In-house development as a cost trap: Why most companies are completely misguided in their approach to AI and are saving money in the wrong place

The 80/20 rule for AI: Those who ignore this strategy are jeopardizing the future of their company

The era of expensive but useless AI experiments is over. While billions are being poured worldwide into building in-house artificial intelligence, a recent study by the Massachusetts Institute of Technology (MIT) reveals a stark truth: 95 percent of these pilot projects fail miserably to generate real business value. Instead of optimizing processes, they degenerate into endless and extremely costly "science projects." This painful realization is currently resulting in an unprecedented shift in the enterprise market. The new, unavoidable motto is: buy instead of build. Instead of tying up scarce developer resources in proprietary systems that are already obsolete by the time they're finished, pioneers are now relying on the so-called 80/20 rule and modular platform approaches. This analysis reveals why conventional “one size fits all” software is obsolete, why customized AI services – such as those from the up-and-coming startup Unframe AI – are revolutionizing the market, and which strategic decisions will determine success or failure in global competition by 2026.

Anyone still relying on in-house development in the age of AI is not only burning money, but also their future

The question of whether companies should develop their AI solutions in-house or purchase them from specialized providers is among the most pressing strategic decisions of 2026. While billions are flowing into generative AI, a widely cited study by the Massachusetts Institute of Technology (MIT) found that a staggering 95 percent of all AI pilot projects in companies fail to generate measurable business value. At the same time, current market data reveals a dramatic shift: within just one year, the ratio of in-house development to outsourcing of AI solutions has almost reversed. It is in this dynamic environment that companies like the Israeli-German startup Unframe AI are positioning themselves with a radically new business model that fundamentally challenges the traditional rules of enterprise software.

The following analysis examines the economic, technological, and strategic dimensions of the build-versus-buy debate, drawing on recent market data from Menlo Ventures, Gartner, McKinsey, and MIT, and places the findings in the context of a real company operating in the midst of this transformation process.

A market in flux: 37 billion dollars and an inconvenient truth

The numbers speak for themselves. According to Menlo Ventures' third annual report on the state of generative AI in enterprises, organizations worldwide spent approximately $37 billion on generative AI in 2025, a threefold increase from $11.5 billion the previous year. This means that generative AI already represents six percent of the entire global software market – a rate of market penetration unprecedented in the history of the software industry. At least ten AI products now generate annual recurring revenue exceeding one billion dollars, and more than fifty have surpassed the $100 million mark.

But behind these impressive aggregate figures lies a far more nuanced reality. Gartner forecasts global AI spending of $2.52 trillion for 2026, a 44 percent increase over the previous year. However, Gartner explicitly places the AI ​​industry in the so-called Trough of Disillusionment for 2026 and warns that AI will, in most cases, be sold to companies through existing software vendors, not as part of bold moonshot projects. Improved predictability of the return on investment must first materialize before AI can truly scale, according to Gartner analyst John-David Lovelock.

The gap between investment volume and actual value creation is the central contradiction of the current AI boom. Companies are investing at a record pace, but the majority of these investments are wasted on experiments, pilot projects, and proofs of concept that never reach production readiness. This raises the fundamental strategic question: Is it wiser to develop AI solutions in-house or to purchase them?

The major turnaround: Why companies are massively ceasing to build their own AI

Perhaps the most striking finding of 2025 is the complete reversal of the build-versus-buy ratio for AI solutions. According to Menlo Ventures, 76 percent of all AI use cases in companies are now covered by purchased solutions, with only 24 percent being developed internally. As recently as 2024, the ratio was almost 50:50, with 47 percent developed in-house and 53 percent purchased. Within just twelve months, the market has thus shifted radically.

This shift is no accident, but the result of painful experiences. S&P Global Market Intelligence found in a survey of more than 1,000 companies in North America and Europe that 42 percent of companies will have abandoned the majority of their AI initiatives by 2025—a dramatic increase from just 17 percent in 2024. On average, 46 percent of all AI feasibility studies were discontinued before reaching production readiness. The RAND Corporation confirms that over 80 percent of all AI projects fail—twice as many as non-AI technology projects.

