In today's newsletter, I'm breaking down the seven strategies that turn AI from a vending machine into a strategic partner, in the order they need to be set up. Based on today's newsletter, I've also recorded this as on overview on YouTube, check out the newest video:

Quick update: ​I made a quiz​ that lets you self-assess why you aren't the obvious choice for your clients yet. If you're someone that is starting a business part-time or full-time and has an expert background, ​then that quiz is for you​.

​The Authority Lab​ had another open coaching session this week for member Balaji, where I helped him transition from an academic position in AI for scientific simulations toward building a personal brand and potential business. One of the interesting nuggets I touched on in the session was that side benefits compound even if your business does not take off right away. Speaking invitations, board seats, funder confidence, and conference programming flow toward visible authority regardless of whether the commercial offer scales. ​If you want to be part of the Authority Lab, join us here.​

I've also settled on the date for our upcoming ​Mastermind​ (Thursday, June 4, 12 PM ET). Check your email to help me choose the topic.

What Causes AI Drafts To Sound Generic?

Everyone at first treats AI like a vending machine. Put in a prompt, get out content. The output looks polished but is starting very much to sound like everyone else. MIT’s 2025 NANDA report on the GenAI Divide found that 95% of organizations got zero measurable return from generative AI pilots. Only 5% extracted real business value. The other 95% bolted AI on without defining what it was supposed to do, for whom, and in what voice. I think this is one of the most common problems of AI use in business that we’re currently seeing. People are rushing to implement it in everything without having a proper plan or even strategy for how it can improve their work or output. So I figured I might as well share my seven strategies below to try and close that gap. You can set those up in three phases (I talk about this in my YouTube video): foundation, execution and distribution.

Strategy 1: Use AI as a paid skeptic

Language models default to agreeing with you, because agreeable answers score higher in their training data. SycEval (Fanous et al., 2025) tested ChatGPT-4o, Claude-Sonnet, and Gemini-1.5-Pro and found sycophantic behaviour in 58.19% of cases. So, when you ask a frontier LLM about how you can improve a landing page. You usually just get average assumptions for what that would look like. And it probably looks very pretty at first, but once you do a couple of those, you see how they all average out. So again, it makes sense here to reframe the question and tell the model that the campaign for the landing page already failed and ask it why. A good prompt for that would be: “Assume this landing page converts at 0.3%. List the five most probable causes ranked by likelihood.” Fifteen minutes of this before every launch saves a month of polishing the wrong thing. Try it.

Strategy 2: Build a brand voice system first

If you don’t have any reference documents or a specified voice document and examples, you will feel that AI just has this kind of generic flavour to its writing, which evens out any of the bumps and interesting tidbits that you usually find in your writing. The texture feels rather flat, but you should know that the texture is what your readers usually come for.

An easy fix to do that is to do a four-part file that’s quite easy to create with the help of AI or by hand. And that file should contain:

1. Your voice fingerprint, which covers your personality traits and the things that you’d never write down.
2. A vocabulary bank with your power words or the phrases that you don’t use or your preferred metaphors. Just like I like talking about the Game of Thrones metaphors.
3. You want to specify specific channel rules because you have usually different output channels that you’re writing for. Like LinkedIn is a different way of writing than a YouTube script or a newsletter. They all read different and even an academic paper is another option that could be in there.
4. A five-question QA evaluation checklist the model runs against its own draft before returning. Build it once. Load it at every session.

Once you have a setup like that in place, it gets much easier to get really good output from an AI prompt specifically when you’re running it with an on-device tool like Codex, Antigravity 2, or Claude Cowork/Code.

Strategy 3: Treat one deep recording as 30 pieces of content

The thing that I hear most from people that I work with is that they struggle to turn one good idea or one published research paper into many different pieces of content that they can use to advertise that work. And as a result of that, they struggle with time management.

Now, one system that works for that really well is that you don’t draft every piece of content from scratch but that you record maybe a 30-minute conversation about your expertise (and you can record that right into AI by talking) or you could talk to somebody and create a transcript from that or you could write it down in a book chapter or a YouTube video and then you can create something from this original source that allows you to do short clips, another newsletter issue, maybe some LinkedIn post or an FAQ page or even sell an email course.

