Trent Gillespie
August 12, 2026
I've talked to a lot of people about AI over the last few years, and even today they almost all end up at the same question: what do I do about AI?
They're usually not asking which app to buy or how to write a better prompt.
They're asking about their own future: what should I actually be doing about AI right now? How much should I invest? How do I ensure my company will be sustainable, and grow? Do I need to be worried about my financial security, and my family's?
I've had these conversations with high school students and tenured professors, with interns and PhD scientists, with entrepreneurs and CEOs of global companies. Almost none of them feel like they have a clear answer, and even worse, no trusted source to help them figure it out.
It's been a few weeks since I last sent this newsletter, mostly because I've been building something to help with that question. And I'd like you to be part of it.
I’ll introduce that here, and then keep reading to get a write-up on Cowork, the most significant change to AI applications in the last 18 months.
At noon Pacific on August 21 I'm holding the first AI SPRINT Monthly Event for the Stellis AI Leader Lab, a new program for leaders who want to take control of their future with AI, not just follow the pack.
It's where I'll share and test my latest guidance, not just on using AI, but on building organizations that can succeed with it. We’ll learn together what actually matters and what steps to take next. And I’ll bring our research, tools, and partners I actually trust.
It’s based around the AI SPRINT, to show progress monthly. And in that spirit, we’re kicking off with the following:
Monthly AI SPRINT virtual events with updates, new concepts, guests and open Q&A
Monthly vendor and use case deep dives
Weekly office hours with our AI strategists
Online forums, including industry-specific ones
Research, tools, and templates as we build them
Partner offers, including our 20% rebate on new OpenAI licenses
On August 21, I'll cover what Leader Lab is, the latest on AI that leaders need to know about, bring in an AI startup (TBC) to go deep on their top use case, give beta access to our new platform, Clarity by Stellis AI, and then open it up for what help you need. I'll be joined by Niraj Desai, another former Amazonian, who runs Leader Lab, and others on my team.
The founding member price is $299 a month. As an AI SPRINT subscriber your first 30 days are free, starting with the August 21 event.
One other invite, and then we’ll cover Cowork.
I'm in Indianapolis this weekend for a couple of sessions, and I'd like to see you at either one.
Saturday I’ll be leading the CEO Summit, and I'm going out on a limb with it. There will be up to 125 association CEOs in the room, each of them serving a different industry. I'm not going to stand up there and build one strategy for them to watch. Every one of them is going to build a job-creating AI strategy for their own industry, in the room, at the same time.
That's a hundred and twenty-five industries in one session. I don't know of anyone who has tried it at that scale, and I'll be honest that it could go sideways. But if it works, it's the clearest thing I can show you about how this actually operates. The value isn't me having the answer for your industry. It's you having a method that produces one, and being able to run it again next quarter without me.
Monday I’ll lead the AI in Your Hands session, sponsored by the ASAE AI Coalition, on using AI in your own role.
I'll also be at our booth, #2510, in between. Stop by.
Now the news, and this is the big one from the last several weeks.
In July, ChatGPT made its biggest product change in a year and a half. It added “Work”, right in the interface you already use.
What's notable is that OpenAI was late to this game. Claude shipped Cowork in January, Microsoft brought a limited version to Copilot in May, and Google launched Spark in July as well. When four platforms ship the same thing inside seven months, it's not a trend anymore, and I expect Cowork to be the primary experience across every major AI app within the next three to six months. Time to start using it and preparing your teams for it.
The idea is simple enough. Instead of sending messages back and forth, you hand it a task and it works until the task is finished. It connects to systems you already have access to, starting with email and calendar but also your CRM, knowledge systems, Slack or Teams, ERP, and so on. You give it skills, meaning instructions on how you want recurring things done, like how your monthly report gets written or how you want a deck built or what to check before a client call. And then you work alongside it, giving it recurring jobs like watching a competitor's site or running background research.
The part that is important is that the skills are the durable piece. Whatever model you're using now will be replaced within a quarter, that's just how this goes, but the skills you write don't expire and they don't belong to a vendor. That's why when someone asks me what to learn now, my answer is skills, not prompting.
Importantly, Cowork is a different kind of AI risk than we've dealt with so far, because you and you're staff are not just delegating work. You're delegating your access and some amount of your judgment to software.
The specific problem is called indirect prompt injection. It doesn't look like hacking at all. It looks like an ordinary email or a shared document or a web page that happens to contain an instruction, which your agent reads and then acts on. There's no malware involved and nobody stole a password, so there's nothing for your security stack to catch. The exploit is written in plain English.
I want to be careful here, because there's a lot of overheated writing on this and I don't want to add to it.
As far as I can tell there's still no publicly confirmed case of a company losing data because someone hijacked an AI agent this way. Which is good, and it's the argument for getting your controls set up now rather than later.
But the risks are established and demonstrated in proof-of-concept. In April, Google scanned billions of public web pages looking for these traps and found people already laying them, including hidden instructions telling an AI agent to send a specific dollar amount to a specific PayPal account. To Google's credit they were honest that most of what they found was crude, more prank than professional, though malicious detections were up 32% over four months. And separately, researchers have proven the attack works against real production systems. There was a zero-click flaw in Microsoft 365 Copilot where a single crafted email could pull internal files out of a company's tenant with no user action at all. Microsoft patched it and said they found no evidence anyone had used it.
So the technique works, the traps are getting planted, and the tools are landing on your team's desktops this quarter. Those three haven't intersected in public yet. I'd plan now for the quarter when they do—probably faster than we expect.
One more thing that I think is underappreciated: these agents now run scheduled work in the cloud, on your credentials, while your laptop is closed. That means productivity gains, but also less supervision. Both are important to plan for.
Check whether you already have it. “Work mode” in ChatGPT, “Cowork” in Claude or Copilot, “Spark'“ in Gemini. If you're in ChatGPT, the desktop app got rebuilt to merge Chat, Work and Codex into one place, and the old app is now called ChatGPT Classic.
Connect low-risk systems first, calendar and email before anything financial. Then give it an actual task, the kind of thing you'd hand to an assistant, not something you'd type into a chatbot.
Before you go wider than that, put a few things in place: Require a human approval for anything that sends, deletes, or shares outside the company. Give each agent the narrowest access it needs. Check scheduled tasks routinely.
We wrote a research brief on Cowork for our clients last month and I'm sharing it with you today. It's the guidance on what it is and how to prepare for it organizationally, and is a fair example of the kind of thing we will produce and share inside Leader Lab.
ChatGPT released the GPT-5.6 family, with “Sol” as the top tier. GPT-5.6 is good and it should probably be your default if you're on ChatGPT. Claude released Opus 5, which I've been using as a daily driver.
Keep an eye out for “Astra”. OpenAI named it publicly on August 1 and showed it to regulators in Washington, but hasn't shipped it, priced it, or said whether it'll carry the GPT-6 label. It'll be another step change when it arrives. Expected to arrive shortly, but no date is confirmed.
Hope to see you there, or at ASAE!
Trent Gillespie is CEO of Stellis AI and a keynote speaker helping business leaders understand and operationalize AI in their companies. He spent almost nine years leading global innovation efforts at Amazon before leaving to help other companies build the capabilities they need to compete. Book Trent to speak to your group or book a call to discuss using AI within your business.
Tell us what you thought of today's email. |
Did someone forward this newsletter to you? If you're not already signed up, you can subscribe to AI SPRINT™ for free here.