The reasons for the failure of internal development projects are multifaceted. McKinsey reports that around 85 percent of all AI proof-of-concepts never progress beyond the pilot phase. An analysis by the Boston Consulting Group of 1,000 executives from 59 countries found that only 26 percent of companies have even developed the capability to move beyond the proof-of-concept stage, and a mere four percent consistently generate significant AI value. Gartner analysts go so far as to predict that by 2027, over 40 percent of agent-based AI projects will be abandoned due to escalating costs, unclear business value, or insufficient risk controls.

Against this backdrop, the massive shift towards outsourcing appears as a rational market response to a wave of failures. The message from corporate buyers is clear: speed to value creation trumps perfect customization. Purchased AI solutions reach production readiness significantly faster and boast a conversion rate almost twice as high as that of traditional software. According to Menlo Ventures, 47 percent of purchased AI deals make it into production.

The MIT study and the failure of enterprise AI: An anatomical examination

The MIT NANDA study, "The GenAI Divide: State of AI in Business 2025," led by Aditya Challapally at the MIT Media Lab, has become the most cited reference on the structural failure of AI projects in businesses. The study is based on 150 interviews with executives, a survey of 350 employees, and an analysis of 300 public AI deployments. Its findings paint a stark picture of failure: 80 percent of organizations explore AI tools, 60 percent evaluate enterprise solutions, 20 percent launch pilot projects, but only five percent reach production with measurable business impact.

The study's key finding is remarkable because it refutes common excuses. The problem isn't the quality of AI models, inadequate infrastructure, or primarily regulatory hurdles. The real bottleneck is what the MIT researchers call the "learning gap": enterprise systems that don't adapt, don't store feedback, and don't integrate into workflows. Generic tools like ChatGPT work brilliantly for individual users because they are flexible. In enterprise contexts, however, they become static academic projects that neither learn from the context nor improve over time.

Another finding of the study is particularly revealing: Purchasing AI tools from specialized providers and building partnerships succeeds in approximately 67 percent of cases, while in-house development is only successful about a third as often. This finding is especially relevant for the financial sector and other highly regulated industries, where many companies were still attempting to build proprietary generative AI systems internally in 2025. The MIT data suggests that companies fail far more frequently when going it alone.

Another systematic error concerns the misallocation of resources. More than half of the budgets for generative AI flow into sales and marketing tools, while the MIT study identifies the highest ROI in back-office automation—that is, in eliminating business process outsourcing, reducing external agency costs, and streamlining processes. Companies are therefore not only investing incorrectly in the type of implementation, but often also in the wrong application areas.

The 80/20 Rule of Enterprise AI: A New Strategic Paradigm

From the convergence of various data sources and industry analyses, a strategic paradigm is increasingly emerging, which can be described as the 80/20 rule of enterprise AI. Industry observers and data from analysts such as Gartner and Deloitte suggest that most companies should pursue a hybrid approach: 80 percent of AI requirements are covered by purchased or subscription-based solutions, while 20 percent are addressed by custom-developed in-house solutions where deep integration or unique intellectual property is crucial.

This 80/20 split is also reflected in practice. Use cases that are ideally suited for procurement include IT ticketing systems, knowledge-based search functions, marketing content generation, data extraction from unstructured documents, and standardized reporting solutions. In-house development remains sensible where there are concerns regarding intellectual property or where the AI ​​solution represents a strategic differentiator, such as in core banking systems, proprietary trading algorithms, or business-critical decision models.

The economic logic behind this division is compelling. Outsourcing offers faster time-to-value, predictable costs through subscription models, continuous innovation cycles from the provider, and the avoidance of internal development backlogs. In-house development, on the other hand, ties up scarce developer resources, creates technical debt, and carries the fundamental risk that an internally launched solution will already be technologically obsolete by the time it is completed because the underlying AI models will have evolved in the meantime.

