Stefan’s AI 2025 Mid-Year Brief

The Enterprise Is In, The Agents Are… Trying Their Best

Blog Stefan’s AI 2025 Mid-Year Brief

The first half of 2025 made it official: AI is no longer a novelty demo tucked into your QBR deck. It’s now the thing keeping your business upright, like duct tape—but smarter and more expensive. Enterprises have stopped “experimenting” and started budgeting. And while agents have spent the year promising to do all your work for you, it turns out they still need a babysitter. What’s new is working. What’s overhyped is quietly being re-scoped. Welcome to the awkward teenage years of AI maturity (I think we can all agree that 1 year in AI feels like 6.5 years in real time, no?)

Let’s dig into what has caught my attention over the past six months when I wasn’t worrying about the end of the world.

Foundational Models: Smarter, Cheaper, Multimodal-er

The models got sharper—Gemini 2.5 Pro “thinks,” Grok 3 reasons, Claude 3.5 codes like your "meeting expectations” staff engineer—but the real story is accessibility. Running a GPT-3.5-level model is now 280x cheaper than it was in 2022¹. Let that sink in. Open-weight models are nearly indistinguishable from closed ones. And tiny models like Phi-3-mini are pulling off tricks that used to require 100x more parameters.²

This means the AI arms race beyond frontier models is no longer just for the Googles and OpenAIs. The field’s leveling—SMEs can now play. Which means providers have to compete on something other than “we have the biggest model.” Usability, integration, and actual results are the new differentiators - we’ll talk more about that below.

Oh, and yes—everyone’s still pouring money into it. Q1 alone saw $60–73B in AI investment. Nvidia stock is basically my retirement plan at this point. And the buildout is on: liquid-cooled data centers, energy-intensive GPUs, less energy-intensive model-based GPUs, and yes—small nuclear reactors. Because putting AIs next to nukes never went wrong and if it did we wouldn’t know about it because Skynet would have sent a robot back in time to make sure we never learned about it (nerds - please don’t attempt to correct my causal loop paradox, I’m tired.)

Enterprise AI: Beyond Pilots, Into the Ledger

75% of companies are now using AI (up from 55% in 2024). Not “evaluating.” Not “running a pilot.” Using. And 92% say more investment is coming.³ Why? Because it’s finally paying off and because leaders are probably tired of their board members sending them incessant links from CIO Magazine showing how their competitors are crushing it. Some fun facts:

Software Dev: Copilot and friends are cutting dev time by 50% (per McKinsey).
IT Ops: AI-based provisioning is 10x faster.⁴
Knowledge Work: Contextual retrieval (meaning the systems understand intent of your question, not just the keywords you’re using)is reducing failed queries by half, per Anthropic ⁵
Retail & Marketing: Chatbots are driving 15% higher conversions, and Coca-Cola’s AI-fueled ad campaign led to an 870% boost in engagement and a 2% bump in sales.⁶

In short: AI’s gone from “science project in the basement” to “top-line growth machine.” Enterprises are realizing it’s not just a tool for trimming fat—it’s the engine that when coaxed can drive the next wave of revenue.

Agents: Hype, Meet Reality

2025 was supposed to be the year of autonomous AI agents. And in some ways, it was—if by “autonomous” we mean “kind of need a human to check their homework.” But we’re only half way through, so we’ll reserve judgment til EOY.

Early hype painted AI agents as fully autonomous sidekicks—able to book your dinner, balance your budget, and maybe even break up with your therapist. But as IBM’s Maryam Ashoori points out, today’s “agents” are mostly LLMs duct-taped to basic planning and tool-use—not the independent, reasoning entities the term implies. While 99% of developers are either building or exploring agents, some experts argue we’re just rebranding old-school orchestration with a shinier label and calling it the future.⁷

Real talk: success rates for agentic systems in production hover around 50%—and that’s generous. Many crash and burn. Multi-step processes? Even worse. Math time: A 10-step workflow with 90% accuracy per step nets you a 35% success rate. This is why sometimes it pays to be a liberal arts major and not let math get in the way of good narrative.

That said, focused agents are delivering—especially in:

Customer support (routing, resolution)⁸
Cybersecurity (detection and reporting)⁹
Compliance (real-time checks)¹⁰

Salesforce’s Agentforce is already showing real ROI. But if you want to know the unlock after studying nearly two dozen agent deployments? Narrow beats general. The dream of a general-purpose, autonomous enterprise agent is still just that—a dream (well, that one that when you wake up, a sock is gone and someone opened the garage door - so SOMETHING is happening). What’s working are bespoke agents, built for one job, doing it well.

