The Big Bold Enterprise AI Failure

Hundreds of billions of dollars have poured into AI infrastructure, tools, and talent over the past three years. 

Whether that investment delivers real value or becomes the next cautionary tale depends on one thing: whether enterprises can generate a meaningful return on their AI investment. 

So far, the early numbers are not encouraging. At least, as per McKinsey’s The State of AI 2025 survey, AI implementation isn’t translating into AI-driven revenue and outcomes. 

Across 1,993 respondents in 105 countries. 88% report regular AI use in at least one business function. Yet, only 39% attribute any EBIT impact to AI, and most report contributions of under 5%. Fewer than 6% qualify as high performers.

The Failure Patterns

“It’s a failure of imagination, and it’s a failure of execution,” Leslie Joseph, principal analyst at Forrester, tells AIM.

Joseph outlines four distinct modes of AI adoption in enterprises.

The first is automation, where AI replaces human labour in functions such as call centres, data entry, and first-line support. The second is augmentation, where copilot-style tools assist employees and accelerate existing tasks. 

Then come enrichment and reimagination, stages where AI enables capabilities that were previously impractical and allows organisations to redesign how they operate. This includes new ways of operation, creating new decision models, revenue streams, and new forms of customer engagement that couldn’t exist before.

Joseph argues that most enterprises remain stuck in the first two stages of AI adoption—automation and augmentation. 

He notes that many companies claim they want to reduce full-time employees and cut costs, but they often haven’t invested the necessary time or effort to truly understand how much of their work can be automated, to what degree, and what dependencies, changes, and transformations are required to achieve it.

Joseph characterises these as immature organisations swept up in the excitement around generative AI, assuming they could simply replace people with models. In reality, many of these initiatives have since been reversed.

“We’ve seen that trend roll back because they did all those mistakes,” He says.

There are real-world examples of this retrenchment. 

Last year, Swedish fintech giant Klarna laid off roughly 700 customer support and related staff after automating parts of its operations with AI tools. As customer satisfaction declined and service issues mounted, the company later acknowledged the move had gone too far and began rehiring human workers to stabilise service quality. 

Last August, Commonwealth Bank of Australia eliminated around 45 customer service roles after introducing an AI voice bot to automate calls. That decision was also reversed, with the bank apologising to affected employees and reinstating roles after discovering that call volumes and handling complexity had not fallen as expected.

Joseph says these failures tend to share the same root cause. Companies rushed deployments, failed to map AI to the right processes, didn’t redesign workflows, and neglected to assess whether their data environments were suitable for AI systems to perform reliably.

Besides, Joseph explains that companies able to scale generative AI are typically those that already operate machine-learning systems in production. 

These organisations invested early in fundamentals, developed operational expertise over time, and built engineering teams capable of deploying models reliably in real-world environments.

By contrast, companies without that foundation often calculate ROI using what Joseph calls “very emotional, primitive” methods—reducing impact to simplistic time-saved arithmetic, multiplying estimated hours saved per employee and treating that figure as business value.

“If you thought implementing copilot would suddenly make your people 50% more productive, and you failed at it, I'm sorry, it’s on you,” says Joseph, suggesting that the quality of SaaS products, foundational AI models and tools isn’t the problem. 

The Codebase Problem

Software development should be AI's proving ground. It’s the domain where the tools are most advanced, the metrics are easiest to track, and the potential is the highest. 

AI-powered coding platforms and products have collectively raised billions in capital, with companies like Cursor raising over $3 billion and achieving unprecedented revenue growth. 

Yet, it’s also where the gap between promise and reality is becoming increasingly visible.

Amar Goel, founder of PubMatic and CEO of Bito AI, hears the same story from engineering leaders across India's largest tech companies: developers may be writing code faster, and everything else becomes a bottleneck.

“I was meeting last night with one of the most famous tech companies in India. They said it’s taking an average of six days for code review, before it gets merged into production,” he tells AIM. 

Six days in a world where code can be generated in minutes is not what the AI hype promises.

For prototypes and greenfield applications, the tools can be remarkably effective. But production-grade sofware is different, and enterprise codebases aren’t blank canvases. 

They’re accumulated complexity spanning years of decisions, edge cases, and institutional knowledge baked into every system. 

Goel explains that AI models see the code that is currently written, but they don’t automatically understand the organisational history embedded in it. “They don’t have the full context and understanding.” 

This involves examining why a service exists, what failed approach it replaced, and which edge cases it handles—details that aren’t obvious from simply reading the implementation.

This requires enterprises to invest in and integrate specialised tools, workflows, and pipelines that compensate for the blind spots of AI models, going beyond basic AI-enabled command-line interface tools. 

“I do actually think a lot of ROI is being achieved,” Goel says. “But the expectations that have been sold by the model companies—probably ourselves included—are that you’re gonna get 100x productivity improvement.”

A quick glance at social media, or a conversation with any working developer, makes it clear that AI coding tools are genuinely impressive. At an individual level, the productivity gains often feel real.

But Goel highlights a common misunderstanding: faster coding does not automatically translate into faster delivery. In one company he spoke to, AI reduced pure coding time by 40%. 

On its own, that sounds significant. In practice, however, coding accounted for only four days in a roughly two-month journey from PRD to production.

Even after the improvement, the overall delivery timeline moved by less than two days. 

At the individual level, the economics still made sense. 

Since coding accounts for only about half of a developer’s workload, a 40% boost in coding speed translated into an overall productivity gain of roughly 20%, enough to justify the cost of the tools.

What it did not deliver was transformation. Coding is only one part of the software development lifecycle. The bigger delays often sit in unclear requirements, architectural decisions, cross-team coordination, testing, and deployment. 

Change Management

Goel also points to change management and employee adoption as an overlooked constraint in enterprise AI rollouts. 

Using AI tools effectively requires shifts in how people work, how teams communicate intent, and how organisations define expectations. Simply deploying a tool is rarely enough.

He describes one customer piloting Bito’s AI-powered code-review agent and tracking how often developers accept its suggestions. 

About half the team engaged heavily with the tool. “But there were a third of the developers who had accepted zero suggestions,” Goel says. “They were basically just ignoring it.”

In many cases, teams had not clearly communicated why the tool was being introduced, how it fit into existing workflows, or what developers were expected to do with it.

“When you roll out ChatGPT Enterprise, have you shown your employees some example use cases and provided some training? You can just give everyone a login. A lot of people aren’t that interested in trying out new things—some are a little hesitant,” said Goel. 

The Discipline Requirement

These bottlenecks, unless properly addressed, pose bigger problems in service delivery environments where revenue depends on tight deadlines and there’s little room for experimentation and exploration. 

Sudipta Chandra, associate vice president of technology at Calsoft, told AIM, “AI-assisted coding has become a strong productivity layer for our teams—especially for boilerplate code, unit tests, documentation, and quick prototyping. For well-defined engineering tasks, it helps developers move faster without compromising quality.”

But AI output gets no special trust. “We treat AI-generated output the same way we treat a junior engineer's contribution—it must go through review, testing, and validation.”

“The biggest risks come from over-reliance and silent errors. The biggest gains come from speeding up repeatable tasks so teams can focus more on architecture and correctness… Our approach is to use AI as an accelerator inside a disciplined engineering workflow, not a shortcut.”

That mindset demands from engineers both curiosity and restraint. “Engineers must be willing to work with AI tools, understand their failure modes, and question outputs when something feels off.”

AI is delivering localised efficiency gains, but those gains often fail to compound into meaningful enterprise outcomes.

“[Because] it's never about buying the technology,” Joseph says. “It's all about the cultural transformation that you do around it.”