Venture capital firm Andreessen Horowitz (a16z) confirms this trend in its analysis of 100 enterprise CIOs: Recently, there has been a significant shift from in-house development to outsourcing, as the AI ​​application ecosystem begins to mature. In particular, the dynamic performance differences between various models and the decreasing costs make it increasingly sensible to outsource the continuous evaluation and optimization for each use case to a dedicated AI application team at an external provider, rather than handling it internally.

The end of one size fits all: Why standardized software is obsolete

For decades, traditional enterprise software followed a simple principle: one product for all. Standardized solutions were designed to serve the largest possible audience with the same range of functions. This paradigm is under massive pressure in the age of AI. The formula has changed: "One Size Fits All" is becoming "One Size Fits None.".

This shift has profound economic causes. Companies increasingly have diverse requirements that generalized solutions can no longer meet. The growing complexity of business processes, the heterogeneity of IT landscapes, and the rising expectations of users who are accustomed to a personalized experience from their private use of ChatGPT and similar tools make tailored approaches essential.

AI-powered personalization enables software platforms to adapt in real time to each user's behavior, preferences, and specific business challenges. The marginal cost of personalization decreases dramatically through AI-driven code generation, refactoring, and testing—not to zero, but low enough to fundamentally rethink the software delivery business model. This opens up models where each customer, upon registration, receives a logically isolated, cloud-based version of the software precisely tailored to their specific needs.

In parallel, pricing models are changing. Outcome-based pricing is increasingly replacing the traditional license or seat-based model. Gartner predicted that by 2025, over 30 percent of enterprise SaaS solutions would integrate outcome-based components, compared to around 15 percent in 2022. Bessemer Venture Partners describes in its current Pricing Playbook how AI-native companies are largely abandoning seat-based SaaS pricing in favor of usage-, output-, and outcome-based models that directly link revenue to measurable results. Examples such as Intercom, with $0.99 per resolved request, or Salesforce, with $2 per conversation, illustrate the direction this is heading.

The modular principle: How modular AI platforms are conquering the market

A key architectural paradigm gaining traction in the enterprise AI segment is the modular approach, often described as a Lego-like building block principle. The basic idea is that instead of building monolithic, rigid AI systems, solutions are assembled from reusable, interchangeable building blocks that can be flexibly combined and replaced as needed.

This principle offers three crucial advantages: First, the flexibility to add and replace components as better technologies become available. Second, the ability to update AI tools without rebuilding the entire infrastructure. Third, the speed at which value can be created while maintaining adaptability. In an industry where the underlying models evolve weekly, this flexibility is not a nice bonus, but an essential necessity.

The practical implementation of this principle can be illustrated using the example of data extraction. An initial module is being developed for processing commercial lease agreements, i.e., complex documents of 80 to 90 pages. This module is designed to be so generic that it can be used with minimal adjustments for financial reports in Excel, resumes, or image-based use cases. Each new module expands the library and is immediately available to subsequent customers. This principle of scalable reusability is the economic core of the platform model: The marginal costs of each additional implementation decrease dramatically, while quality increases through the growing body of experience.

In practice, a modular AI architecture also means that different Foundation models can be used for different tasks—for example, GPT for logical reasoning, Gemini for architectural tasks, and Claude for precision work—without affecting the overall solution. This LLM agnosticism is another key differentiator compared to in-house development, which is typically tied to a specific model and incurs significant migration effort with each model change.

Here you will learn how your company can implement customized AI solutions quickly, securely and without high entry barriers.

A managed AI platform is your all-inclusive, worry-free solution for artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a ready-made solution tailored to your needs from a specialized partner – often within just a few days.

The key advantages at a glance:

⚡ Rapid implementation: From idea to ready-to-use application in days, not months. We deliver practical solutions that create immediate added value.

