Everything Goldman Sachs Taught Me About AI (In 10 minutes)

0:00 So, I'm going to give you everything that I learned after working at Goldman Sachs, one of the biggest financial firms on the planet. These are principles I still follow to this day in

0:06 everything that I do with AI. A company like this doesn't have the luxury of making mistakes, and one wrong data entry or hallucination can cost them millions of dollars, or even worse,

0:15 their reputation. So, if one of the biggest and most trusted financial firms on the planet follows these principles, it's because they work. And even if you're not working at a big firm, these

0:23 principles will completely change how you use AI and the results that you get out of AI. So today I'm going to share the five AI principles I learned from Goldman Sachs which I call vault and

0:31 show you how to apply them to your AI projects. So let's get into it. Okay, so V is to verify the output. And just a reminder, this is the framework I created from the principles I learned

0:40 while working there. And these principles apply whether you're using claude code, codecs, or whatever new tool comes out next month. So Verify, I think a lot of us, myself included, can

0:49 get a little too comfortable with AI because it gives you something that looks finished and it will deliver it to you super confidently. It's formatted perfectly and it probably took about 30

0:57 seconds. But something that looks finished and something that's actually correct are obviously two very different things. So Marco Agenti, Goldman's CIO, has made this distinction between the

1:06 reasoning of a model and its final output. Basically, the way a model breaks down and analyzes a problem can still be useful even when the final answer it gives you is wrong. So you can

1:14 use the breakdown to get value from it, but you still need to verify the result. And this actually starts before the AI even generates an answer. So my role at Goldman was in business intelligence.

1:23 And every business intelligence analyst in my role had to pass something called data school. We had to understand how data gets pulled in, how it gets cleaned, how it gets updated, and how

1:31 you turn all of that into something that people can actually use to make a decision because every report, every dashboard, every automation, and now every AI system is only as good as the

1:40 data that's actually powering it. And your company's data is probably one of the biggest advantages that it has because nobody else has that exact information. They don't have your

1:47 customers, your sales history, your internal processes, your support tickets, all of that kind of stuff. It's gold. But if that data is outdated or duplicated or messy or just wrong, then

1:56 plugging it into AI doesn't magically fix it. It just gives you the wrong answer faster. So when I use AI tools, I like to build verification directly into the processes. You could give the model

2:05 a prompt like, "Hey, you know, before you give me the final answer, recheck every number and factual claim. Cite the source for each one and flag anything that you're not fully 100% confident

2:14 about." And you don't have to manually verify every single word that it gives you. If you're creating a report, maybe there are only two or three numbers that could actually change the decision. So

2:22 have the model site those numbers and then you can spot check them yourself. And obviously for higher stakes projects you can even build a separate AI reviewer or multiple you know like a

2:30 team of AI agents that are just there to review. They'll check the first output before you see anything and then the whole system can iterate once again before you see anything. And that's an

2:38 additional check. But important claims still need original sources, deterministic tests or human review. So verify the data going in, verify the important outputs coming out and build

2:48 review into the process. But once you can trust the inputs and the outputs, you still have to decide whether AI should be in the workflow at all. So moving on to a augment, which means

2:57 augment don't replace. Now my full-time job at Goldman was basically building automations, reports, dashboards, and systems that made teams more efficient. And across the teams I worked with,

3:06 there were people doing some version of that work. None of this was some brand new AI initiative. You know, Goldman had been building automated data and risk systems for decades. and teams across

3:15 the firm were doing this kind of work before the current wave of AI agents hit the market. Which means there were already a crazy amount of problems worth solving before anyone started talking

3:24 about AI agents. There were manual reports being built that took hours. You know, data had to be moved between systems. Teams were waiting on information from other people. People

3:32 would enter something incorrectly with a fat finger or skip a step or or two people would follow the same process in two slightly different ways. Because at the end of the day, humans are

3:40 inconsistent. We get tired. We miss steps. communication breaks down and everybody makes mistakes. But a properly tested deterministic automation will run the same process the exact same way

3:50 every single time. Now, that doesn't mean an automation is automatically correct because obviously if you build the logic wrong, it's going to repeat that same mistake perfectly every single

3:57 time. But because it's deterministic, meaning predictable, it's much easier to test, to audit, and to fix. So, if a task follows a clear set of rules, you probably don't need an AI agent. Use AI

4:08 when you need judgment, interpretation, flexibility, or the ability to work through messy information. But use normal automation when the steps and correct answer are already known. And

4:16 this is really important because in this hype wave of AI, everyone thinks, "Oh, we have this problem. Let's use AI to solve it." But if you can come in here and say, "You know what? Actually, if we

4:24 didn't use AI, this would actually be cheaper, faster, and probably better." That perspective is really valuable as well. And honestly, a lot of the best systems will combine both. Let's say

4:32 you're creating a weekly business report. A normal automation can pull the numbers from your database, clean them up, run the calculations, and verify that nothing is wrong. And then you can

4:39 have an AI model look at those verified numbers and explain what changed in plain English. The deterministic system handles the facts and AI helps you tell a story from it. That's what augmenting

4:49 an existing process looks like. You take something that already works, break it into steps, and then you improve the specific steps where AI can actually help. Once again, don't use AI to solve

4:57 a problem that a simple automation can already solve. And choosing the right approach gets a lot easier when you understand why you're building the system in the first place. Which brings

5:04 us on to the U of Vault, which is to understand the why. One of Goldman's engineering principles is called build with purpose. Marco Agenti explains it by saying that engineers can't focus

5:14 only on the how. They need to prioritize the why. And I see people do the exact opposite with AI all the time. They open up an AI tool thinking, I want to build an agent, or I want to make a workflow,

