How Companies in the Middle East Can Get Payback from Gen AI Investments

News

1. AI Research Forum and Summit Focused on Agentic AI Announced

2. Cisco's New Quantum Entanglement Chip Aims to Accelerate Distributed Computing

3. 80% of Middle East Cyberattacks Result in Data Breaches, Warns New Study

4. Global leaders convene in Abu Dhabi for inaugural Governance of Emerging Technologies Summit

5. Gulf Nations Fast-Track AI Ambitions, UAE Leads Regional Readiness, says BCG Report

6. AI Professionals form a Redefined Workforce. But Systemic Roadblocks Persist, Survey Finds

In January, when DeepSeek, a Chinese AI startup, released its R1, claiming it had achieved its generative AI (Gen AI) large language model (LLM) for just $6 million, the billions being spent by AI market leaders, including Microsoft-funded OpenAI, came under scrutiny.

Although DeepSeek’s cost analysis remains dogged by skepticism, some question whether the AI spending levels are too high.

As expectations for AI development rise, companies pour money into AI development, and researchers push the ability of LLM building, companies investing significantly in Gen AI are recalibrating, and many are still looking for returns on that spending. 

Current Limitations of LLMs

Experts say the key is identifying LLM use cases where new efficiencies outweigh the tools’ cost and risk. First, they suggest starting by examining the current limitations of large language models (LLMs). 

That’s a critical point. 

“Enterprise business problems can only be solved by the right governed decisioning, which is explainable, traceable, controlled, and risk mitigated,” says Celal Kavuklu, Director of Customer Advisory for Middle East, Turkey, and Africa, at SAS. “LLMs are powerful, but do not solve the business problem independently. And they’re not flawless.”

While LLMs are incredibly powerful, a deeper understanding of LLMs’ fundamental limitations can help organizations address their shortcomings with complementary technologies and human governance.

Bias: LLMs are trained on vast amounts of text data from the internet, which inherently contains societal biases. “This means models can inadvertently reinforce stereotypes or provide biased outputs. These biases can have real-world consequences, particularly in sensitive applications like hiring tools or healthcare,” says Alessio Bagnaresi, Vice President – Artificial Intelligence, Middle East, Africa, and Turkey, Oracle.

Ethical Concerns: LLMs can raise ethical concerns regarding their usage. Misusing these models to generate misinformation, deepfakes, or spam is a growing concern. “Additionally, the potential to replace human workers poses ethical questions about job displacement,” says Bagnaresi. 

Hallucinations: LLMs sometimes generate responses that sound plausible but are factually incorrect or entirely made up, which is known as hallucinations. This issue stems from their reliance on patterns in data rather than a true understanding or reasoning. However, new techniques are in place to reduce hallucinations.

Troubleshooting:  Diagnosing and resolving issues with LLMs can be challenging due to their black-box nature. “It’s often unclear why a model generated a particular response, making it difficult to debug or improve specific aspects,” says Bagnaresi.

Security: The deployment of LLMs introduces unique security challenges, including adversarial attacks, where malicious inputs cause the model to behave unexpectedly. Also, LLMs can inadvertently expose sensitive or private information they were trained on, especially if the training data was not carefully curated.

For too long, Saad Toma, General Manager of IBM Middle East and Africa, says AI has been seen as a game of scale, where bigger models meant better outcomes. But the real breakthrough is as much about size as it is about efficiency.“Today, only 1% of all enterprise data sits within LLMs. Because, while LLMs carry enormous potential, their limitations become clear the moment they’re applied in real business settings.”

Toma adds that enterprises don’t just need AI that can generate content or answer questions; they need AI that can be trusted with their most critical processes, that can run cost-effectively at scale, and that adheres to their industry’s regulatory and operational boundaries. 

“They need transparent models, so they know what’s inside them. They also need flexibility to adapt those models to their proprietary data and domain expertise. They also need efficiency because scaling generative AI shouldn’t mean a significant cost burden.”

Economic Implications of Building LLMs

While LLMs offer tremendous potential, understanding their economic 

implication is crucial for businesses considering their adoption. First, building and training LLMs is expensive. It requires thousands of Graphics Processing Units, or GPUs, offering the parallel processing power needed to handle the massive datasets these models learn from. 

The cost of the GPUs alone can amount to millions of dollars. According to a technical overview of OpenAI’s GPT-3language model, each training run required at least $5 million worth of GPUs. Also, licensing costs and the talent needed to implement and manage these systems should be considered.

According to an IBM Institute for Business Value report, compute costs are expected to increase by 89% between 2023 and 2025, with 70% of executives citing Gen AI as a key driver. This cost surge is largely driven by the enormous compute power required to run these models. These resources extend beyond servers, including storage, data centers, networking, and the energy required.

“Costs are an inevitable part of the equation of any technology implementation,” says Kavuklu. “And applications vary, depending on the business need – whether it is creating content for marketing, creating content for customer experience teams, or analyzing customer service requests by processing the input text and being able to come up with the right answers with the LLMs.”

When LLMs are used in a hybrid fashion, they can be deployed to various business problems and applied to many use cases. But what is the cost of it? Kavuklu says, “We analyze where we want to invest. What problem do we want to tackle, and to understand what kind of investment it requires?”

However, the indirect costs can be even more significant: trial-and-error cycles, rework, privacy concerns, and most importantly, the cost of choosing the wrong use case. “That’s why we push for upfront assessments to avoid waste and ensure ROI,” adds Kavuklu.

