What are GenAI and Enterprise AI?

Generative AI, or GenAI, is a subset of artificial intelligence (AI) and machine learning technologies that can create original content, such as text, images, audio, or software code, in response to a user’s prompt or request. The most common examples include ChatGPT, Google Gemini, and Adobe Firefly.

Enterprise AI is the application of AI technologies such as natural language processing, GenAI, computer vision, or automation across an entire organisation to generate tangible business value, enhance decision-making, and increase efficiency. Rather than focusing on specific, narrow tasks, enterprise AI is designed to work at a large scale across an entire organisation and integrate with existing systems and workflows. 

The Productivity Paradox and the GenAI Divide

Despite an estimated $30 to $40 billion invested into enterprise GenAI pilots and initiatives in just the last two years, 95% of organisations have seen no measurable return, according to MIT’s report on ‘The GenAI Divide: State of AI in Business 2025’. Meanwhile, only 5% of integrated pilots are producing real financial impact, such as reduced outsourcing spend, improved customer conversion, or workflow automation that sticks. This emerging divide is now being referred to as the GenAI Divide.

This divide also speaks to a productivity paradox wherein the gap between large investments in GenAI and the expected, significant productivity gains that are not yet fully realised. One of the reasons is that consumer-facing AI tools are designed for individual productivity with real, tangible, and easy-to-feel benefits such as:

• Faster research 
• Cleaner writing
• Quick ideation
• Assisted creativity 

However, these benefits do not directly improve a company’s profit and loss statement. Individual productivity does not necessarily scale into organizational efficiency. As a result, companies see widespread usage and high user satisfaction, but almost zero business impact. 

Are we using the wrong metrics to measure AI’s business impact?

ROI is a financial metric that measures the profitability of an investment. It’s calculated as the percentage of profit obtained for each currency (e.g., $ or £) invested. However, the ‘profit’ obtained can mean different things to different organisations, such as a reduction in headcount, increased revenue, or a direct cut in operational expenses. 

In addition, concluding that an investment in AI had ‘zero business impact’ is a strong phrase. Some organisations operate in regulated industries (e.g., healthcare or finance) where the impact of AI is measured through risk reduction, compliance accuracy, or improved patient outcomes rather than profit. Others may have invested in projects with intangible or long-tail value (e.g., sustainability monitoring or R&D acceleration) where quantifying returns is equally challenging.

David Gallacher (Industry Fellow and Faculty Director at UC Berkeley) argues that the MIT report—and many AI critics—are measuring AI success with the wrong metrics entirely. Firstly, there may be a timing issue: Value may accrue over longer cycles than the timeframe captured in the study. 

More importantly, Gallacher asserts, the problem isn’t that investing in AI doesn’t work; it’s that we’re applying industrial-era metrics (profit and loss) to a cognitive-era transformation, which creates a critical measurement gap: Organisations don’t understand how to quantify AI’s value creation in knowledge work environments, therefore their ROI calculations fail to capture it. The examples he offered include:

• A customer service team that handles 20% more complex inquiries without additional staff, because an AI tool handles routine questions
• Engineers who explore more design alternatives with no additional cost to the client, because AI accelerates prototyping 
• Analysts who provide more extensive reports because AI expands their research capabilities 
• Managers who make better decisions because AI enhances their access to relevant information.

The above scenarios demonstrate that AI delivers real business value (productivity, quality, satisfaction, innovation) and a return on efficiency or execution (ROE)—but not always immediate profit. This distinction underpins and adds much-needed nuance to the GenAI Divide: success depends on understanding where and how AI creates value, not just measuring short-term profits.

From ROI to ROE: What the 5% are getting right 

If the issue isn’t the technology itself, but rather how we define and measure its value, then the GenAI Divide becomes less about who has access to advanced models and more about who understands where to apply and how to execute them effectively. 

The 5% of high-performing organisations highlighted in the MIT report share several defining traits:

1. Targeted Scope: The focus is on a well-defined process rather than everything at once.

They select a clear business process — often a back-office or customer-facing workflow — where AI can automate, augment, or accelerate specific tasks. Instead of chasing enterprise-wide transformation, they aim for measurable depth over breadth.

2. Process Alignment: Tools are integrated into existing workflows

Successful deployments don’t sit in silos. AI systems are integrated into CRM systems, design workflows, or customer support platforms, allowing the technology to become an integral part of the process, rather than an external add-on.

3. Continuous Learning

Unlike static chatbots or one-off copilots, these systems capture feedback, adapt to user context, and improve continuously. MIT notes this as the key to sustained business impact.

4. They measure outcomes, not outputs.

The winners focus on tracking metrics that drive business impact, such as improved ticket-resolution time, increased design throughput, or higher customer retention, rather than limiting measurement to prompt accuracy or model latency.

5. Hollistically manage overall cost and risk.

They account for compute requirements, data governance, and environmental costs from the outset, ensuring scalability while mitigating operational and reputational risks.