🔒 Maximum data security: Your sensitive data stays with you. We guarantee secure and compliant processing without sharing data with third parties.

💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.

🎯 Focus on your core business: Concentrate on what you do best. We take care of the entire technical implementation, operation, and maintenance of your AI solution.

📈 Future-proof & scalable: Your AI grows with you. We ensure continuous optimization and scalability, and flexibly adapt the models to new requirements.

More information here:

• Managed AI Platform

Unframe AI: A case study of the new enterprise AI business model

The Israeli-German startup Unframe AI provides an instructive case study for the practical implementation of the described market trends. The company was founded in April 2024 by Shay Levi, Larissa Schneider, and Adi Azarya. Levi had previously co-founded Noname Security and, as CTO, transformed it into the first unicorn in the API cybersecurity sector before its sale to Akamai for approximately $500 million. Schneider brings over a decade of experience in the enterprise tech sector, including leadership positions at Nutanix and Noname Security, coupled with an academic background from Hult International Business School in San Francisco.

In April 2025, Unframe emerged from stealth mode with a total funding round of $50 million, divided into a $20 million seed round and a $30 million Series A round led by Bessemer Venture Partners. Other investors included TLV Partners, Craft Ventures, Third Point Ventures, SentinelOne Ventures, Cerca Partners, and Terra Nova Ventures. In less than a year, the company achieved millions in annual recurring revenue (ARR) and acquired dozens of large enterprise customers worldwide, including Cushman & Wakefield and Nomura.

What Unframe apart from many competitors is its business model. The platform is based on the so-called Blueprint approach, a methodology that provides large language models with the necessary context to generate domain-specific results without requiring extensive model training or fine-tuning. The company is LLM-agnostic, meaning customers can switch between different public and private models without being locked into a specific ecosystem. Pricing is per person per year in tiers (Small, Medium, Large, Extra Large), with all customization services and the work of the AI ​​product leaders included in the subscription – no hidden costs or additional fees.

Perhaps the most radical aspect of the business model is the principle of results-oriented payment: customers only pay when they see real impact. In an industry where 95 percent of AI projects fail, this is a bold promise that can only work if the implementations actually create value. According to the company, the lead time from the initial consultation to a production-ready, fully customized solution is typically days, rather than the months or years that are standard in the industry.

1,670 use cases and no end in sight: The reality of AI demand in large companies

The scale of the challenge facing large corporations in AI implementation can be illustrated by a concrete example. A senior AI executive at one of the three largest investment banks on Wall Street reported a backlog of 1,670 AI use cases that had been brought to her department by operations and needed to be implemented by the end of 2026. This executive's assessment was unequivocal: even with unlimited internal development resources, it would be impossible to handle this volume internally. What was needed was a scalable approach.

This example is by no means an outlier. JPMorgan Chase now operates over 1,000 AI use cases in production, spread across risk management, marketing, fraud detection, and customer service. Bank of America has earmarked $4 billion of its $13 billion technology budget for AI by 2025. Citigroup has piloted agent-based AI for 5,000 employees and launched a company-wide initiative to systematically integrate AI into all its processes. These figures illustrate that the demand for AI implementations in large enterprises far exceeds available internal capacity.

McKinsey data shows that while 88 percent of organizations are using AI in at least one business function, only seven percent have scaled AI company-wide. The vast majority are in an intermediate stage between experimenting (32 percent), piloting (30 percent), and scaling (31 percent). The gap between what companies want to do with AI and what they can actually implement is the biggest bottleneck in the current AI transformation.

In this context, it becomes clear why hybrid models, which combine the advantages of in-house development (adaptability, control) with the benefits of outsourcing (speed, scalability, lower maintenance burden), are gaining importance. Partnering with a specialized platform provider allows companies to systematically address the exponentially growing backlog of AI use cases without overwhelming internal teams.