Also, apart from regulatory compliance costs, which continue to escalate as governance frameworks evolve worldwide, another significant cost center involves workforce transformation initiatives, says Zakaria Haltout, AVP MEA, UiPath.”Employees require technical training on new AI tools and guidance on reimagining their workflows and decision-making processes. This organizational change management aspect often exceeds the direct technology costs, especially when considering productivity losses during transition periods.”

The build-versus-refine decision for language models also carries substantial financial implications. While proprietary LLMs offer greater customization and data security, they require massive computational resources and specialized expertise. “Hence, building LLMs in-house is typically beyond the scope of many organizations, which find that refining existing foundation models with proprietary data offers a more cost-effective approach, though this calculation varies based on use case specificity and data sensitivity,” adds Haltout.

Experts say companies that understand these cost drivers can make smarter investment decisions, balancing innovation with cost efficiency. “The challenge is not just about acquiring the right cloud capacity, but optimizing across the entire infrastructure,” says Toma. “Companies that fail to manage these operational costs may quickly see their budgets outpace their ROI.” 

Disciplined Approach Focused On Measurable Impact

To play the game, Gen AI investments must begin with a systematic approach to identifying high-impact, appropriate use cases, particularly in the enterprise, and of course, the involvement of domain experts working with AI engineers to define the business KPI to optimize and the business case.

According to Bagnaresi, it is important “to ensure that use cases, once in production, will remain in production through proper observability and adoption.”

As organizations move from experimental AI efforts to scalable, production-ready deployments, Haltout says, “From what we’ve seen so far, the FOMO-driven ‘try everything’ approach that characterized initial GenAI adoption is giving way to a more disciplined approach focused on measurable impact.”

Generative AI represents a significant technological advancement but is not a silver bullet. As a subset of AI, it must be used judiciously. This involves asking the right questions: What challenge are we trying to solve?

“This is why we emphasize a systematic approach that starts with a Gen AI maturity assessment,” says Kavuklu. “This allows organizations to evaluate their AI readiness to see where they stand— technically and strategically.”

“This helps identify the right business problems to solve. It’s about making informed, not impulsive, investments. Identifying whether it is the technology or the people’s education and training where investment should be made,” adds Kavuklu.

Focusing on targeted, high-impact use cases yields operational improvements and builds institutional knowledge. Enterprises deriving the most value are adopting an empirical approach to use case selection; they’re examining processes where people currently spend significant time on data synthesis, pattern recognition, and decision-making supported by large volumes of information. 

These areas, Haltout says, typically yield the highest returns when augmented by GenAI.

That said, he adds, pilot projects remain essential, acting as controlled experiments that allow organizations to test feasibility, refine implementation approaches, and demonstrate value before committing to full-scale rollouts. “This disciplined ‘start small, scale smart’ methodology significantly improves success rates compared to more scattered approaches.”

However, as organizations refine their approach to agentic automation, how can companies identify which processes best suit GenAI or agentic automation?

IBM, Toma says, emphasizes that AI’s value is not in experimentation alone but in driving tangible outcomes. That starts by looking at core workflows where AI can enhance productivity, reduce costs, or improve experiences—whether in customer service, IT operations, or software development.

“You need to ask: What problem am I solving? And is this where AI, especially generative AI, can provide a meaningful advantage? That’s why we built Watsonx, to give organizations the tools to identify, validate, and scale AI use cases responsibly, with transparency and governance embedded from the start, while using trusted data,” adds Toma.

In the region, IBM is working with e& to enhance their AI governance framework to promote compliance, oversight, and ethical practices across their growing AI ecosystem. In Saudi Arabia, SDAIA has introduced IBM’s watsonx.ai and the Arabic LLM ALLaM to Saudi Arabia’s DEEM Cloud, enabling government entities to deploy and scale generative AI securely. 

“In each of these cases, the common thread is intentionality. These companies are aligning AI with business priorities, governing it responsibly, and choosing use cases that generate value from day one,” he adds.

Finding Meaningful Efficiencies

Finally, it is critical for companies to assess processes and tasks to identify areas for gaining meaningful efficiencies.

Experts say finding meaningful efficiencies requires moving beyond superficial process analysis to identify the “long tail” of automation opportunities—those complex workflows that have traditionally resisted conventional automation approaches but offer significant potential value.

The key is aligning Gen AI capabilities with real business needs, says Kavuklu. He adds:

• Start with readiness assessments
• Understand your data, your infrastructure, and your goals 
• Look for high-impact areas—customer service, marketing content, internal knowledge search—where Gen AI can enhance efficiency without jeopardizing compliance or data security

According to Toma, IBM has seen how Gen AI can drive meaningful productivity improvements, realizing over $3.5 billion in productivity savings through thousands of client engagements.

IBM has identified key use cases that deliver the highest returns.

App modernization and deployment: With tools like watsonx. Code Assistant, developers can generate code 60% faster, drastically accelerating development and innovation.

Customer service:  AI-powered assistants have transformed client interactions by improving response times and customer satisfaction. For example, by reducing call wait times by up to 30% and boosting satisfaction scores by up to 40%, businesses can offer superior service with fewer resources.

Digital labor: By automating key business processes—such as HR, IT, and procurement—generative AI frees up human capital for more strategic tasks. In fact, IBM Consulting managers saved 50,000 hours last year using watsonx. Orchestrate in their promotion process. Also significantly sped up contract workflows, with watsonx.ai making them 80% faster. 

Ultimately, meaningful efficiency requires looking beyond task-level optimization to reimagine end-to-end workflows. Haltout says, “Organizations achieving the greatest gains are those willing to fundamentally reconsider how work gets done when intelligence can be embedded throughout their processes.”