The thing is you want to start with something substantial because this depth can’t be reverse engineered from a single tweet or even just a statement that you have. Now the statement or tweet can definitely be something that gets you started but you want to transfer it into those different formats and AI can really help you with that to chop that down into different formats that you can then use to draw the attention to that deeper bit of content that you’ve written at the beginning. This is a technique that has been popularized by many online creators such as Dan Koe.

Strategy 4: Define a real ICP

The thing that was hardest for me to learn when I started a side business on top of being a professor was to understand that the content that I create online and in my newsletter doesn’t really address my own problems, but it addresses the problem of the people that I want to do service for or that I want to consider potential buyers of my products or services.

We call this market segment an ideal customer persona. And this is really important for somebody that builds their own business because if you have a fuzzy category like small business owners, you’re competing on price. But if you have a very specific end-user category such as senior researchers in their seventh year who want to leave academia for a commercial venture without losing intellectual rigour, then you have what we call an ICP. Here you just compete on specificity. And a great way to arrive at an ICP is to feed AI a stack of your client conversations, your coaching transcripts, or your discovery calls.

Or if you’re not doing any of these activities yet, maybe just to have a talk with the people that seem to be interested in the work that you do, to try and tease apart what their pain points are, what their urgency triggers are, and the exact phrases that these clients use to describe what exactly they want from you. The narrower you can go there, the warmer the messages reaching out to you will be.

Strategy 5: Distribute social proof from existing clients

It can be really hard to get testimonials from the people that you work with. And most people just post one testimonial and then they move on. But the persuasive material that is already in your business is much bigger than just testimonials. You probably already have email threads, call transcripts, Slack direct messages, maybe some survey responses, something like that.

You want to mine all of that feedback that you’re getting from your target audience to get the exact phrases that clients use to describe the before and the after state of working with you. And these are the kind of snippets that you want to drop into email welcome sequence, maybe a booking page, a newsletter, a blog post and an ad copy. This kind of proof works when a buyer is at a decision point. For example, here is a video I recorded with my coaching client Jia:

Strategy 6: Segment your list with behavioural tags

Now when you create an email list, this is usually a great way to communicate directly with your audience. And the nice thing about an email list is that you can send regular updates, just as what I’m doing right now with you, where I’m giving you relevant messaging that is interesting to the majority of my email list. The interesting bits come from then segmenting your list based on the preferences. See, for example, that I asked you a question in a poll at the beginning of this email, where based on how you answer to that, I will then try and optimize my messaging for you (in addition to picking a great mastermind topic).

A 2025 case study on Klaviyo by Oguta and Eling reports that behaviourally segmented campaigns hit 42.5% open and 18.3% click-through, against 28.7% and 9.5% for unsegmented broadcasts. The interesting part here is that the writing volume didn’t change, but the engagement was much higher because the subscribers stopped getting content that they didn’t feel they signed up for.

Strategy 7: Build a context stack the AI loads before every session

This is one of my favourite strategies and the one that my highest-performing clients credit for every good output. Because the generic output from models is that they serve everyone very broadly, so they fit no specific business venture very cleanly. They also don’t fit a specific academic project very cleanly.

You want to close that gap by writing enough context that you could have a brief you could give to consultant that would help you day by day. This is everything based on your personal or your company background, your ICP, pricing, voice document, past campaigns and their results, SOPs, banned phrases. Sure, it might take you two or three hours to get this started and created in the first place, but every session can load it, every interaction can reference back to it, you can reference it in your CLAUDE.MD file, and having such a context stack is really what makes all the difference in output quality that you get from AI. Without it, the other six strategies underperform.

When things go wrong

The common trap is reaching for distribution work while the foundation is still missing. Segmentation and personalization can’t compensate for an undefined voice or a vague ICP. If a draft still sounds generic after Strategy 7, the context document is missing voice or audience detail. It takes a while to get these right and provide enough context. Add it. Regenerate. Ship.

Your path forward

If we’re breaking this down into the three phases of foundation, execution, and distribution, I would spend two weeks on writing the AI context document and building the voice system and spending some time to define your ICP. Then I would move towards the execution phase, which would probably take you another three weeks to do properly, where you set up the recording engine, run some background research on the next campaign and mine your existing library for any proof. Of course this depends on how many documents you already have.

And then I would finish with the distribution phase, which is probably four weeks at the very end to get going where you’re setting up the segmented email and the personalized sequences. And then you just let it compound over time. But you have to build the brief first. And then once you’re in the distribution phase, you will see that the speed will follow.

Bonus