilding the infrastructure will require governance mechanisms that don’t exist in browser automation: audit trails, rollback capabilities, and accountability frameworks. It will be costly, but the efficiency gains from eliminating browser automation overhead will make it worthwhile.
THE BROADER TRANSITION
This transition will affect companies’ entire digital structure. For the last 20 years, businesses have relied on a plethora of SaaS products to function. Each solved a narrow problem, and each introduced licensing costs, integration burdens, and additional management overhead.
But as AI agents become commonplace, they will carry out the same processes that previously required several SaaS subscriptions, making these clunky and rigid products difficult to justify. That creates structural risk for many SaaS companies as agents change the unit economics of enterprise software.
Meanwhile, the companies building for agent-native interaction will need to rethink data access policies.
At the moment, many publications block LLM crawlers, platforms restrict automated interactions, and enterprises want agents running on their own infrastructure rather than handed over to Amazon, OpenAI, or Google. These barriers exist for legitimate reasons, but as companies adapt their systems, they will need to distinguish between unwanted scraping and legitimate business use cases.
The bridge period is temporary. Infrastructure always recalibrates to the most efficient pattern, and cost pressure will determine the timeline.
CREATE A STRUCTURAL ADVANTAGE
Once companies realize how much compute is being burned to simulate keystrokes and mouse movements, the pressure to clean up the architecture becomes unavoidable. But as IBM has noted, most organizations aren’t agent-ready on the inside, either. Their proprietary data isn’t structured for agent access, their internal systems lack the exposed APIs that agents need, and their workflows weren’t designed with automated decision making in mind.
The companies doing this preparation now will have a structural advantage. The rest will be left trying to retrofit their legacy workflows while the market shifts under them.
Jito Chadha is CEO at Nventr.