Why 9 in 10 AI projects end in failure and how the Technology-Enabled Service model is changing that

Reports from RAND and MIT indicate that as many as 80–95% of AI projects fail to deliver tangible profits. The key to success lies not in better technology, but in shifting from IT tool implementation to process transformation through the Technology-Enabled Service model. We reveal how to avoid structural pitfalls and deliver measurable business outcomes instead of more “dead” pilots.

The Scale of the Problem – Numbers Impossible to Ignore 

Companies worldwide are pouring tens of billions of dollars into AI projects, yet most see no return on investment.

A 2024 RAND report indicates that the vast majority of AI initiatives fail to deliver a measurable business impact, with failure rates reaching 80% or higher. Separately, the MIT NANDA research project’s report – “The GenAI Divide: State of AI in Business 2025,” widely cited by industry media including Fortune – states that as many as 95% of GenAI pilot projects yield no measurable financial impact. 

⚠  Methodological Note: The 95% figure specifically refers to GenAI pilots and is defined as a lack of measurable financial impact rather than a “failure” in the traditional project management sense. While the methodology itself has been debated by analysts (such as Futuriom), even the more conservative RAND estimates of around 80% paint the same picture: the vast majority of AI investments do not translate into measurable business results.

Deloitte’s State of AI in the Enterprise 2026 report confirms this diagnosis: only a fraction of companies truly translate AI deployments into business model transformation. Most are merely experimenting, without achieving scalable financial results. A significant portion of IT leaders report seeing no measurable improvement in performance despite ongoing investments in digital transformation.

This is not a technology problem. It is an approach problem.

Why Do AI Projects Really Fail?

The most common answers you hear are: “the wrong model,” “not enough data,” or “too expensive.” All of these are wrong – or at least incomplete.

MIT NANDA points to what it calls the “learning gap”: most deployed AI systems do not retain context, do not adapt to company processes, and do not improve over time. Tools like ChatGPT or Copilot increase individual employee productivity, but they fail to integrate with operational processes at a level that translates into a company’s bottom-line financial results.

Anonymous IT Director, MIT NANDA 2025: “We’ve seen dozens of presentations and demos over the year. Maybe one or two turned out to be actually useful. The rest were just wrappers on existing models or academic projects.”

The report also highlights a structural flaw in deployment strategy: organizations concentrate their AI budgets on highly visible areas – marketing, sales, and customer service – while real ROI is also generated in the back-office: automation of reporting, contract analysis, and internal support systems. These areas are less flashy but deliver measurable savings.

Four Structural Errors Killing AI Projects 

There is no single reason why AI projects fail. However, there are four recurring patterns that appear regardless of the industry or organization size.

1. Treating AI as an IT Project, Not a Business Transformation “Let’s build a chatbot,” “Let’s deploy Copilot,” “Let’s integrate GPT-4 into our system”—these are sentences heard at the start of many AI projects that end up in a drawer a year later. The problem is that AI requires process redesign, not just software installation. Organizations that overlook cultural change, role adjustments, and workflow shifts end up with a tool that no one uses – or uses differently than intended.

2. Building Instead of Buying – The Path Selection Error MIT NANDA indicates clearly: projects implemented in deep partnership with external AI specialists have a significantly higher success rate than internal projects built from scratch. Despite this, most companies interviewed by the MIT research team tried to build their own tools. Why? The belief that a custom solution provides more control. In practice, this often leads to the opposite: blurred ownership, revealing significant skill gaps in teams lacking experience in scaling AI, and a systematic underestimation of both financial and time costs. As a result, projects drag on, consume more resources than planned, and rarely deliver real business value.

✓ MIT NANDA 2025 Conclusion: Projects carried out with an external AI partner have a significantly higher success rate than projects built exclusively in-house – regardless of industry or organization scale.

3. Investing in Front-Office, Ignoring Back-Office Where do companies put their AI budgets? Mostly in customer-facing areas: chatbots, marketing content generation, personalization, or prospecting tools. The catch is that the ROI from automating reporting, contract analysis, or HR processes is statistically higher and faster to achieve. MIT explicitly points to this “investment bias” as one of the four structural factors of the divide.

4. Pilots as a Goal, Not a Starting Point Gartner predicts that over 40% of agentic AI projects will be canceled by 2027 due to unclear business value and rising costs. Why? Because many organizations treat a pilot as proof that “we are doing AI.” Once the pilot ends, the project dies. There is a lack of a mechanism to transition from experiment to production: a scaling plan, system integration, and real-time monitoring. 

What Do Successful AI Projects Have in Common?

