DatE
July 2, 2025
Reading Time
9 Min.

When AI lies and twiddles its thumbs: How MCP changes that

API
KI-Integration

By

Andreas Siegel

Generated with ChatGPT

Imagine asking your AI assistant for the current revenue of your online store. The reply comes instantly: “Based on typical e-commerce patterns, I estimate your monthly revenue to be around €50,000.” Sounds professional — but it’s pure speculation. The AI has no access to your real data and instead invents plausible-sounding numbers.

And that’s the crux of the problem: artificial intelligence has a credibility issue. Worse still, it can write perfect emails but can’t actually send them. It can draft brilliant project plans but can’t update tools like Jira. AI is like a smooth-talking intern who sometimes lies — and never lifts a finger.

Anthropic’s Model Context Protocol (MCP) changes that fundamentally. Since its launch in November 2024, adoption has been exploding: over 5,000 active server implementations are already online, and tech giants like Microsoft, Google, and OpenAI are integrating MCP into their platforms. At Block (Square), development teams report “massive productivity gains,” while Twilio benchmarks show tasks completed 20.5% faster and API calls reduced by 19.3% thanks to MCP integration.

When AI Turns into a Storyteller

A lawyer in New York learned this lesson the hard way. He used ChatGPT for legal research and received six detailed court decisions — complete with citations, judges’ names, and docket numbers. The problem? All six rulings were entirely fabricated. The AI had written them so convincingly that even an experienced attorney was fooled. The case ended up in court, and the lawyer was sanctioned.

These so-called “hallucinations” aren’t a bug — they’re a systemic issue. AI models work as probability-based text generators, choosing at each step the statistically most likely next word. When the training data lacks certain information, the system simply invents the most plausible-sounding answer. In academic references, recent studies show that 47% of ChatGPT citations are completely made-up, while another 46% contain incorrect details drawn from real sources.

Microsoft experienced this collateral damage firsthand when its AI recommended Ottawa’s “Food Bank” as a popular tourist attraction. The embarrassing headlines came quickly — after all, it’s a charity serving people in need, not a sightseeing spot.

The root of the problem lies in how AI works: it operates solely on static training data — a frozen snapshot of knowledge as of a fixed date. Without access to current, verified sources, even the most advanced AI remains a smooth-talking but unreliable conversational partner.

MCP Brings AI Back to Reality

This is where MCP comes in — turning speculation into fact. Picture the shop manager from earlier: with MCP, Claude connects directly to the Shopify database and responds with precision: “Your current monthly revenue is exactly €73,245.67, as of 2:30 p.m. today.” No estimates, no fabrications — just real, verified data in real time.

MCP works like a “USB-C port for AI applications” — it allows any AI system to communicate with any business application without the need to develop custom integrations for each connection. This dramatically reduces the complexity of a fundamental IT problem: the so-called “N×M integration problem” becomes an “N+M problem.” Previously, every AI application needed a dedicated interface to every data source. With 10 AI tools and 20 data sources, that meant 200 separate connections — a maintenance nightmare. With MCP’s standardized client-server architecture, only 20 implementations are needed, all compatible with the same 10 AI tools.

From Advisor to Actor

The second problem with AI is even more frustrating: it can talk, but it can’t act. Today’s AI can draft the perfect email, but it can’t send it. It can create detailed project plans, but it won’t update tools like Asana or Jira. It can brilliantly analyze financial reports, but it has no access to live market data.

This limitation forces users into tedious double work: AI generates the content, but humans have to manually move it into different systems. The result is inefficient workflows with a high risk of human transcription errors.

MCP breaks through this barrier. Traditionally, if you ask ChatGPT to “send an invoice to customer@example.com,” you’ll get a well-written suggestion: “Here’s the perfect text for your invoice. Please copy and send it manually.” With MCP, Claude instead activates the email tool, generates the invoice, sends it directly, and confirms: “Invoice successfully sent to customer@example.com. Confirmation ID: #12345.”

