Op-Ed: Massive Promise of MCP Meets High-Stakes Threat of 'Context Rot'

In an industry drowning in three-letter acronyms, a new one has arrived: MCP, or Model Context Protocol. Travel suppliers and intermediaries have been scrambling to release MCP servers while travel managers are wondering 'what does this mean for me?'

Model Context Protocol has incredible promise, but it also has a flaw: poorly designed servers that trigger a hidden performance killer known as 'context rot.' As with all emerging technology, distinguishing the functional from the fluff is critical.

The 'USB-C' of Artificial Intelligence

Remember the nightmare of charging your electronics 15 years ago? Every phone, camera, and laptop required a unique, proprietary cable.

Until recently, connecting an AI agent to corporate travel systems was exactly like that. MCP changes everything.

Popularized as an open standard, it acts as the USB-C cable for AI. A travel supplier can package their data and tools so that any modern Large Language Model—whether built by Google, Anthropic, or OpenAI—can instantly plug in and understand how to use it.

Suddenly, the dream of a seamless AI travel agent that can check a flight, verify a corporate rate, log an expense and check a safety alert in seconds seems within reach.

It’s Not Just a Protocol, It’s a 'Server'

Let's clear up a massive technical misconception. In marketing pitches, tech partners will throw around 'MCP' as if it is just a set of rules and standards—a 'protocol' like NDC or QR codes. But in the real world, you don't just 'connect' to a protocol. You connect to an MCP server.

The word 'server' is the most important part of this name because an MCP server is not just a passive rulebook; it is an active compute resource. To use another analogy: A protocol is a blueprint for a car; the server is the actual, physical engine.

An MCP server is a piece of software running on a computer somewhere. It doesn't just sit there; it has to do heavy lifting. It runs code, translates complex database queries on the fly, and actively executes commands. Because it is active infrastructure, it requires hosting, maintenance and security.

And because it is a physical compute resource, it can experience physical failures. It can suffer from latency, run into connection timeouts, crash under heavy search loads or even become a single point of failure for hackers looking to exploit multiple customer API keys.

When a travel supplier says they have built an MCP, they haven't just written a document—they have spun up live, active digital machinery that your AI agent now relies on to survive.

The High Stakes of Corporate Travel

In our industry, "seamless" isn't enough. It has to be flawless. If a retail AI assistant recommends the wrong shoes, the customer processes a return. If a corporate travel AI assistant makes an error, a business traveler can miss a multi-million-dollar merger meeting.

Worse, if an AI lags or misinterprets data during a geopolitical crisis or a severe weather event, a company fails its fundamental legal and moral duty of care to its employees.

For booking and safety, accuracy and timeliness are non-negotiable prerequisites. And that is exactly where poorly engineered MCP servers pose a silent, catastrophic threat.

The Hidden Trap: Memory and 'Context Rot'

To understand why these integrations can fail, you need to look at how an AI 'thinks,' using two simple concepts: tokens and context windows.

Tokens are the basic syllables or building blocks of text that an AI reads. The context window, meanwhile, is the AI’s short-term working memory. Picture it as a physical desk. When you ask an AI a question, it lays out all the relevant information on this desk to formulate an answer.

Tech companies love to brag about the massive size of their digital desks (their context windows). But here is the catch: AIs suffer from a practical memory limit of about 50 percent to 60 percent before their reasoning starts to break down.

If you clutter the desk with too much paperwork, the AI experiences 'context rot.' Its attention becomes diluted. The crucial instruction you gave it—like 'Book the negotiated corporate rate only'—becomes a needle in a massive haystack of data.

When context rot sets in, the AI becomes sluggish and highly expensive to run. It suffers from the 'lost in the middle' effect, completely missing vital details right in front of it. Ultimately, it begins to hallucinate—inventing non-existent flights, picking the wrong hotels, or failing to execute the task entirely.

How Bad Servers Cause the Rot

How does an MCP server trigger this disaster? Through lazy engineering.

Many technology providers build 'thin wrappers.' Instead of building a highly sophisticated, streamlined tool, they take their massive, messy raw databases and dump the entire technical dictionary into the MCP server.

When your AI agent connects to one of these unoptimized servers, the server forces the AI to read its entire instruction manual upfront, before the user even asks a question.

The math is alarming. A single unoptimized MCP server can instantly devour 10,000 to 17,000+ tokens just explaining what its tools do.

A travel program that connects just three of these lazy servers—say, one for flight bookings, one for hotel inventory, and one for traveler tracking—can naively waste 143,000 tokens.

That means 72 percent of the AI's working memory is completely filled with structural jargon before your employee even types: "I need to go to Chicago tomorrow."

In stress tests, when an AI is overwhelmed by this kind of tool clutter, its accuracy in selecting the right tool drops to a horrifying 13.62 percent. In a corporate travel setting, an overwhelmed AI could easily confuse a "routine weather lookup tool" with a "critical emergency broadcast tool," failing to alert a traveler to danger because its memory desk was buried in junk data.

Treat Version 1.0 With Caution

Travel managers should be thinking "so how do you protect your program from this?" But the brutal truth is that an MCP server is incredibly difficult to evaluate without hands-on, live testing.

Anyone will know the exact moment you connect your AI to a badly written server: your system will suddenly lag, costs will spike, and the AI will start giving bizarre, sluggish responses. The problem, of course, is that by the time a buyer plugs it in and see sit fail, it is already too late. The contract is signed, the budget is spent, and your system is compromised.

Buyers and suppliers alike should be conscious that, like traditional enterprise software, version 1.0 could well be a bloated, unoptimized minefield. It takes years of real-world mistakes, architectural rewrites, and rigorous performance tuning to build an efficient server that doesn't cause context rot.

Before signing on the dotted line, buyers should hand their prospective vendor this three-question technical audit:

• Is your MCP server a 'thin wrapper' that exposes hundreds of low-level database endpoints, or has it been distilled into high-level, intelligent tools?
• What is the baseline token footprint required just to load your tool schemas into our LLM? Will it eat 10 percent of our memory, or 70 percent?
• Does your architecture support selective tool activation, or does it dump your entire system vocabulary into our AI's active memory all at once?

The Bottom Line

The Model Context Protocol is an incredible milestone that will undoubtedly shape the future of corporate travel automation. But treating it as a magic bullet is a recipe for operational failure.

Corporate travel leaders must pivot from simply asking, "Can our AI connect to this vendor's data?" to asking, "How efficiently is that connection engineered?" Meanwhile, marketers should refrain from hyping up immature releases.

In an industry where traveler safety and bottom-line accuracy are non-negotiable priorities, keeping your AI’s digital workspace clean isn’t just a tech issue—it’s a business imperative.