5:23 or I want to use this new tool that everyone's talking about, but what's the problem you're actually solving? You can build a really impressive workflow that nobody needs. You can spend an entire

5:30 week getting an agent to work perfectly, and if it doesn't save time, make money, reduce mistakes, or improve some actual outcome, then what's the point? So start with the problem and then choose the

5:38 tool. Before you open an AI tool, just write one sentence. The problem I'm trying to solve is blank. And a good result would look like blank. And if you can't write that sentence, then you're

5:46 probably not ready to start prompting an AI model yet. Because if you don't know what you want, then how is the model supposed to know that either? And if you're still figuring it out, that's

5:53 fine. You can use AI as a brainstorming thought partner. You know, give it the context, explain the bottleneck, and ask it to help you find the simplest way to solve it. This is especially important

6:01 if you're trying to make money with AI. Inside my communities, I've seen people build every workflow imaginable and still struggle to get clients because they started with the tech. I've also

6:09 watched completely new members focus on one painful problem, build one useful system, and replace their job income within months. The difference was the problem that they chose to solve. And

6:17 speaking of the community, if you guys want to learn how to use AI tools and get free resources that help you move faster, then you can join my community for completely free. I've put courses in

6:26 there and I've also put everything that I'm talking about in today's video with the vault framework, a full resource guide around that in my community. There's also stuff in there about how to

6:32 start a business using AI. So, if you want to join, the link for that is down in the description. We're almost at half a million members in there, which is just awesome. So, I'd love to see you

6:39 guys in there. Let's get back to the video. Okay, so now we're moving on to the L, which is loop humans in. Think about AI kind of like a megaphone. Whatever you give it can be amplified

6:46 across an entire workflow, including mistakes. A slightly unclear instruction might give you one slightly wrong answer inside a chat, but once you connect that same instruction to an autonomous agent,

6:55 it could send the wrong email. It could update the wrong record. It could publish something publicly or message an entire client list before you even see it. We actually had an autonomous AI

7:03 agent that looked at a task list, proactively took one off the list, interpreted that wrong, and ended up sending a discount code to almost 200,000 people on our email list.

7:11 Obviously, big mistake. Marco Agenti has warned about this exact problem. A small miscommunication can get amplified by an AI system. And when agents take action, hallucinations can lead to incorrect or

7:20 even dangerous actions. I'm sure we've all heard those horror stories of agents deleting massive databases at massive companies. But anyways, Argenti's conclusion is pretty simple. Until these

7:28 tools are consistently reliable, humans have to stay in the loop. That doesn't mean a person needs to approve every single tiny little action, your level of oversight should basically match the

7:36 consequence of the action. If an AI tool is organizing your personal notes, let it run. If it's preparing something for a client, you know, changing important data or spending money or communicating

7:44 with a large group of people, then add an approval step, please. One of the easiest ways to do this is to set your automations to draft, not to send. Have the AI write emails into your Gmail

7:52 drafts. Have it put proposed replies into a document. Whether you're using cloud code, codecs, or whatever else, have it show you the plan before it changes any important files, deploys

8:00 anything, or takes an action that you can't easily undo. You still obviously are getting most the speed, but one weird output doesn't immediately turn into a much bigger problem. The more

8:08 people an action can affect, the more important the human checkpoints become. A guiding rule that we talk about inside my team is if the agent could potentially do something, then you have

8:16 to assume that it will because 999 times it might do what you want, but that one time it could be bad. And the thing is that human can't properly approve the work if the system doesn't show them how

8:25 it reached the result. Which brings us on to the T which is for transparency. At a company like Goldman, if you build something that touches important data, reporting or risk, then you need to be

8:33 able to explain where the information came from, what happened to it and how the final result was actually produced. They have regulators, clients, managers, and other teams that may need to review

8:42 those decisions. You can't defend something that you don't understand. And you should build your own AI systems the same way. If an agent works, but nobody understands how it works, then it's

8:49 going to be a nightmare when something breaks, when the data changes, or when somebody else needs to take over that project, because you might not know what to fix or which step failed, and you

8:57 probably won't be able to repeat the results consistently. So, ask the AI to document what it's doing as it builds. Every single execution of these systems needs to be logged somewhere, listing

9:06 its data sources, assumptions, tools, validation checks, and any important decisions. If it produces a report, ask it to show which source supports each conclusion. If it creates an automation,

9:15 have it explain what triggers the workflow, what happens at each step, and where a human needs to approve something. And transparency doesn't mean asking a model to reveal some hidden

9:21 internal thought process. What you need is evidence that you can actually inspect the sources, the inputs, the actions, the assumptions, and the checks that it ran. That makes the system

9:29 easier to trust, easier to improve, and so much easier to hand off to another person. Okay, so those are the five principles inside Vault. Verify the output, augment the systems that already

9:38 work, understand the why, loop humans in, and keep the entire process transparent. And I think the biggest lesson that I took from Goldman is that you don't need to make everything an AI

9:46 agent. You need to understand the problem, protect the data, use the simplest system that works, and keep control over the actions that actually matter. That's how you build AI systems

9:53 that you can trust. And if you want to turn these principles into actual projects, my free community has courses and resources that can help you get started. There's a full path to building

9:59 an AI portfolio in 7 days, learning AI tools, and creating your first agents and systems. And if you want more direct help and a path to getting your first client and spinning up your own

10:07 oneperson AI agency, then you can check out my road map in the link in the description. But anyways, that is going to do it for this one. So, if you guys enjoyed the video or you learned

10:14 something new, please give it a like. It definitely helps me out a ton. And as always, I appreciate you guys making it to the end of the video. I'll see you in the next one.