The MIT report identifies four characteristics of projects that actually generate business value:

• Deep Process Integration – The AI tool is not a standalone app but an integral part of procedures and systems. Data flows automatically between systems, and tools communicate in real-time.
• Ability to Learn and Adapt – Systems adapt to the company context (leveraging internal knowledge via RAG), remember previous decisions, and improve based on user feedback. This is the opposite of a static chatbot that starts from zero every time.
• Evaluation via Business Outcomes, Not Technical Metrics – Projects are measured by time savings, cost reduction, or retention increases – not just how well the model performs technically.
• Partnership Over Building Solo – Purchasing a specialized solution from a partner with deployment experience significantly increases the chances of success compared to building from scratch.

The classic software sales model looks like this: a company purchases a license, gains access to the platform, and what they do next with the tool is entirely up to them. If the implementation fails, the blame is placed on the client for “not knowing how to use it.”

Software house implementations exacerbate this problem: they take months or years, consume massive budgets (often in the hundreds of thousands of dollars), and the software house bears no responsibility for the business outcome – their role ends with delivering the code, without any guarantee of ROI or scalability in real-world conditions.

This model worked well for simple SaaS tools or individual IT deployments. For AI projects, however, it fails structurally. This is because AI is not a tool you simply install. It is a transformation you conduct.

In contrast, the Technology-Enabled Service model used by Extentum AI combines two elements that are typically separated in traditional approaches:

• Proprietary Tech Stack – A custom GenAI platform designed specifically for AI implementations within the business processes of large enterprises. Not just a generic language model, but an environment that enables building, deploying, and managing AI agents within the context of a specific organization.
• Forward Deployed Engineers (FDE) – A team of highly skilled engineers who don’t just configure the tool, but understand the client’s business processes, design the implementation architecture, and ensure the system operates and scales after launch.

✓ This approach directly addresses what MIT identifies as the source of success: deep process integration + partnership with an external specialist = a significantly higher project success rate.

Why a GenAI Platform Alone Isn’t Enough?

Buying access to a GenAI platform is easy today. Dozens of providers offer language model interfaces, AI agent builders, and ready-made connectors. Why, then, do projects still fail to deliver a measurable business impact – even when a company chooses a high-quality product?

The answer lies in what happens after the purchase.

A GenAI platform provides possibilities, but it is the engineering teams who are responsible for turning those possibilities into real business processes. In practice, this involves four key areas:

• Process Mapping – Analyzing which processes are truly suitable for AI automation and which are not. Not every business problem is an AI problem. An experienced AI engineer knows when a simple script is enough, when an AI assistant is needed, and when a full AI agent with an autonomous workflow is required.
• Integration with Existing Systems – Without connecting to ERP, CRM, or company document databases, a GenAI platform answers questions based on general knowledge, not your organization’s data. RAG (Retrieval-Augmented Generation) – a technique that combines a language model with the company’s own knowledge base – is a prerequisite for a meaningful enterprise deployment.
• Post-Launch Management and Oversight – AgentOps (managing AI agents in production: monitoring, debugging, version control) is a discipline in its own right. A deployed AI agent must be monitored, updated, and managed by someone who knows what to track and how to react to anomalies.
• Security and AI Governance from Day One – On-premises deployment, data anonymization mechanisms, and the auditability of agent actions are not options for an enterprise – they are requirements. A good AI engineer designs these mechanisms from the start, rather than adding them as a “plugin” at the end.

According to industry analysis (Futuriom / CRN, January 2026): Platform engineering is becoming the critical factor in enabling the transition from AI experiments to operating systems. It provides the structure, governance, and integrated workflows necessary for AI to function as a digital teammate rather than a detached tool.

How to Choose an AI Transformation Partner – What Should You Look For?

The market is saturated with companies offering “AI implementation,” but the differences between them are fundamental.

Extentum AI operates precisely in the Technology-Enabled Service model, combining its proprietary tech stack (a visual Agent Builder with RAG, including graph databases, a user portal, and SDK) with a dedicated deployment team (FDE). This team is not an external integrator; it consists of engineers working directly on their own product. The client is never left alone with the technology. This approach eliminates communication errors, accelerates delivery time, radically reduces implementation costs, and significantly minimizes the risk of failure.

Questions worth asking a potential partner:

• Do you have your own GenAI platform and proprietary tech stack, or do you build exclusively on third-party technology?
• What does the path from pilot to production look like – what are the milestones and success criteria?
• Can I deploy the system on-premises? How do data anonymization and auditability work?
• Who manages the system after launch, and what does real-time monitoring of agents look like?
• How will you solve the problem of Shadow AI and hallucinations within my organization?

