Elon's AI Agent Has a Massive Cost Problem

Elon Musk just said Grockbot costs too much to serve for the money it brings in. At first glance, this looks like Grockbot is in trouble. Now, I know what you're thinking here. If the richest men in the world can't afford to run agents, what hope do the rest of us have? But look at the wrinkle. XAI has been publishing writeups on how it uses Grockbot inside the company. And part of what it's talking about is saving money. It says Grockbot found 43 paid software seats that nobody had used for 90 days and downgrading them saved over $14,000. So the agent is clearly doing economically valuable work at XAI and yet it can still be too expensive to serve to customers at scale. How are both true? Because the value depends on the work you give the agent. And you're going to have to decide what your agents work. A customer had burned through their Grockbot allowance with 4 days left before the reset. They were unhappy. Elon's response was very, very direct.

The cost of serving Grock, quote, was unsustainably high relative to revenue. Okay, fair. But what was the bot doing? Because if it spent 3 hours rearranging a dock nobody needed, yeah, we have our problem that the bots are not being used well. But if the bots are finding things that saves the company tens of thousands of dollars, we would want to know what else we could give it and we would pay whatever it take. So same usage limit, very different conversation depending on the economic value the agents are generating, which is the hard thing right now. And you don't have to run a business for this to matter. Think about a report that might eat up your week. The project that keeps getting delayed because nobody has time to work through it. I've been there. There's work in there that could be worth a whole lot more than AIA costs if you can frame it correctly. So, in this video, I want to show you how to dig out that work, how to tell whether the agent actually added value, and when giving it more time and money is a decision that makes sense.

Because getting really efficient at work that did not need doing is a great way to get into a position where you're like, "Yeah, but the money I spent on Grockbot wasn't worth it, and so I'm just walking away." which some people are doing in response to Elon's tweet and other people are saying I will pay what it takes because this is valuable to me and that is not just a rockbot story that is a story about agents writ large that is a story about many other agents in the market today including pay as you go agents as well let's start with a meeting that you can probably already picture let's say it's Thursday we're going to decide whether the project is ready have you been in a meeting like that I have the plan is in a dock somewhere the dates change there's a ton of AI I slot volume around this. Somebody raised a problem last week. Nobody knows whether it got resolved. Two or three agents went to work on it. There's a spreadsheet everyone keeps on referring to. Maybe you haven't read it. Maybe your agent has. Uh maybe there's two competing spreadsheets and they disagree. You get eight people into the meeting and 45 minutes later you've agreed that someone should find out what's happening and that's a ninth person that isn't there.

I would be very very interested in paying an agent to help me avoid that [clears throat] mess. Now, you could start this by asking AI to summarize the docs. Sure, that's a very 2025 response. Now, you've got yet another doc that you have to bring to the meeting that's AI generated along with 10 other ones. But think about what you need to walk into the room with. What's actually ready? What's not ready? Which unresolved thing is actually going to change the date? Give your agent the context it needs to cut the slot, to show you where dates conflict, to show you what matters in this project when it finishes. If you can't tell, don't assume that it works because the summary looks nice. What I'm saying is that if you can't fight with your agent a little bit and tell your agent, "This is where it works, this is where it doesn't. These are my constraints," you're unlikely to get that disproportionate value out of the agent that would let you avoid the meeting. Now, let's tell a more positive story. Let's say you do give it that clarity. And imagine your agent comes back and says, "You know what? The two teams are fundamentally misaligned.

They're working on different dates. this is the set of discussions they've had. They're not aligned on their inputs. Well, you can run something down, but you avoided a meeting. You avoided the confusion. You used the agent to cut through the cloud that goes through most corporate decisioning and you moved the ball forward on the work. What I find overwhelmingly when I see these situations is that the person using the agent knew they needed to use the agent as a clarifier, not a complexifier. And so often I think people take this idea that the agent can read faster than me and they end up turning it into an amplifier of complexity rather than a cutter through something that drives clarity. It drives me nuts, guys. I've seen it over and over and over again. You have to push for clarity. And if you don't push for clarity, you're not going to get your money's worth. And when people tweet things like the agent cost too much to run, like Elon did, you won't have a clear picture of the value.

