Are you feeling smug about your team or company’s AI adoption? Well, I am sorry to be the bearer of bad news (#sorrynotsorry): You might be using the exact same tools as the top 10% of companies and getting roughly the value of a decent spellchecker.
OpenAI released data in August showing that the heaviest AI-using firms now get 8.3 times more output per person from AI than the rest of the pack. Six months earlier, that gap was 2.6 times. So basically, it tripled while the rest of us were busy congratulating ourselves for finally getting the whole team onto paid Copilot licenses.
What separates the people at these companies comes down to what they’re asking AI to do.
Earlier this week, Neo Aplin and I put on a webinar unpacking 8 habits that separate top AI users from the rest of us. (You can watch the replay here).
Here are four of the strategies we discussed.
1. Start a new chat thread. Every single time.
You know that feeling when your AI seems to get dumber the longer you talk to it? You’re not imagining it.
Every chat thread has a context window, which is geek speak for short-term memory. For most of the major platforms, this is (currently) around 200 A4 pages, and that includes everything you’ve typed or attached (PDFs and images are token hungry!), and everything your AI has written back. Once you fill it, the AI starts forgetting what you told it. However, it VERY confidently fills in the gaps (most likely incorrectly).
Long threads have a second problem. If you ask about a people strategy question and then a tech question in the same thread, your AI will try to weld them together, whether or not they actually belong together. Needless to say, this results in poor quality output.
So: start a new thread for every new purpose. It costs you nothing and it’s the single easiest change you can make today.
When you genuinely need to keep going with your train of thought, write yourself a “handoff prompt”. Something like:
Summarise what I was trying to do, where we got to, the decisions we made, and the dead ends we hit, so I can paste this into a fresh thread.
2. Stop typing to your AI. Talk to it instead.
Most of us speak at 120 to 140 words a minute. I type at about 80 wpm on a good day, and I’m fast (and also not very modest about my fast typing speed).
Speed is an obvious benefit. But context is the biggest one.
When you type, you write like you’re using Google (i.e. brisk and caveman-like - “latest fair work legislation victoria”). When you talk, you speak like you’re briefing a human: here’s the background, here’s who it’s for, here’s what we tried last time and here are some other details that are on my mind.
So instead of typing something like “latest trends private health insurance 2026”, you might talk: “I want you to tell me about the private health insurance industry because I’ve got a meeting with a major player in that space. I want you to tell me what the 5-10 biggest trends going on right now are that I might need to know about in an Australian context in 2026. I want you to only look at reputable sources like the Australian Financial Review, for example. Maybe tell me if there have been any major legislative changes or any really big news stories that have made a material difference to the industry.”
Spoken version = roughly the same amount of elapsed time. Also = much more context.
And because of the context, AI will give you a much much more useful response.
Use the dictation button inside your AI tool, or download something like Wispr Flow. I use it because it works everywhere: a browser bar, a Word doc, my Substack drafts. It also fixes your speech as you go. Say “coffee on Wednesday, I mean Thursday” and it writes “coffee on Thursday”.
3. Your AI doesn’t know you nearly as well as you think
People fall into one of two camps. They either assume AI knows nothing about them (which is absolutely true when you first start using it), or they assume it knows EVERYTHING - which is oh so wrong.
Go and look at your memory settings right now or simply ask your AI to tell you what it knows about you. Most people get about 2 pages: some of it accurate, some of it badly quite out of date, and some of it complete fiction.
Memory builds itself passively by reading your old chats overnight (yes, AI dreams about you), which means it collects fragments rather than guaranteeing it remembers the things you actually need it to know.
The fix is a Briefing Document*. A briefing document covers things like: Who you are, what you’re working on, who your clients are, what your priorities are this quarter.
You can also include how you want to be treated. For example, do you want a hard, critical sparring partner that doesn’t let you get away with sloppy thinking. Or maybe you might want a warm comms person. Those are two completely different assistants and you need to specify this to get what you want.
My advice is to write one briefing document per project, paste it in or attach it at the start of a thread, and your AI will have much needed context (and an infinitely greater chance of doing a great job for you).
*A more advanced version is setting up a series of Projects and Skills (or Agents), or setting these context areas up within Cowork, all of which we cover in AI Mastery.
4. Ask it to argue with you
Almost everyone uses AI to produce things. More emails, more reports, more (very generic) slide decks. Very few people use it to really interrogate things.
Left to its own devices, your AI will agree with you and then start generating. This is not particularly valuable.
At Inventium.ai, one of our most used agents is “Charles”, a thinking partner whose entire job is to push back. Charles restates what I was trying to do, then scores my work against 8 criteria, from factual accuracy through to whether it’s actually practical.
Every problem that Charles identifies goes in a table with how serious it is, how confident Charles is on its view, and the fix. Anything he can’t be 100% sure of gets flagged as unverified, along with where I should go to check it.
What makes Charles particularly useful is the list of things he’s forbidden from doing: No skipping sections, no vague opinions, no letting a “decent” draft off the hook. Even when the work is good, he has to name at least one improvement.
If you’d like more tips (and demos of all of it), check out our webinar recording from Monday’s session.
And if you are keen to level up your own AI skills and get in that top 2% of users, I highly recommend checking out our latest (and, in my humble opinion, best and most comprehensive) course: AI Mastery. Enrolments close tomorrow, Thursday, September 10.
And if you would like a copy of our mate Charles, you can download the text to paste into an Agent/GPT/Gem/Skill below (for paid subscribers only).
Download Charles, our Critical Thinking GPT, right here.