Mid-Year Report: AI’s Biggest 2025 Surprises Revealed
This summer, we’re checking in on our TechArena predictions for 2025 to see how they are holding up.
Today, TechArena correspondent Will Torresan sat down with Intel’s Lynn Comp to discuss how her AI predictions for 2025 have performed.
Will: You predicted that in 2025, data management, security, and governance would be more important than ever for companies looking for great AI results. Six months in, how have you seen companies investing in data structure? And is it possible yet to see the results of those investments?
Lynn: I am seeing companies really focusing on figuring out how to ensure predictable, well-governed results in their agentic AI workflows, which by definition is forcing them to methodically “trace” their data flows across an agentic system, especially where the data sources are common between different agents specializing in different operations. One big surprise in what agentic seems to have shifted to is the adoption of Anthropic/Claude—partly due to their model context protocol (MCP) that results in a structured and predictable communications substrate between different agents and partly because of the growth in using AI within coding tools. Menlo Ventures recently published an update in the enterprise large language model (LLM) application programming interface (API) usages and OpenAI/GPT fell to 25% from 50%, where Anthropic/Claude grew from 12% to 32% (2025 Mid-Year LLM Market Update: Foundation Model Landscape + Economics | Menlo Ventures).
Will: The “treasure hunt” problem you described—where users must discover AI value through trial and error—was a key barrier to AI adoption that you identified. What are some solutions you’ve witnessed to this self-discovery challenge?
Lynn: There has been a lot of reporting on best practices and success stories by different corporations in the leading business publications—a few examples in the Wall Street Journal include Walmart, Johnson & Johnson, and a few of the financial insutions who have been working with AI for many years. NetApp recently deployed their “AI Mini Pod” after extensive inputs from their customers on the most common, successful AI use cases for enterprises. I recently published an article in the MIT Technology Review summarizing the most common templates/uses I’ve seen across all the industries I’ve met with over the last 6 months. The goal of a template rather than exact recipes is to allow flexibility and customization for enterprises to realize “that sales challenge described in this specific article isn’t my exact problem, but the description sounds like it could apply to challenges I am facing in my own business.”
Will: You discussed how AI can only amplify existing greatness, not create it. What’s the most compelling example you’ve seen this year of a company using AI to enhance their differentiation?
Lynn: Some of my favorite examples right now boil down to AI in retail use cases. I recently posted a video of a very chipper and polite inventory bot going through the aisles of a local Harbor Freight store, and Walmart has done significant innovation not just in customer experience and catalog management but using digital twin technology to execute HVAC and refrigeration maintenance with less downtime: Walmart Bets Big on AI with “Super Agent” Strategy.
Will: What’s been your biggest surprise in AI trends in 2025 that you didn’t predict, whether it’s a technology breakthrough, a market shift, or an unexpected challenge that emerged?
Lynn: I did not understand in the beginning of 2025 just what was continuing to happen with AI training bots and the consequences on the open internet/content creators who faced large cost increases in doing business because of the openness common to the internet of the 2000 through 2010s. I have been curious what would happen as a result of copyright infringement cases and ongoing settlements between large content creators (New York Times, Wall Street Journal, etc.) and the hyperscaler/model owners racing to train new models in their pursuit of either profits or artificial general intelligence (or both). I had branch predicted that individual content creators who had previously accepted the informal contract with social media platforms so they could reach a large global audience would move more of their content behind a paywall of some sort. I did not expect CloudFlare’s move in “Content Independence Day” vs. an individual flood of content creators going to Patreon and Substack or Medium.
Will: Looking back at your predictions, which one has played out most differently from what you anticipated?
Lynn: Freemium business models. I thought in February that the introduction of DeepSeek would help normalize and balance out the pricing from major LLM vendors. Instead, at least one of them introduced a concept of a price shift for business users to $20,000, and later that same company introduced enterprise services as a revenue generator.
This summer, we’re checking in on our TechArena predictions for 2025 to see how they are holding up.
Today, TechArena correspondent Will Torresan sat down with Intel’s Lynn Comp to discuss how her AI predictions for 2025 have performed.
Will: You predicted that in 2025, data management, security, and governance would be more important than ever for companies looking for great AI results. Six months in, how have you seen companies investing in data structure? And is it possible yet to see the results of those investments?
Lynn: I am seeing companies really focusing on figuring out how to ensure predictable, well-governed results in their agentic AI workflows, which by definition is forcing them to methodically “trace” their data flows across an agentic system, especially where the data sources are common between different agents specializing in different operations. One big surprise in what agentic seems to have shifted to is the adoption of Anthropic/Claude—partly due to their model context protocol (MCP) that results in a structured and predictable communications substrate between different agents and partly because of the growth in using AI within coding tools. Menlo Ventures recently published an update in the enterprise large language model (LLM) application programming interface (API) usages and OpenAI/GPT fell to 25% from 50%, where Anthropic/Claude grew from 12% to 32% (2025 Mid-Year LLM Market Update: Foundation Model Landscape + Economics | Menlo Ventures).
Will: The “treasure hunt” problem you described—where users must discover AI value through trial and error—was a key barrier to AI adoption that you identified. What are some solutions you’ve witnessed to this self-discovery challenge?
Lynn: There has been a lot of reporting on best practices and success stories by different corporations in the leading business publications—a few examples in the Wall Street Journal include Walmart, Johnson & Johnson, and a few of the financial insutions who have been working with AI for many years. NetApp recently deployed their “AI Mini Pod” after extensive inputs from their customers on the most common, successful AI use cases for enterprises. I recently published an article in the MIT Technology Review summarizing the most common templates/uses I’ve seen across all the industries I’ve met with over the last 6 months. The goal of a template rather than exact recipes is to allow flexibility and customization for enterprises to realize “that sales challenge described in this specific article isn’t my exact problem, but the description sounds like it could apply to challenges I am facing in my own business.”
Will: You discussed how AI can only amplify existing greatness, not create it. What’s the most compelling example you’ve seen this year of a company using AI to enhance their differentiation?
Lynn: Some of my favorite examples right now boil down to AI in retail use cases. I recently posted a video of a very chipper and polite inventory bot going through the aisles of a local Harbor Freight store, and Walmart has done significant innovation not just in customer experience and catalog management but using digital twin technology to execute HVAC and refrigeration maintenance with less downtime: Walmart Bets Big on AI with “Super Agent” Strategy.
Will: What’s been your biggest surprise in AI trends in 2025 that you didn’t predict, whether it’s a technology breakthrough, a market shift, or an unexpected challenge that emerged?
Lynn: I did not understand in the beginning of 2025 just what was continuing to happen with AI training bots and the consequences on the open internet/content creators who faced large cost increases in doing business because of the openness common to the internet of the 2000 through 2010s. I have been curious what would happen as a result of copyright infringement cases and ongoing settlements between large content creators (New York Times, Wall Street Journal, etc.) and the hyperscaler/model owners racing to train new models in their pursuit of either profits or artificial general intelligence (or both). I had branch predicted that individual content creators who had previously accepted the informal contract with social media platforms so they could reach a large global audience would move more of their content behind a paywall of some sort. I did not expect CloudFlare’s move in “Content Independence Day” vs. an individual flood of content creators going to Patreon and Substack or Medium.
Will: Looking back at your predictions, which one has played out most differently from what you anticipated?
Lynn: Freemium business models. I thought in February that the introduction of DeepSeek would help normalize and balance out the pricing from major LLM vendors. Instead, at least one of them introduced a concept of a price shift for business users to $20,000, and later that same company introduced enterprise services as a revenue generator.
.png)