⚠ If any of these questions cause your partner embarrassment or confusion – keep looking.

FAQ – Frequently Asked Questions 

Where does the “9 out of 10 AI projects fail” figure come from?

It is a simplification based on several independent sources. RAND estimates that the vast majority of AI projects do not yield a measurable business impact – with failure rates reaching approximately 80%. The MIT NANDA research initiative report (“The GenAI Divide,” July 2025), widely cited by outlets like Fortune, indicates that up to 95% of GenAI pilots do not produce a measurable financial effect-though the methodology itself has been a subject of debate among industry analysts. However, the consensus is clear: most AI projects do not deliver the expected ROI.

Does this mean AI doesn’t work?

No. It means that most organizations implement AI in a way that structurally makes success impossible. Projects that combine deep process integration, external partnerships, and evaluation based on business outcomes do work and generate tangible value.

Which AI projects have the highest chance of success?

According to MIT NANDA: back-office projects (automation of reporting, contract analysis, internal HR helpdesks) that are implemented with an external partner and measured by specific business KPIs such as cost reduction, decreased processing time, and customer retention. These are less flashy than front-end chatbots but significantly more effective.

How does Technology-Enabled Service differ from a standard SaaS or software house implementation?

In the SaaS model, you get access to a tool; what you do with it is up to you. Similarly, a software house delivers a finished product and moves on. In the Technology-Enabled Service model, you receive the tool along with a team that guides you through the entire journey: from process analysis, through build and integration, to managing the system in production. It is the difference between buying building materials and commissioning a construction project with a guarantee.

Is Shadow AI a real threat to my company?

Absolutely. MIT NANDA indicates that a large portion of employees use private AI tools (like ChatGPT) for work tasks – often without the IT department’s knowledge. This means sensitive corporate data may be processed on external servers without any control. Implementing a proprietary AI assistant with proper AI governance mechanisms radically mitigates this problem.

How should the success of an AI project be measured?

Three measurable criteria: (1) Process handling time – how long the same task takes before and after AI implementation; (2) Unit cost – the cost of handling a single query, document, or ticket; (3) User retention – whether the team is actually still using the system 3 months after launch. Projects measured by these indicators from the start have a significantly higher rate of remaining in production.

Summary

AI project statistics are unpleasant to read – but they are vital to know before making your next investment decision.

The majority of AI projects fail to deliver a measurable business impact, and it is rarely due to poor technology. They fail because organizations treat them as IT projects rather than process transformations. Because they build in isolation instead of partnering with a firm that specializes in AI. And because they measure success by the number of pilots rather than the impact on the bottom line.

The Technology-Enabled Service model is the answer to these structural errors – not as a mere promise, but as a deployment architecture. A GenAI platform without engineers is a tool without context. Engineers without their own platform offer a service without scalability. Together, they form a complete solution capable of moving from pilot to production and staying there for the long term.

If you are planning an AI project and want to see which approach makes sense in your context – book a call with us. We would be happy to walk through your processes and tell you directly where AI can bring real, measurable value. 

Sources

1. MIT Project NANDA, „The GenAI Divide: State of AI in Business 2025”, July 2025 (cited by Fortune, among others) – https://fortune.com/2025/08/18/mit–report–95–percent–generative–ai–pilots–at–companies–failing–cfo/ 

2. RAND Corporation, AI Project Failure Rates, 2024 (via Super Biznes / Industry Media) – https://superbiz.se.pl/wiadomosci/zabka–zarabia–na–ai–a–inni–traca–ekspert–radzi–jak–uniknac–milionowych–strat–aa–1N72–XVVk–TRRb.html 

3. Deloitte, State of AI in the Enterprise 2026, 2025 – https://www.deloitte.com/us/en/what–we–do/capabilities/applied–artificial–intelligence/content/state–of–generative–ai–in–enterprise.html 

4. Gartner, „Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027”, June 2025 – https://www.gartner.com/en/newsroom/press–releases/2025–06–25–gartner–predicts–over–40–percent–of–agentic–ai–projects–will–be–canceled–by–end–of–2027 

5. Futuriom / CRN, „AI Platform Engineering Will Be Key to Deploying AI in 2026”, January 2026 – https://crn.pl/aktualnosci/inzynierowie–platform–beda–kluczowi–dla–wdrazania–ai–w–2026–roku/ 

6. Futuriom, Methodological criticism MIT NANDA, August 2025 – https://www.futuriom.com/articles/news/why–we–dont–believe–mit–nandas–werid–ai–study/2025/08