Why 95% of enterprise AI projects fail and what the other 5% do differently

Sep 9

5 min read

An MIT study published in 2025 as State of AI in Business, built on 300 public implementations, 52 executive interviews and 153 survey responses, found that 95% of organizations record no measurable return on their P&L, despite enterprise investment in generative AI estimated at between 30 and 40 billion dollars. Over the same period, global corporate investment in AI reached 581.7 billion dollars during 2025, growing 130% year over year, according to Stanford's AI Index 2026.

For the 5% that make it to production, the deciding variable is approach, meaning how the problem gets defined, how the data gets organized and how the path to production gets designed before anyone signs up for the first cloud service. The foundation models available on Amazon Web Services (AWS) comfortably exceed what most enterprise use cases require, and cloud compute stopped being a limit several cycles ago. Here are the five points where the difference is made.

According to the MIT report, 80% of organizations explored generic generative AI tools and close to 40% went on to deploy them, with the impact concentrated in the individual productivity of the people using them. An assistant that saves twenty minutes per person moves the P&L only if someone defined what happens with that time. Adopters that start from a narrowly scoped use case report an average revenue increase of 15.8% and a productivity gain of 22.6%, according to a Gartner survey of 822 business leaders. What changes the outcome is having someone on the business side own the metric.

The metric before the code

97% of organizations struggle to demonstrate the business value of their generative AI initiatives, according to a 2025 Informatica study. An Economist Impact survey adds a revealing data point, 37% of executives believe their applications are production ready, and the figure drops to 29% among the people actually building them. MIT itself defines success as sustained impact on productivity or financial results, a standard that forces the metric to be agreed before the first line of code is written.

89% of organizations redesigned their data strategy because of generative AI and only 26% deployed solutions at scale, according to an IDC survey. The usual pattern is a pilot that works on a hand curated dataset and collapses once it is connected to real systems, where duplicates, incompatible schemas and nonexistent lineage all live together. On the other side, organizations with mature data infrastructure and governance achieve a 24.1% revenue improvement and 25.4% in cost savings compared with their peers. That is why the cloud data layer gets sorted out before the model layer.

F5's report on AI application strategy places 77% of companies at an intermediate readiness level, with marked gaps in governance and in security consistency across clouds. When questions about access traceability, encryption and retention surface weeks away from going live, the usual answer is a costly redesign. Among the most mature organizations, 62% increased their security budget for AI initiatives over the past year, against 16% of the least mature ones, according to IDC.

The funnel MIT describes puts the scale of the problem in order, 60% of organizations evaluate enterprise AI systems, 20% reach the pilot stage and 5% make it to production. Gartner projected that at least 30% of generative AI projects would be abandoned after the proof of concept, because of poor data quality, inadequate risk controls, rising costs or unclear business value. Designing for production from day one of the pilot means resolving observability, cost control, model versioning and failure recovery while the pilot is still running.

On AWS, an AI architecture that reaches production rests on a data layer consolidated with Amazon S3 and AWS Lake Formation, on foundation models managed through Amazon Bedrock when there is little reason to train from scratch, and on Amazon SageMaker when the use case demands custom development and deployment. Governance is implemented from the start with AWS Identity and Access Management and AWS Key Management Service. The AWS Well-Architected Framework, and its Machine Learning Lens in particular, offers a concrete review framework for catching gaps before they turn into technical debt.

How we work on this at Teracloud

As an AWS Advanced Tier Services Partner with AWS Migration Competency, at Teracloud we run these projects in three stages, in that order. We start with an assessment of the real state of the infrastructure and the data layer, which produces a map of what sources exist, what condition they are in and how much normalization work lies ahead before a model can rely on them. Then we run a Well-Architected review focused on security, cost and reliability, grounded in the AWS shared responsibility model so that it is explicit what the cloud provider controls and what stays on the client side. Only then do we design the path from pilot to production, with agreed business metrics, defined exit criteria and a cost budget that holds up as volume grows.

That order comes from practice. In the project we did with Colven, an Argentine manufacturer of fleet monitoring devices, the raw telemetry from its GESTYA platform was first consolidated into a Data Lakehouse on Amazon S3, with a daily ingestion of more than 20 million rows, and only afterward did we integrate the generative AI agent that today answers questions about that data in plain language, with row-level security so each Colven client sees only their own operation. The agent works because the layer underneath was already sorted out.

To sustain that standard in production we have added specific capabilities. As Datadog partners we bring managed monitoring and alerting, plus threat detection integrated into daily operations, to the architectures we design and run.

Our entry into Anthropic's Claude Partner Network addresses the other end of the problem, working with state of the art models through Amazon Bedrock while keeping data inside the client's own account, which is often the condition that unblocks projects in regulated sectors such as Fintech and Utilities.

One nuance that comes up often in our practice, a good share of the projects that reach us are moves between clouds, from GCP or Azure to AWS, where there is already an architecture, decisions made and accumulated debt. That inherited context tends to weigh more than the choice of model. With teams in Argentina, Chile and the rest of Latin America, and projects across Oil & Gas, Utilities, Financial Services and startups, the assessment almost always starts there.

As the debate shifts from model capability toward the quality of the cloud architecture, the competitive advantage will sit with those who did the unglamorous work of sorting out the foundation.

If your organization is evaluating an AI project on AWS, or needs to review why a pilot cannot scale, you can learn more about our approach here.

Marcos Loguercio

Marketing Analyst