6. Vendor/partner strategy

Many successful enterprise AI deployments result from partnerships with specialist vendors or the use of targeted solutions, rather than attempting to build everything in-house. For teams exploring how to design learning systems responsibly, expert guidance at the outset can prevent costly missteps. Schedule a free consultation with an expert on our team. 

By relying on these principles, the “5%” shows that AI value comes from intelligently designing feedback loops between humans, processes, and machines. Unlike flashy demos that appear polished in the boardroom but often fail in real-world use, the successful few build adaptable systems that learn from friction, retain context, and integrate deeply into workflows. 

The financial and environmental costs of learning-capable AI

The idea that AI systems should learn and remember context sounds simple enough. In principle, it’s what separates static automation from genuine intelligence. But in practice, learning comes with its own costs, both financial and environmental.

Financially, building AI that learns (rather than remains a fixed productivity tool) requires:

• Ongoing data-ingestion pipelines, feedback loops, and retraining cycles
• Integration with existing enterprise systems (ERPs, CRMs, operational workflows)
• Monitoring, governance, model drift management, and lifecycle management — all of which add cost beyond the “license the model” budget. 

From an environmental/resource perspective, the challenge is scale. Training or fine-tuning large language models (LLMs) demands significant computational power, and with it, energy: training a single large AI model can emit as much carbon as five cars over their entire lifetimes. More recent analyses by the International Energy Agency suggest that data centres supporting AI workloads could consume up to 10 times more electricity than traditional cloud systems by 2030.

Factoring in the financial and environmental implications of learning-capable AI is important for ROI because:

1. If you build a GenAI system that learns, you incur higher upfront and ongoing costs. If the business value hasn’t been rigorously defined and embedded, you risk a major cost centre with no return.
2. The environmental/performance risks (higher cost of ownership, potential regulatory/ESG problems) can amplify this risk: projects may be delayed, scaled back, or abandoned — contributing to the 95% no-ROI statistic.
3. From a product-developer viewpoint: if you promise “learning-capable AI” but don’t factor in integration and lifecycle cost (financial + environmental), you might be overselling.

Mitigating AI’s costs at enterprise level 

The financial and environmental costs of AI are significant, but at an enterprise level, most organisations crossing the GenAI Divide aren’t building or training foundation models from scratch. They’re building on top of existing ones — leveraging OpenAI, Anthropic, or Meta models and then creating lightweight, domain-specific layers.

These smaller systems don’t require massive compute resources. They use techniques like:

• Retrieval-augmented generation (RAG) to access relevant company data on demand instead of retraining models.
• Low-rank adaptation (LoRA) and parameter-efficient fine-tuning (PEFT) to adjust models with minimal computational overhead.
• Feedback loops that capture human corrections to refine future outputs — a process that’s computationally modest but operationally transformative.

This is what makes learning-capable systems more efficient than they seem. The additional compute required is often marginal compared to the value of the data they generate — data that continually improves decision-making, accuracy, and customer experience.

In other words, the real cost of “learning AI” isn’t in the electricity bill, but in the design discipline.

It takes careful planning to build systems that learn efficiently, ethically, and with a clear understanding of the environmental footprint of every layer added.

Ironically, the bigger sustainability issue may come not from advanced, adaptive AI systems — but from ineffective deployments that fail to produce results. When organisations implement static tools that can’t improve with feedback, employees end up redoing tasks, maintaining duplicate workflows, and burning more energy and time across the board.

So, is the AI bubble about to burst?

If by “bubble” we mean the hype — the belief that AI would rewrite every business model overnight — then yes, that phase is already deflating. The market is sobering up, and we seem to have gotten past the gold rush moment when most startup pitch decks included “powered by AI”.

But if by “bubble” we mean an overvalued technology with no lasting utility, then no — this is not another dot-com moment. The correction we’re seeing is not the collapse of an illusion. It’s the maturing of a toolset that is beginning to find its proper scale, rhythm, and relevance.

The MIT report reveals a hard truth: most companies aren’t failing because AI doesn’t work — they’re failing because they don’t know how to make it learn. They invest in tools, not systems; in experiments, not integrations. They chase automation, not adaptation.

Final Thoughts: GenAI is a process partner, not a product

Those on the right side of the GenAI Divide are taking a different approach. They treat AI less as a product and more as a process partner — something to be shaped, trained, and taught over time. They understand that the goal isn’t to replace human intelligence, but to extend it.

This reframing changes everything:

• For investors, it means shifting expectations from quick ROI to measured learning returns — systems that quietly compound value as they evolve.
• For product developers, this means designing AI features that integrate, listen, and adapt — tools that improve because of the people who use them, not in spite of them.
• For businesses, it means recognising that transformation doesn’t come from deploying AI, but from teaching it.

The GenAI Divide will narrow, not because the technology improves, but because our understanding of what to build and how to measure it finally does. If you’re looking to deploy enterprise AI and want to ensure that you end up on the right side of the divide, we’re here to help – please get in touch.