AI context systems

You wouldn't hire a senior level employee and then hand them a broom. Right, but that's exactly how we're treating AI right now.

I think businesses are just starting to realize there's a huge gap between what AI models are now capable of and how we're actually using them. The reason for that gap is that we haven't figured out how to work AI into our processes the right way.

Most businesses are just looking at their current processes and trying to figure out which individual steps could be handled by AI. And when you look at it that way, the answer is typically not a whole lot.

Most of the important steps in your existing processes require some degree of judgment or context or trust that these models just don't have yet. And so we end up giving them only the most basic tasks where there's little to no risk of them making a mistake.

Things like writing the first draft of a document that a human then still has to review, or brainstorming a list of ideas that a human can then select from...

in a thousand different ways with the ability to be multiplied a thousand times. If you were to hire an exceptionally talented person to your team right now, you'd be foolish to just give them your existing processes and say, hey, follow this checklist.

No, you would start from first principles. You'd start with the goal in mind and you determine a process that works with the talents of this person.

So I would challenge you to pick one process in your organisation. Not an individual task, an entire process.

Scrap the existing checklist and think through what would it look like if AI handled this from start to finish. We've been doing this at my digital marketing agency for the past six months or so, and I can tell you it's difficult.

When you first start things with the process, you're gonna say, no, I need a human for that. I need a new human for that.

I really can't hand it over to AI. But if you really force yourself to think, what if I had to hand this over to AI?

Gun to my head, why do I feel like the AI isn't capable of this specific step?...

. Maybe a specific step requires a human to go into a specific piece of software or multiple pieces of software.

Well, AI can already use most popular software, and I think it's safe to say that within a few years, if your software can't be used by AI, it will be obsolete. Now, I'm not saying you can replace all the humans in your company.

I think as you go through this process, you're gonna find the places where you still do need humans, and I think you should double down on those areas. And if the humans are freed up from all these other tasks, they can get really good in those areas.

I'm confident that over the next one to two years, we're going to see a huge delta between the companies that just keep replacing individual tasks with AI and those that redesign their processes from first principles with AI in mind.

If you're using a large language model for work, Gemini, ChatGPT, Claude, it's important that you give it the necessary context. All the model makers understand this, which is why all of them have made it easier to connect things like a Google Drive.

But I keep seeing people making the same mistake. At least once a week I hear of someone very excited that they've connected Google Drive to ChatGPT or Gemini or Claude and they've given it access to all their call transcripts and documents and now they can ask questions and produce strategy and brainstorm with all that context in mind.

And this is great. It really is.

You'll get much more useful high quality output from the model once it has that context. But as someone who's been doing this for about two years, I see people making the same mistakes.

The problem is when you give the model that much unstructured content, we're talking about hundreds of documents in some cases, you're leaving it up to the model to decide what context is relevant for a given prompt. Now some models are better at this than others...

content. So while having access to all this context might be better than no context, there's a much better way to do it.

I recommend setting up a project in Claude or Gemini and then just take the time to decide what context you actually want the model to have. Now, based on what you're asking the model to do, you'll want it to have a different set of context.

So the second half of the strategy is to clearly label all the documents you give it and then provide clear instructions on when to use each document. It does take a little time to set this up and maintain the project, but you can have the project help you with that part as well.

Some documents can just be added as is. For example, the transcripts of each of these videos is added to one master document.

And in the project instructions, we tell Claude or Gemini to pull from these transcripts whenever it needs an opinion from me...