When AI Becomes a Teammate: Real-World Success Stories

The transformation from “talking AI” to “working AI” is already visible in documented business results:

  • Block (Square), the payments company, found its developers overwhelmed by numerous APIs, each requiring separate integrations for data access. This fragmentation slowed efficiency and stifled innovation. By introducing MCP servers, Block streamlined access to critical databases and integrated them into developer tools like Replit. The results were striking: faster development cycles, greatly improved collaboration, and a 25% increase in project completion rates. Teams can now focus on innovation rather than wrestling with integration problems.
  • Codeium, an AI coding platform, struggled to give developers simultaneous access to multiple codebases, documentation, and debugging tools. MCP integration enabled its AI to dynamically tap into all needed resources. The outcome: a 30% reduction in time spent troubleshooting and hunting for resources. Developers can focus entirely on writing code, resulting in a significantly better user experience.
  • Atlassian wanted to integrate AI features into its project management tools without impacting performance. Its MCP server now enables real-time updates, project status tracking, and effective user feedback integration. The results: higher customer satisfaction scores and a 15% increase in product usage. This dynamic interaction has been a game-changer for project management workflows.
  • An e-commerce company implemented an MCP server to unify various customer support tools, enabling AI chatbots to access real-time inventory and customer data. The results speak for themselves: a 50% reduction in average response times for customer inquiries and a measurable increase in sales conversions thanks to higher customer satisfaction.
  • A leading manufacturer leveraged MCP to optimize its supply chain by connecting AI algorithms directly to supply chain management systems. This enhanced integration led to a 25% reduction in inventory costs and improved efficiency through better demand forecasting and real-time data analysis.

These examples reveal a key insight: MCP is not just an IT topic — it’s a company-wide transformation. Whether in customer service or procurement, development or logistics, MCP revolutionizes every business function. Marketing can directly launch campaigns into social media tools, finance teams receive automated reports from all systems, and sales can update CRM data in real time. This isn’t about better IT — it’s about better work for everyone in the organization.

The Silent Revolution in the IT Department

For IT leaders, MCP solves a decades-old problem. By introducing a single standard for all integrations, development costs and maintenance overhead drop significantly. Unified security and authentication protocols, centralized logging of all AI interactions, and granular access control to corporate resources suddenly become a reality.

The market dynamics are impressive. Since its launch late last year, MCP has seen extraordinary adoption that has surprised even veteran tech insiders. All major AI providers have announced integrations: OpenAI joined in March 2025, Google DeepMind in April, and Microsoft is developing official SDKs. More than 1,000 community-built MCP servers are already available — and that number grows daily.

Industry leaders are already positioning themselves strategically. In the financial sector, PayPal and other payment providers are using MCP for payment automation. Development platforms like GitHub, Cursor, and Replit have integrated it into coding assistants. Enterprise giants such as Microsoft Dynamics 365 and HubSpot are extending their CRM systems, while cloud providers like AWS and Cloudflare are leveraging MCP for infrastructure management.

The forecasts are clear: in 2025, MCP will become mainstream in development tools; by 2026, it will be the de facto standard for AI integration. Companies that act now can expect a positive ROI in as little as six months.

The Path to the Intelligent Enterprise

Companies now face a strategic choice. While software firms and tech startups are already experimenting, many traditional industries are still hesitating. Yet they can benefit from a structured approach — and the entry point is easier than it seems. Successful implementations tend to follow a proven pattern:

Phase 0: Mindset

Recognize that AI marks a paradigm shift — especially once it gains access to data and tools.

Until now, machines were known for deterministic and repeatable results: same inputs, same outputs. We learned to rely on that certainty. This changes with the rise of language models, which literally calculate their outputs based on probabilities — freshly, each time.

The same input will ideally tend toward similar output, but it’s not guaranteed. Sound familiar? It should — it’s exactly how humans have always worked together. Adopting the right mindset for AI means, in a sense, recognizing the machine as a peer and extending the same allowances: it can make mistakes, just like people do. Our goal should be to learn how to collaborate.

