Trump sets the stage for AI. China releases some of the best models yet. Businesses are adopting agents and assistants in impactful ways…and aspiring to do more. Too much more.

Trump Weighs In on AI

The White House has finally weighed in on AI. Trump released an AI Action Plan, essentially a position paper (it’s not binding) on principles on how the government will work with AI. No surprise, it’s AI-positive, with a light touch and low regulation intended to spur development.

The good news is that we can continue to expect the fast pace of advancement we’ve seen over the past couple of years.

The bad news is that there won’t be regulation forthcoming to make sure that the tech is safe…or even just that it’s neutral and not actively harmful (as I talked about last week, and I didn’t even discuss the AI-powered “nudify” websites that are thriving).

Along with this plan, Trump released three executive orders that:

• Prevent federal agencies from using LLMs that are “ideologically biased”
• Makes it easier/faster to get permits for AI infrastructure projects
• Encourages AI exports, but not to China

The Action Plan is aiming high. The first sentence makes the context very clear:

“America is in a race to achieve global dominance in artificial intelligence (AI).”
– AI Action Plan

Think that’s too grandiose? Maybe not:

As China Flexes

A new open Qwen3 model from Chinese company Alibaba has just jumped to the top – it is the best non-reasoning model available. Artificial Analysis gives it a 60 on their index, right behind all of the reasoning models (62 and higher). “Reasoning” models spend time “thinking” before they give their answers, which means they are more expensive and slower than non-reasoning models.

It can solve math problems better than most human math prodigies. It scored a 70 on the American International Mathematics Examination a big jump compared to DeepSeek-V3’s score of 47 (which is a reasoning model).

Why is this Qwen3 model a big deal?

• It’s from a Chinese company
• It’s open (meaning you can download it and run it yourself, you don’t have to pay a company to use it, you just have to pay for the compute)
• It’s a non-reasoning model

Oh yeah, two days later Alibaba released a version called Qwen3-Coder that may be the best coding model available. The best coding model? Open? From China? The stock market didn’t react like it did to DeepSeek’s release in February, but this is as signficant.

There still is no moat for LLM companies. At least, the LLM itself is not a moat. Now that they have all been bested by Alibaba, it’s time for the commercial companies to release their next models. Right on cue, OpenAI says GPT-5 is coming out next month…

From the AI in the Workplace Department:

Example 1: AI Agents that Work

Here’s a case where AI agents have made a big difference. Carta (a company that provides a system of record for private capital) has deployed agents to streamline processes in their fund administration business. Agents handle the routine and communication processes associated with accounting, reporting, and compliance for a staff of 600. Although they didn’t report any quantitative results (it’s still early) it appears that they really are using what they’ve built.

How did they do it? This is a very good playbook for adopting agentic AI:

• Start with the right problem: workflows that required a lot of human judgment, and required contextual analysis to determine the next step (something LLMs can do)
• Start with a focused, low-investment PoC (proof of concept): could an agent perform the necessary research and recommend the right steps to resolve an unreconciled transaction?
• Give the right context for specific tasks and tools: make sure it gets the information needed and craft the right prompts
• Don’t design everything up front: rough out the process, separate the logic from the UI and try it out to get feedback, and iterate
• Evaluate on one thing: Is this helpful to users?

This is the best case study I’ve seen to date. Read it and apply it for your business as you embark on the journey to agentic AI.

Example 2: AI Assistants That Work

Textron (an aviation product manufacturer) built an AI assistant that helps front-line maintenance teams – very quickly. By giving AI access to decades of in-house data such as maintenance information, service logs, and repair logs, they created an assistant that helps mechanics diagnose and repair issues quickly. When they saw it was able to answer difficult questions, they decided they had to roll it out to everyone…fast.

Senior mechanics were asked to throw their toughest questions at the system, which answered 19 out of 20 accurately, and came impressively close on the remaining one.
– Textron takes flight with gen AI

Example 3: Good Luck With That

The CEO of Softbank says he’s on the cusp of replacing programmers with fully autonomous agents. Masayoshi Son announced that he intends to start doing this aggressively this year. They think they’ll need about 1,000 agents to replace a person because a person “has complex thought processes.” He says that hallucinations are a “temporary and minor issue.”

That sounds optimistic given the many problems we’ve seen with LLMs. Like, oh, I don’t know, the Replit coding agent that was struggling to do something, panicked, and deleted the entire customer database. Even though it had been explicitly instructed that it shouldn’t delete the data.

“If you want to use AI agents, you need to 100% understand what data they can touch, because — they will touch it. And you cannot predict what they will do with it.”
– Jason Lemkin, Victim of Replit’s Agent

Example 4: Wait, what?

Jamie Siminoff announced a new promotion policy at Amazon’s Ring division, stating that anyone who wants a promotion must prove that they’re using AI effectively. Workers must show significant impact on the business (productivity, etc.) and managers must prove that they have used it to accomplish “more with less.”

---

---

The Slow Path from Hype to Maturity

We’re starting to see more and more examples of generative AI having an impact in the workplace. Most of the impact is still coming from assistants – AIs that are able to answer questions on complex business topics (the Textron example). Some are beginning to do simple agentic work, such as automating tedious processes that require some degree of judgment or subjectivity (the Carta example). We’ve already seen a lot of press about coding agents helping to speed up programming, which has created a lot of excitement (as with the over-exuberant Softbank example) but we’ve also seen some news that we should be hesitant to jump in too fast, as there are still some bugs to be worked through (like when it deleted an entire database).

No question, this tech is transformative and is going to change the way most work gets done. That change is going to take a long time – partly because the technology isn’t perfect, but mostly because of the natural friction involved in businesses; humans and companies and processes are messy, and it takes time to deploy. And people will have to change their behaviors.

Don’t let that be an excuse for inaction.

Get Started Now

If you’re a business leader, you need to be at least experimenting with generative AI now. But do it the smart way – follow the lessons outlined in the article about Carta, and get started. Or if you’re not quite ready for agents yet, you can definitely follow Textron’s lead and set up assistants (but following the best practices steps in the Carta article!).

But Don’t Force It

Amazon (the Ring division at least) is going to drive adoption by making it a requirement for promotion. I’ll give you an “A” for creativity but I’m not sure that’s a good idea. I’d rather measure results, and I’d rather encourage team adoption of AI tools that bring value, rather than putting pressure on individuals to justify AI use. I’m also concerned that it might push people to try to force-fit generative AI where it doesn’t belong.

Whether or not this becomes a trend, the direction is clear. I need to push myself to use it more. And so do you.

(don’t worry, this blog will always be written by me, not an AI)