Bespoke AI and the race to specialize

Artificial intelligence has spent the past several years getting bigger. OpenAI, Anthropic, Google and other developers have competed to build frontier models, highly advanced AI systems trained on enormous collections of text, code, images and other data and capable of working across a wide range of tasks.

Alongside these general-purpose systems, another approach to AI is gaining ground. Companies are buying or adapting models for particular industries, departments and processes, connecting them to proprietary data and designing them around narrower sets of tasks.

The change is occurring as AI itself becomes commonplace in business. Some 88% of organizations surveyed reported using AI in at least one business function in 2025, compared with 78% in 2024, according to Stanford University’s 2026 AI Index. Seventy percent reported using generative AI.

As access to powerful general-purpose models spreads, competitive advantage may increasingly come from what companies can make those models uniquely capable of doing.

One giant model or many specialists?

A frontier model is built for breadth. The same underlying system can summarize a report, write computer code, and identify objects in an image because it has been trained to recognize patterns across vast quantities of information.

Its breadth also creates a limitation for businesses. A publicly available model has not been trained around the knowledge and practices of a particular company, but existing models can be adapted or connected to company-specific information without training another frontier model from scratch.

An analysis by IBM, the U.S. technology company, identifies three main approaches for incorporating proprietary data into generative AI. The simplest is prompt engineering, in which proprietary information is supplied directly with a request, such as attaching a call-center transcript and asking the model to summarize it. Retrieval augmented generation, or RAG, connects the model to a private database, allowing a customer-service assistant, for example, to consult company documentation before answering a question.

Fine-tuning trains the model further on additional data. IBM gives the example of an insurer fine-tuning a smaller model to process claims using the terminology, classifications and procedures required for the job.

The economics favor such adaptation because companies can begin with intelligence that someone else has already spent heavily to develop, then concentrate their own investment on the information and tasks particular to their business.

Where companies are spending

Money flowing into business AI suggests that specialization is already becoming a substantial commercial market.

Menlo Ventures, a U.S. technology venture capital firm, estimated in its 2025 State of Generative AI in the Enterprise report that companies spent $37 billion on generative AI that year, up from $11.5 billion in 2024. More than half of that spending, $19 billion, went to applications used by businesses and employees rather than underlying models and computing infrastructure.

Within it, specialization is already visible. Companies spent an estimated $7.3 billion on departmental AI designed for particular jobs such as software development, marketing, customer support and human resources. Another $3.5 billion went to vertical AI designed for particular industries. Healthcare accounted for about $1.5 billion of vertical spending, while legal AI reached $650 million.

Specialization does not mean building everything internally. Menlo found that 76% of corporate AI use cases were purchased from outside vendors in 2025, up from 53% in 2024, suggesting companies are combining external AI products with their own information, expertise and processes.

From answering questions to doing work

The difference becomes more consequential when AI moves from generating information to acting inside a company.

Deloitte’s January 2026 State of AI in the Enterprise report found that 23% said their organizations were already using agentic AI at least moderately. Within two years, 74% expected to reach that level, while 85% expected their organizations to customize agents for their own business requirements.

AI agents carry out sequences of actions rather than responding to one prompt at a time. Deloitte describes an airline connecting agents to its reservation and baggage systems to help passengers rebook flights and reroute luggage, tasks a general-purpose chatbot could discuss but not independently complete.

The distinction illustrates where customization can add value. The underlying model supplies general capabilities, while the airline supplies the data, software access and operating rules needed to complete the task. Two companies could therefore use the same model while building substantially different systems around it.

Building those connections remains difficult. Deloitte found that only 25% of surveyed organizations had moved 40% or more of their AI experiments into production, while data privacy and security was the most frequently cited concern, identified by 73% of respondents.

A different opportunity for the Middle East

The spread of customizable AI offers a different route into the technology for economies unlikely to compete with the enormous capital expenditure of the leading frontier laboratories. 

Rather than building a rival to the largest global models, companies can concentrate resources on adapting existing systems to languages, regulations, industries and business practices that general-purpose models may not understand in sufficient detail.

AI adoption is already substantial in parts of the Middle East. Stanford’s 2026 AI Index found that population-level AI usage reached 64% in the United Arab Emirates in the second half of 2025, the highest rate among the economies measured. Qatar reached 38.3% and Jordan 27.1%.

None of this guarantees that smaller models will displace frontier systems. Many specialized applications will continue to depend on powerful models developed by OpenAI, Anthropic, Google and their competitors. Instead, a division of labor is emerging in which the frontier model supplies general intelligence while companies and specialist developers determine what information it can access, what tasks it performs, and how it fits into a particular workplace.