Phase 1: Exploration

Your entry into MCP should be gradual, as the technology is still young and evolving rapidly. The most effective approach is to begin with low-stakes experiments in areas where mistakes carry no serious consequences. Internal documentation, for example, makes an excellent playground — teams can experience how AI interacts directly with Confluence or other knowledge bases without touching production systems.

A particularly useful starting point is to experiment with existing MCP servers that fit your current toolchain. The community has already built a wide range — from simple file-system access to complex CRM integrations. This exploration phase should be deliberately playful, giving teams the freedom to discover possibilities without performance pressure. Why not make the company cafeteria menu available to AI, just for fun?

The learning effect takes center stage: How does AI behave when it can access real data? What new workflows emerge? Where do unexpected problems or opportunities arise? Tools with read-only access provide a safe on-ramp before later enabling write capabilities.

Phase 2: Focused Integration

After the initial experiences, concrete ideas for where MCP can deliver real value will likely have emerged. This phase is characterized by the targeted identification of repetitive, manual processes that can be optimized through AI integration.

At the same time, internal know-how must be systematically developed. Autonomous access to data and tools is undeniably a powerful capability — and with great power comes great responsibility.

The challenge lies not only in technical implementation, but also in organizational learning. Teams must develop new workflows and adapt to working alongside probabilistic systems.

Monitoring becomes essential — not just from a technical standpoint, but also in terms of process. How is work quality changing? Where are new sources of error emerging? How can we ensure high-quality results — meaning low error rates and precise, targeted tool use by AI systems? What governance structures are needed? How can we ensure compliance with existing policies?

Equally important at this stage is building expertise within the organization. External consultants can help accelerate the start, but the understanding of MCP-specific challenges must grow internally. This includes developing standards for security, data quality, and compliance.

Phase 3: Scaling

The scaling phase requires strategic decisions about the company’s long-term AI architecture. This is no longer about isolated experiments, but about systematically embedding MCP into business processes. GDPR-compliant solutions become mandatory once mission-critical data is involved.

Industry-specific MCP servers may come into play here, tailored precisely to your unique needs and constraints. Such an investment signals that MCP has matured from an experiment or novelty into a strategically relevant technology.

It also means moving beyond the “felt” productivity gains and early wow moments. Measurable KPIs should be established to provide a reliable foundation for further investment decisions.

The greatest challenge at this stage is often cultural, not technical: How does the role of employees change when AI systems take over a significant share of routine work? Which new skills will become essential? How can teams remain motivated when their ways of working change so fundamentally?

These are the questions we will increasingly have to face in the future …

Setting Realistic Expectations

With MCP only introduced in November 2024, it is still difficult to make reliable, broadly applicable, and quantifiable claims about the benefits of this relatively new technology. However, the examples from the Arsturn case studies reveal a clear trend — one that aligns with my own experience, even if I can’t yet quantify it precisely.

Artificial intelligence is already part of everyday work for many tasks. Without granting AI direct access to data and tools, however, there is inevitably additional overhead — for example, copying and pasting input and output into or from the AI tool. Simply eliminating this extra step through autonomous access to data and tools can, over time, lead to significant productivity gains.

The effect grows even stronger when entire workflows are integrated into a single tool (such as GitHub Copilot). For instance, moving from understanding an implementation task in a Jira ticket, reviewing related documentation in Confluence, implementing it in IntelliJ IDEA, and opening a pull request in GitHub could take as little as 30 minutes — naturally depending on the task’s complexity and the quality of the provided information and instructions.

The potential is undeniably there.

However, the success of MCP implementations depends less on the technology itself and more on organizational factors.

Executive sponsorship is fundamental — without leadership support, AI transformation initiatives can fail due to cultural resistance or lack of resources. At the same time, new AI-driven workflows require continuous employee training: teams must learn to collaborate with AI tools rather than fear them.

