When a business starts to integrate an AI assistant into its customer service, the initial use case may appear relatively simple: summarizing conversations, answering questions based on a knowledge base and composing responses for customer service representatives. Once the business expects the same assistant to check an order in an ERP system, retrieve a customer’s complete record from a CRM system, determine whether a refund is permitted or update another system after the interaction, however, the requirements become considerably more complex.
Once a business has a model that can perform all of the required tasks well, the real work begins. The rest of the business must give the AI assistant access to relevant data, decide what functions the AI should perform and what it should not, and ensure that all surrounding systems and business processes can function correctly with the new assistant.
For most companies, business processes are not organized within one single, well connected system. Customer data for example is stored in a CRM system, financial transactions in an ERP system and inventory management in yet another system. Most business processes are organized around legacy applications that have been around for many years.
While evaluating the capabilities of individual models, there is a larger gap between the functionality offered by a single model and the rest of the technology stack that supports it in actual business processes.
From AI pilots to business processes
The scale of investment shows how quickly companies are moving in that direction. Gartner expects worldwide AI spending to reach $2.7 trillion in 2026, representing a 49.5% increase from the previous year. At the same time, another Gartner survey found that only 22% of organizations had successfully scaled AI across multiple business units or adopted an AI-first approach.
The gap suggests that spending on AI is advancing faster than many companies’ ability to blend it across the business. Building a working demonstration is one thing. Making the technology part of a process that involves a few teams, applications and decisions is another.
A similar difference between individual productivity and broader business results appears in McKinsey’s 2026 research. 80% of respondents said AI had improved their productivity, while only 37% reported a positive impact on EBIT. A company can make an employee faster without necessarily changing how the organization operates, which makes the difference important.
The snag is shown by the customer-service example. An employee may save a few minutes by using AI to recap a conversation, but if that employee still has to move information manually between the CRM, billing system and order management platform, much of the original process remains unchanged. The AI has improved one step without necessarily improving the entire process.
This is where code becomes more complicated, and companies aren’t only adding AI to existing software. More and more, they have to find out which parts of a process should change because AI can now run tasks that previously needed people to move information between systems.
The also explains why the technology underneath AI has become part of the adoption question.
The infrastructure underneath AI
In projects that never make it beyond the pilot stage, the snag is already showing up.
GFT Technologies commissioned Wakefield Research to survey 945 CIOs and CTOs at companies with at least $500 million in annual revenue across 19 countries. The results show how much old systems are holding AI back. Some 84% had canceled at least one AI pilot or project because of legacy-system limitations, and 95% said legacy technology had slowed their ability to use or scale AI. For 56% of them, that delay was moderate or major.
As the applications surrounding an AI system often contain the information and processes that make the system useful in the first place, those limitations matter. An answer can be created by a model. And yet, it can’t independently solve a problem caused by disconnected databases, outdated applications or restricted access to business information.
The trouble is currently not only about deployment speed. 93% of the executives surveyed by GFT said that running AI on legacy systems without modernization would eventually trigger an enterprise-wide security crisis. At the same time, 89% were concerned that global investment in AI could be expanding faster than the business value the technology can realistically deliver.
That puts modernization in a different position. Replacing an old application or reorganizing data can easily be treated as a separate IT project, particularly when a new AI application is more visible to the rest of the business. But if that application depends on older systems, the two investments are, after all, increasingly connected.
The more AI is typically projected to do maybe, the harder it becomes to keep those issues separate.
When AI starts taking actions
The difference becomes even more apparent, as businesses move toward agentic AI.
An AI system that recaps a contract can remain relatively separate from the rest of an organization. An agent anticipated to review that contract, spot an issue and begin the next step needs access to the relevant information, permission to act and a way to interact with the applications that control the process.
That means the move from AI that creates an output to AI that dos a task also changes the infrastructure requirements around it. An agent working across a business inherits the limitations of the systems it has to interact with, which makes data access, application integration, security and workflow design increasingly important.
A pilot can operate in a controlled environment with bounded data and human supervision. A production system has to deal with exceptions, permissions, security policies, older databases and the employees already responsible for the process. Pretty soon, companies moving toward more autonomous AI have to consider not only which models they want to use, but whether their existing technology can support those models once they begin taking actions.
Meaning also changing the role of technology partners working around enterprise AI. Microsoft recently announced its 2026-2027 AI Business Solutions Inner Circle recognition, which brings together partners working with the company on AI-powered business transformation. Sonata Software was, as a matter of fact, among the companies recognized, receiving its sixth Inner Circle recognition.
The recognition comes, as the broader Microsoft ecosystem connects AI adoption with the modernization of technology and business processes across Dynamics 365, Azure, Fabric, Power Platform and Copilot. More recently, the relationship expanded through Microsoft Frontier Partner status and participation in Microsoft’s Copilot Agents & Platform Engineering Depth Partners program, with a focus on advancing agentic AI.
This signals a broader shift in enterprise AI. As companies hand these systems more responsibility, the technology around them gets harder to ignore. Updating an application, cleaning up data or redesigning a process may not sound like an AI initiative, yet any of them can determine whether an AI system holds up once it’s plugged into the rest of the business.
That may be why the next stage of enterprise AI will depend on more than the capabilities of the latest models. Companies still have to make those models work with the systems, data and processes they already rely on.
The difficult part, in many cases, is no longer getting AI to produce an answer.
It is normally making sure the business is mainly ready for what happens after the answer.