And so, if you want to judge the result, it's like, did it find a real problem that I would not have noticed otherwise? Did it help me resolve something a whole lot faster that would have held up the work? how much of my time went into checking and fixing the answer, going back and forth. You don't need to invent dollar values here. I want you to have something tangible that said, "Wow, this wouldn't have happened without the agent." I want you to be able to point to something and say, "The agent had value." And so often what I find is that people who are able to do that are people who have the instinct to be clarifiers. And look, especially if you're in a company, be honest. And don't straw man. Be honest about the value that your human colleagues bring to the table. Don't just say the agent did the whole thing when in fact the human did most of the work. The human was the one that finalized the budget and the agent did not do much of that work other than figure out some of the misalignments. Be clear about what your agent is doing, where you're applying human judgment, and what the difference is. There's just no other way to do it.

Not if you want to add value. All right, let's say we're back in report land. Let's say you inherited one. I've inherited reports. Have you ever inherited reports? Somebody else did it before. Now it's your job forever. Let's talk about that. Let's say every week I've had to do this. It's like a giant file. You download some things. You copy over some numbers. You fix the formatting in the spreadsheet on tab 37. You figure out why the totals don't match. Maybe there's a staffing issue or a late project. Whatever your industry is, you know where the painful part is there. and it's the same problem every whatever day, Friday for me or whatever day of the week it is. An agent can help you solve that pain and that's a perfectly reasonable place to start. But I'm going to push you on value again. There may be a bigger opportunity here. What would it take to build a tool that handles the repeatable parts of that report every week? What if you gave the agent previous input files and you said these are the business rules I'm keeping in my head and when a project moves this is why I don't count it twice and all those little things that only you know then test it by having it build an old week report where you know the answer and make sure that it's like a tough challenge etc. And find out if your agent can lift that pain for you entirely. find out if the agent can take it to done entirely. That's a huge lift

for teams because teams do not like the pain of doing something that they know is not very valuable. And so much of the time when we pick these things for agents to automate, we're picking things to automate that teams don't really care about. Again, I point back to value. I say, "Where's the value this agent is generating? Suppose this report took 20 minutes before and no one cared about it and it takes two minutes now. It's not going to really be measured as value, even if technically it is. But if it took the team distributed a bunch of pain on a Friday and they're all writing little bits of it and now you take it and you say, "No, we're going to have the agent go after it." Now you've got a map to value. And I bet if you work at a company with over a hundred people, you can find your own version of that with unused seats. I guarantee it. Go have it look for seats that have not been used in 30 days and see if you can get rid of them. You're going to find seats that haven't been used in 30 days. It's just going to happen. That being said, when you are looking to paint the picture of value, make sure you don't just stop at the simple, we found the seats for value. We can measure the direct dollar cost. Some of the examples I've given earlier in this video, they're a little bit harder to value, but intangibly teams are going to point to them and say, "My report that I had to spend time on every week, I just am heads down doing my actual job now, and I'm not worrying about that report because it just gets done." Don't write

off those pieces of value. I know they're not flagship value, but they're value that lifts up teams and helps teams to focus. That's the difference I've seen between people who are taking agents and using them to produce more slop faster and people who are taking agents and saying, "No, I have a direction. I have a goal. I have a north star. That's what I'm going after." And maybe it's I can like clean up my ops in my company and I end up, you know, taking a bunch of seats and cutting those down. And maybe it's reporting and this and that. Or maybe it's upside and maybe I'm working on uh sales and customers and revenue. But either way, I'm obsessed with the net value. And look, I I can hear the objections here. This sounds like more work than just telling the agent to do the work. Yeah, it is more work. It's more human work. It's the kind of work where a little effort now is going to keep paying you back because you set up your agents and they're running in the right direction.