Governance… Is Coming

Enterprises are no longer asking “Can we do this?” They’re asking, “Are we legally safe to?” Governance, risk, and compliance are potentially company (or at least career) ending concerns so they’re front-line requirements. With 60% of AI projects still struggling to show ROI (despite so many people doing them - see above!), the pressure is on.

Regulators are on the move too:

The U.S. is pushing a 10-year moratorium on state-level AI laws (to me, it’s a good thing…)¹¹
Japan, Australia, and Oman launched sweeping AI bills¹²
UNESCO hosted a global AI ethics forum just last week because yes, that’s a thing now¹³

Ethics, explainability, auditability—these aren’t buzzwords anymore. As we talked about at HumanX 2025, a TRUST framework for AI deployment isn’t optional - it is a go/no-go checkpoint.

H2 2025: What’s Next (And What’s Not)

The rest of 2025 is shaping up to be less about big reveals and more about refining what we’ve already started.

Emerging Trends to Watch:

Multi-agent orchestration: Not Skynet—more like a group project that finally works, except for Brad who will continue to phone it in. Imagine a "swarm of AIs" collaborating on tasks, each specializing in different functions.
AI-native apps: Platforms built around AI, not just with it bolted on. No, not every company can just change their domain to .ai and say they’re an AI company. Vertical orientation and truly using AI to make thing more efficient - that is the 2025-era play. Think "AI-native ecosystems" that are designed to redesign workflows from scratch, exemplified by AI-native CRMs that predict sales trends and write proposals, and AI-based HR platforms for candidate screening and feedback management.
Voice-first + emotion-aware AI: Your call center rep may soon be (and in some cases in healthcare already is) an AI that sounds suspiciously empathetic.
Ambient Intelligence: Finally, and I’m not totally predicting a 2025 WIN here, we’re headed to where AI doesn’t just respond, it anticipates. The convergence of multi-agent systems, emotionally aware models, AI-native platforms, and voice-first interfaces isn’t about adding more tech bells and whistles. It’s about AI fading into the background and quietly running the show. We’re talking about systems that reroute your delivery fleet before traffic hits, or flag a failing HVAC unit in a warehouse before anyone notices the temperature change. No dashboards (finally, after 25 years we can stop pretending we look at these!), no prompts—just context-aware systems that understand the moment and act. The shift here is subtle but seismic: from “Hey AI, do this” to “Oh, it already did.” It’s not just automation—it’s intuition at scale. And it’s going to completely reset expectations for how users interact with tech. Which is probably why Sam Altman paid $6.5b for Jonny Ive’s company.

As said above, the race is shifting from “who has the best model” to “who can embed intelligence seamlessly, scalably, and ethically.”

Bottom Line: Less Magic, More Metrics

The first half of 2025 proved that AI is here to stay—but the fantasy version is taking a bit longer. Enterprises that win won’t be the ones deploying the flashiest tech—they’ll be the ones solving real problems, integrating carefully, and building systems that people actually trust.

This isn’t about chasing every shiny object. It’s about building an AI foundation that actually works. The good news? The pieces are finally falling into place and we can’t wait to talk about it at HumanX 2026.

___________________________________________________________________________________

¹https://blog.irvingwb.com/blog/2025/05/the-2025-ai-index-report-ais-influence-across-society-continues-to-transform-the-way-we-live-and-work.html

² https://hai.stanford.edu/news/ai-index-2025-state-of-ai-in-10-charts

³https://www.coherentsolutions.com/insights/ai-adoption-trends-you-should-not-miss-2025

⁴Q1 2025 Tech Trends Report: The Evolution of the AI-Powered Enterprise

⁵https://dexian.com/white-paper/q1-2025-tech-trends-report/

⁶https://teachthe4ps.com/promotion/the-ai-powered-future-of-coca-cola-from-personalized-ads-to-smart-inventory/

⁷https://www.ibm.com/think/insights/ai-agents-2025-expectations-vs-reality

⁸https://www.lyzr.ai/state-of-ai-agents/

⁹ ibid

¹⁰ibid

¹¹https://thehill.com/policy/technology/5355684-ai-moratorium-sparks-gop-battle-over-states-rights/