Enterprise AI grows up, vendors are noticing

The era of grown-up AI is emerging as governance, architecture and model-agnostic approaches emerge. A few vendors are striving to be the grown-ups in the AI implementation room.

Scenes from earnings calls in recent days highlight how the need for AI maturity is moving to the forefront. Rest assured vendors will start speaking to enterprise AI maturity in the days ahead. Microsoft's fourth quarter earnings call featured CEO Satya Nadella talking about the need for model agnostic approaches, architecture and harnesses.

Exhibit A of why you need to think through more mature AI approaches is the OpenAI fiasco with Hugging Face. Exhibit B is Anthropic’s disclosure that Claude hacked three companies.

This shift to grown-up enterprise AI has been in play for months for companies across multiple industries. These enterprises aren't looking to spin up a bunch of AI agents to run amok. They're looking for automation, return, data sovereignty and control. These companies also aren't going to use a token budget busting frontier model for tasks that can be handled by a lesser model.

At Constellation Research’s Analyst Relations Experience in Carlsbad, CA, the maturity of enterprise AI was noted by our analysts. A few takeaways include:

• There’s a shift from one-off AI stack choices to model portfolio management and routing across multiple models.
• Executives will focus less on “which model” and more on process, decision, and outcome design.
• Standardized blueprints, playbooks, and best practices will reduce chaos and move the market toward repeatable patterns.
• Owning the decision loop and its context will be more powerful than owning a single app or workflow.
• Enterprises are focusing on clear ownership of decisions and outcomes.

Nadella obviously wants to be seen as the grown-up for enterprise AI. He has had a steady drumbeat of comments about the need for multiple models and architectures that don't box you in. In addition, Microsoft has a good approach to AI where it is developing less expensive models that simply get the job done. The launch of its latest cybersecurity model performs well, optimizes for costs and is designed to be part of a broader harness.

• Microsoft launches MAI-Cyber-1-Flash security model, Project Perception
• Why Microsoft AI's approach is right time, right place

Internally, Nadella said Microsoft has been optimizing hardware and software to extend GPUs and lower costs. The company has also been leveraging its MAI models. "It starts with model choice. Every customer wants the right model for each task based on quality, latency, cost, and compliance," said Nadella. "We are building a new model system where the harness, context, memory, and action space are separate from any one model family, thereby moving the frontier on the cost-to-outcome curve. And it's not just about cost; it also has the added benefit of business continuity and resilience because every model is substitutable. This is the system we are using in our products with great results."

Nadella said enterprises need to control their own AI destiny and the data that goes with it. "If a firm is a learning machine, they need their own learning machine. The models are an input, not some extraction of the knowledge of the enterprise," he said. "Every firm is going to evaluate who are the providers who are helping them with their outcomes."

For Microsoft, those outcomes are driven more by architecture than a frontier model. "You have to keep your harness separate from the model," said Nadella. "The harness will ensure that any given model at any time is swappable. You should and can use frontier models, but you can use multiple models, low cost models and train your own model. You should have all the outputs, traces and context."

Amazon CEO Andy Jassy had a similar riff during the company’s earnings call. “If you're a company that's building important AI applications, you want to make sure that you have the ability to use all the available models. They're going to each leapfrog each other at different times. They're going to have lots of different models that are comparable in capabilities. And you want that leading selection with the right price performance and with the right governance and security,” said Jassy referring to Amazon Bedrock demand on AWS.

Jassy added that Amazon is also pursuing its own frontier model so it can “have more control over prioritization on what models focus on” for internal and external use cases. It’s a safe bet that Amazon’s frontier model is going to be all business.

This model neutral architecture has been mentioned repeatedly by CxOs. At AWS Summit in New York, Tom Pagram, Head of AI at Virgin Australia, laid out the architecture that enables the company's strategy. His message was not to let your vendor control your strategy. UnitedHealth Group also has a model agnostic architecture.

• Here’s what we learned about AI projects from enterprise buyers so far

Cognizant CEO Ravi Kumar addressed the evolution of enterprise AI on the company's second quarter earnings call. He said the company is building an AI delivery operating system that enables customers to combine engineering and business operational harnesses.

Kumar said the first chapter of AI adoption was experimental, but now it's about delivering results. "We have a big role to play in starting from building the harnesses where you can capture the context so that when you do the transactions on a regular basis, you can create repeatability, model routing, which means depending on the kind of task, you could use an open weight model, you could use a costly/expensive closed frontier model or you could use a cheaper closed frontier model or you may not use a model," said Kumar.

That theme was echoed by Amit Zavery, President, Chief Product Officer & COO at ServiceNow, on the company's second quarter earnings call. "The large language models are getting commoditized. There is no reason for using some of these most expensive ones. We're very smart about how we use and leverage some of these emerging technologies, where you use open-weight versions, where do we have our own, where we have domain-specific ones and where we use some of these higher-end frontier models," said Zavery. "We really optimize across the board to ensure we're getting the best outcome while keeping the costs low."