Another space that I would look at is find a spot where everyone has opinions and nobody has really built a solution in the company. Find a way to help your agent build a rough version really quickly and give it to a few of your colleagues. Can they use it? Is it useful? Cycle on that. Make a small version that people can try. It's like you're a tiny internal startup, right? And you're making a little MVP and you're able to deliver that value quickly. And once you've decided what your learning goal is, how much allocation of time and money that question deserves internally, then that experiment can show you really quickly if your idea has legs. If you can actually get something that has value internally as a result and iterate on it to keep going. This can be a really useful approach where you don't have clarity on what the solution shape should look like. You just know it kind of is painful. So, you've got several places to dig in. Now, if you're trying to figure out where are agents that are worth paying for, look at work that keeps people waiting and bottleneck, look at work that you keep doing by hand. Look for an idea that you can test before making a bigger commitment. Pick a particular job and ask what getting it all the way done would unlock for the business or for you personally. And then focus your agent on getting that done.

Now, at this point, of course, we come back to Elon and the bill. If you include the AI usage and you include all the extra tools, if you include the fully loaded cost, now you can decide, is that actually worth it to you? Is saving two hours worth 10 bucks to you? Is saving three hours worth 50 bucks you? I don't know what your numbers will look like because I don't know the model you're using and I don't know how you value your time. Both of those are going to become some of the most important metrics in the world. You have to understand the value of your time and how you would use it. There's no substitute for that. And I know that we don't always think that way, but you're going to have to. And we also have to understand what is our value for agent cost. Do we value it as a sunk cost because we understand that the subscription gets paid for every month and so if it's under the usage, we don't care. Do we value it API costs? Do we value it at the cost of a cheaper model if we can trade down to that? Figure out what your value system is and take it seriously. And I'm not kidding. I know that feels like work, but even having a working rule of thumb, saying, "I value my time at X an hour, and if I can get an agent to do this for this much, it's going to be worth it." is a huge unlock.

There's no substitute for it because you can't really make progress without deciding what your ROI is going to be. How else are you going to get the agent to be a clarifier? Otherwise, you're going to be in a position where you don't know what your break even is. You don't know when the agent is worth doing the job. You don't have clarity. You don't have a north star. You don't have direction. You need to come up with evidence for what counts as success in your job. This is another part of one of the larger themes I've talked about a lot on this channel, which is that these things that were previously the job of companies and CFOs and people who are very high-powered are everybody's job now. Everybody has to think about the time value of what they're doing. Everybody has to think about whether it's worth it for agents to do that job. And the better you are at that, the more leverage you have. The people I know who are incredible at deploying agents understand the value of their time, understand the value of what the agent is doing. They have run the numbers on the exact model they're using and they say this model doing this work is absolutely worth it. And this is why I know that if you can't say that, it's going to be really hard. And if you're trying to figure out like how do I value my time? How do I value my agents? You can literally tell your AI to interview you and it will help you build a model for that. And if you get stuck, I absolutely have a Substack for you that will help you out that will go with this that will help you walk through that process and develop time value so you

can be clear with your agents about the value. I firmly believe that if we use agents to complexify, there is negative time value. We are wasting time with agents and wasting money. And this brings us back to Elon. Finishing a difficult assignment can be real, real real hard. Understanding the value of agents can be really, really hard when they're solving problems that are nonlinear like that. I expect we are moving into a world where if an agent can do that kind of nonlinear, I can do a report and go after it. It will take time and I don't know how long it will take kind of a job, but it's really valuable. and that same agent can do a useless rewrite of an email, we're going to see really, really different agent unit economics and we're going to see much more precise agent tiering because people are not going to want to pay for the stuff that isn't worth it for an agent to do. People are going to want to pay extra for the stuff where they get extraordinary value. And then the next time Elon says, "I'm cutting usage for Grock because it's too expensive and didn't pay for itself," we'll have a good enough idea of where to go with this problem that we can go to other agents and solve it because we know what our budget is, right? We can say, "Okay, well then I'm going to do, then I'm going to Muse, then I'm going to Hermes, whatever it is, and we can solve it from there." Every single person I know who is AIDS, who is deep on AI, has that

mindset. You should have it, too.