On the surface, it looks as if Anthropic, OpenAI and Google — three leading enterprise frontier AI model vendors — are in a capability horse race. The harder problem for businesses isn’t picking an AI winner but ensuring they can move to another AI model as features, pricing and contracts change.
This is shifting attention to orchestration layers that provide controls and linkages to business systems. Businesses also want to avoid vendor lock-in, IT’s eternal problem.
Model Overload
The AI vendors are acknowledging the orchestration problem, but also potentially compounding it. In February, OpenAI rolled out Frontier AI, a platform for managing agents across business systems. Meanwhile, for example, ServiceNow has its AI Control Tower, its own agent orchestration layer. This adds complexity to enterprise decision-making but also illustrates how attention is shifting from model capabilities to architecture.
“The model is not going to be the special sauce — they all produce spectacular models,” said Alisa Scharf, chief AI officer at Seer Interactive, a digital marketing company. “It’s more about the features of the platform that we care most about now.”
Scharf’s firm has standardized on Anthropic and still has many ChatGPT accounts but might ultimately end up on Google’s AI platform.
One feature that mattered to Scharf was Anthropic’s implementation of its Model Context Protocol (MCP) as well as features such as skills and reusable workflow instructions that sit on top of the model. Anthropic developed MCP as an open standard that provides a vendor-neutral way to connect AI to business data. It’s gotten wide support from other vendors.
But while using Anthropic, Scharf keeps an eye on Google. The firm already uses Google Workspace as well as Big Query, Google’s data warehouse, and the tech giant’s BI application Data Studio.
“We have every reason in the world to go all in on Google,” Scharf said. “But it just simply pales in comparison to what we can do with Anthropic.”
If Seer Interactive shifts to Google, the change should be easier because of MCP, which acts like “connective tissue” to its business data, Scharf said. “We should be able to just kind of point the MCP back at Google, as opposed to rebuilding architecture from the ground up,” Scharf said.
Model Choice
Enterprises want flexibility, and that is happening at the architecture level. Apart from the risk of vendor lock-in, users are frustrated if they can’t access the model of their choice.
Gartner analyst Kjell Carlsson put a slightly different take on the problem, calling it “lock-out” — a situation in which enterprises that have standardized on one AI model are getting pushback from employees. “Lock-out is occurring at every level as IT organizations struggle to cope with demand for an increasingly large AI ecosystem.”
“Organizations are increasingly recognizing that no single vendor can meet the breadth of their AI needs,” he said. To respond, organizations are increasingly investing in AI architecture teams.”
Conducting the AI Orchestra
Jeff McMillan, the former head of firmwide AI at Morgan Stanley, who now runs his own advisory firm, McMillanAI, outlined the multi-layered stack that’s needed.
It includes data, the data model, controls and governance, AI models, orchestration and applications.
“The frontier models are going to shift over time,” McMillan said, meaning enterprises will want flexibility to use different models as their capabilities evolve.
Models will improve on different tasks over time, and organizations want to avoid the risk of a single frontier model falling behind, McMillan said. “You want to keep some pricing power by playing vendors off each other,” he said.
The orchestration layer “gives you the flexibility by avoiding direct connection between your apps and the model, ” he said. “If you’re connecting your apps directly to models, you are absolutely going to get locked in.”
In that framework, the AI model becomes something that can be swapped in or out. This makes the data plumbing and orchestration all the more important.
“Enterprise customers expect enterprise capabilities,” said Ray Wang, founder and principal analyst at Constellation Research.
“We’ve moved from model capability to enterprise requirements of security, governance, agility, support, integrations and workflow fit,” Wang said. “The ecosystem piece will be key in the next 12 to 18 months.”
Governing AI
The stakes can be high for getting governance right.
Frank Dickson, an analyst at IDC, said MCP “is being adopted at a pace that has outrun the security infrastructure needed to govern it.”
Dickson said MCP servers are easy to create, and organizations might not realize how many unsanctioned servers are already in their networks. AI agents can use these services to access systems and data and act autonomously. Many organizations lack the tools and governance needed to manage them properly, he said.
Pavan Madduri, a senior platform engineer at W. W. Grainger, a distribution firm, said many companies, including his own, remain in the evaluation phase, running multiple AI models and tools in parallel. The specific frontier model matters less than the workflows and plumbing around it, he said.
But the real dependency risk might not come from the models themselves, Madduri said. Instead, it comes from the orchestration, workflow and data integration layers built around them.
Avoiding lock-in, Madduri argued, will require a full understanding of data connections and, in some cases, building out a company’s own orchestration layer and data pipelines. Relying on third-party orchestration is where “real lock-ins happen.”
The capabilities of models matter. Madduri’s own usage illustrated the point. He uses Perplexity for research, Claude for development-oriented work, ChatGPT for general-purpose work and Gemini for its context window management and ability to work with large amounts of data. “Every model has their unique pros and cons,” he said.
But broader problems remain.
The AI Value Puzzle
Organizations are spending heavily on AI but struggling to get business value. Many companies are still at an early stage, either experimenting with AI or deploying it in limited and fragmented workflows across the enterprise, according to a research paper by Jeanne McClure, a postdoctoral research fellow at NC State University’s Data Science and AI Academy, and Gregg Gerdau, a managing partner at Matador Advisors, a venture advisory firm.
At that stage, model choice matters less than an organization’s ability to integrate AI into workflows and operations, the researchers argued. Vendor selection becomes more consequential only after an enterprise reaches what they describe as the “orchestrated” stage of AI maturity.
“Vendor choice becomes truly consequential only once an enterprise reaches the orchestrated stage, and at that point, selection is highly use-case specific, not a generalized ‘which model is best’ decision,” the researchers wrote in response to questions.