The Governance Paradox: When AI Agents Get Out of Control

Besides the economic aspects of the build-versus-buy decision, there is a frequently underestimated dimension: governance. This topic is gaining particular importance with the rise of agent-based AI systems – that is, AI agents that not only provide information but can also autonomously execute actions within enterprise systems.

A vivid example from the insurance industry illustrates the problem. The IT manager of a large insurance company on the US West Coast was confronted by his executives with the demand to build AI agents, without a clear definition of their intended use. The idea of ​​simply providing business units with a tool to independently create AI agents carries significant risks: Hundreds of thousands of unmaintained AI agents performing autonomous actions within a company in a highly regulated industry represent a governance nightmare.

Regulatory requirements further exacerbate this problem. The EU AI Act, in force since August 2024, introduces increasing obligations for high-risk AI systems by 2026/2027, including conformity assessment, CE marking, and transparency requirements for general AI models. Singapore's framework for agent-based AI requires the definition of the so-called action space (which tools and systems an agent may use) as well as clear limits of autonomy with human oversight. The NIST AI Risk Management Framework offers a vendor-neutral structure for risk controls, which is increasingly being adopted by US companies.

The governance dimension has significant implications for the build-versus-buy decision. Companies developing AI in-house must independently build and maintain the complete governance infrastructure: lifecycle gates, recertification cycles, model maps, red team testing, post-market monitoring, and incident workflows. Specialized platform providers can centrally address these governance requirements and offer them as part of their standard solution, significantly reducing the workload for individual customers. In an era where regulatory requirements for AI systems are growing exponentially, governance expertise is becoming a crucial competitive advantage for platform providers.

KPIs or flying blind: What distinguishes successful AI projects from failed ones?

The data is clear: The decisive success factor for AI projects is not the technology itself, but rather the definition of clear success criteria before launch. The MIT study identifies the lack of alignment between technology and business processes as the primary cause of failure. Companies have attempted to force generative AI into existing processes with minimal adjustments, instead of first defining the desired business impact and strictly aligning the implementation accordingly.

According to current best practices, a multidimensional KPI framework for AI projects comprises six dimensions: business impact (revenue growth, cost reduction), operational efficiency (process speed, error reduction), risk mitigation (compliance, fraud prevention), strategic value (market position, innovation capacity), economic efficiency (cost per result) and adoption rate (user acceptance, penetration).

Practical implementation is what separates winners from losers. Successful companies define concrete, measurable goals before a project begins – for example, 96 percent accuracy with a response completeness rate of over 90 percent. They establish benchmarks against which to compare and create transparency about exactly what success looks like before the first line of code is written.

In contrast, most companies fail to answer the vague question: "What can we actually do with AI?" This exploratory, unstructured approach leads to what industry experts call science projects: technically interesting demonstrations without any significant business value. The consequence is an endless cycle of experiments that never make it into production.

The implications for the build-versus-buy decision are significant. Internal development teams tend to focus on technological feasibility and consider the business impact as a secondary consideration. Specialized platform providers, on the other hand, who bill based on results, are existentially dependent on delivering business value from day one, because their business model would otherwise collapse. This structural incentive alignment is an often underestimated advantage of the buying model.

The speed advantage: Why time is the hardest currency in the AI ​​economy

In the AI ​​economy, time is the decisive competitive factor. Technological development is progressing so rapidly that an internally developed solution can already be obsolete by the time it's completed. In traditional enterprise environments, the time between the conception of an internal AI system and its production readiness typically ranges from 19 to 24 months: one to two months for needs assessment, three to four months for piloting, and further months for budget approval, vendor selection, legal and security reviews, integration, and finally, rollout.

During this period, dozens of new Foundation models appear, entire product categories emerge and disappear, and benchmark performance improves by orders of magnitude. Menlo Ventures documents that spending on code agents and AI app builders exploded from near zero to several billion dollars, as models can now interpret entire codebases and execute multi-stage tasks completely autonomously. What begins as state-of-the-art in-house development risks becoming a relic upon completion.