Incremental integration strategies have proven particularly effective, starting with non-critical pilot areas to gather experience before expanding into mission-critical systems. “Big bang” migrations often lead to overload and project failure.

The decisive factor, however, is close collaboration between IT and business units, ideally supported by change management. MCP projects are interdisciplinary transformations that must combine technical excellence with organizational readiness for change.

Companies that address these four factors systematically report significantly higher success rates and faster adoption among employees.

The long-term vision is clear: MCP as the standard for all AI integrations, with custom industry-specific servers and sustainable competitive advantages through advanced AI integration.

Our Journey Toward More AI at pentacor

Following the phased approach described earlier, we at pentacor have gradually embraced both the use of AI tools and the integration of MCP into our workflows.

After some team members gained initial hands-on experience through individual experiments and practical use, we decided to bring the entire team together for an AI-themed day. We firmly believe in the disruptive power of artificial intelligence, especially the Model Context Protocol. That’s why we want everyone — from tech experts to back-office staff — to be empowered to apply these tools effectively in their respective areas.

Selected examples from various departments served as a starting point. We wanted to make the potential of AI and MCP tangible — and it worked! Many pentacorians were able to develop ideas for their own work, and together we set up the necessary tooling and took the first steps. The goal was to lower the entry barrier and, in some cases, to reduce initial skepticism.

One thing is clear: Such disruptive changes, which can sustainably transform workflows, naturally come with fears and concerns. We gathered, discussed, and addressed these together. At the same time, we collected potential use cases and scenarios with promise for our daily work. The list we compiled was truly impressive. Manual and repetitive tasks were common across the board — and the idea of offloading these to AI sparked great interest.

It was only logical to continue the conversation in smaller groups afterward, discussing further questions or even starting to implement some ideas in practice.

To be continued ...

Assessing Risks Realistically

As with any transformative technology, MCP also presents challenges that companies need to understand and for which practical solutions must be developed.

Security and Compliance

MCP is still in an early development stage and known vulnerabilities like prompt injection attacks exist. In theory, an attacker could try to gain unauthorized access to corporate systems via manipulated inputs. Anthropic and Microsoft are aware of these issues and are actively working on solutions, with comprehensive security updates planned for Q3 2025.

Especially in regulated industries such as finance and healthcare, GDPR and compliance requirements demand extra attention when integrating AI tools with enterprise data.

Technical Implementation

Integrating MCP into existing enterprise landscapes is more complex than marketing materials might suggest. Legacy systems, which often make up 70% of corporate IT, require careful integration planning. Latency issues can arise with complex toolchains, and token limits restrict processing of large data volumes.

Organizational Challenges

The most underestimated factor is change management. Teams need upskilling to adapt to new ways of working with AI tools. Debugging and monitoring AI interactions require new skills and processes.

Pragmatic Approach

The recommended path remains optimistic: start with non-critical use cases such as internal documentation tools, enforce strict access controls, and proactively plan security updates. Begin with pilot projects in isolated environments before integrating business-critical systems.

The potential benefits justify the manageable risks, especially given how rapidly the technology is evolving. A gradual approach with continuous risk assessment is key.

The Decision Is Yours

MCP transforms the “eloquent intern who sometimes lies” into a “more trustworthy digital colleague with system access.” Companies that act now should be aware that they are at the forefront of the hype cycle. MCP as a protocol is less than a year old, widely discussed, and subject to ongoing changes.

However, those who take action today gain valuable early experience that will matter in the near future. As we’ve seen on the path to becoming an intelligent enterprise: now is the time to work on your mindset. Make friends with AI, learn its strengths and weaknesses, and follow its development closely. The insights gained are invaluable. Business processes will undergo a transformation. Actionable AI systems are our colleagues of tomorrow.

The revolution has already begun. While you’re reading this, thousands of developer teams worldwide are implementing MCP solutions. The question is not if but when companies will become part of this transformation. The technology is available, early success stories are documented, and market dynamics are clear.

What’s holding you back?