Specialized platform providers reduce this timeframe from months to days or weeks. They centrally absorb the complexity of constant model changes, updates, and security patches, allowing individual enterprise customers to benefit without having to allocate their own resources. This pooling of innovation speed is a classic example of economies of scale: what a single company could never manage so quickly becomes possible for many simultaneously through the platform.

Furthermore, the a16z report shows that the performance differences between various models are becoming increasingly marginal, while the cost differences remain significant. In this situation, the competitive advantage shifts from model selection to pure implementation speed and process integration – precisely to the strengths of specialized platforms.

The strategic exception: When in-house development still makes sense

Despite all the arguments in favor of outsourcing, there are clearly defined areas where developing AI solutions in-house remains strategically sound. These areas typically share one or more of the following characteristics: high relevance to the company's intellectual property, direct link to the core business as a strategic differentiator, or use cases where the AI ​​solution itself becomes a product to be sold.

A core banking system based on proprietary algorithms that represents a genuine competitive advantage in risk modeling is a classic example of sensible in-house development. Similarly, proprietary trading strategies where AI logic is central and disclosing them to an external provider poses unacceptable risks. In the pharmaceutical industry, AI-driven molecular research can be so deeply intertwined with a company's DNA that outsourcing is neither practical nor desirable.

The challenge for decision-makers, however, lies in making a brutally honest distinction between genuine strategic differentiators and the infamous "not invented here" syndrome. Many companies overestimate the strategic importance of use cases that are, in reality, merely standard functionalities. An IT ticketing system, a knowledge-based search, or the generation of marketing content typically do not fall into the category of strategic differentiation and, if developed in-house, only create a costly development backlog.

The recommendation of industry analysts is clearly converging: The 20 percent share of in-house development should be strictly limited to those areas that actually create a unique competitive advantage, while the remaining 80 percent should be covered faster, more cost-effectively and with significantly less risk by specialized platforms.

Crossing the valley of disillusionment: A look ahead to 2026 and beyond

Gartner's prediction that AI will be in the trough of disillusionment by 2026 should by no means be misinterpreted as a pessimistic signal. Rather, this stage in the hype cycle marks the healthy point where unrealistic expectations give way to reality and companies begin to understand the technology's actual strengths and limitations. It is the phase in which pure experimentation gives way to the cold calculation of return on investment.

The figures indicate that this maturation process is already well underway. Global AI spending of $2.52 trillion in 2026 and the projected increase to $3.3 trillion in 2027 demonstrate that the willingness to invest remains absolutely strong, despite disappointments with individual projects. AI is expected to account for 41.5 percent of all IT spending in 2026, and this share could rise to over 50 percent in 2027. Infrastructure investments alone will drive a 49 percent increase in spending on AI-optimized servers in 2026.

What's changing isn't the volume of investments, but their structure. Companies are becoming increasingly selective in choosing their AI projects, prioritizing proven results over speculative potential. The era of AI experimentation is giving way to the era of AI production – and this production is being bought, not built. For platform providers that demonstrably deliver measurable business value, a market of almost historic proportions is opening up. For companies still wavering between building and buying, the decision is becoming increasingly clear: In a world where speed has become the most valuable currency and 95 percent of internal AI projects fail, purchasing specialized solutions is not only the more pragmatic, but also the only economically superior strategy for the vast majority of use cases.

The winners of this transformation will be those companies that have the courage to radically focus their resources on the truly strategic 20 percent and rely on smart partners for the remaining 80 percent—partners who deliver faster, cheaper, and with a demonstrably higher success rate. The rest will remain mired in disillusionment, overtaken by their own slowness in an industry that shows no mercy to the hesitant.

I would be happy to serve as your personal advisor.

You can contact me at wolfenstein∂xpert.digital or

Just call me on +49 7348 4088 965 .

LinkedIn