How much should a small business actually spend on AI?

There is no correct AI budget, because a budget only makes sense priced against work. Start with the job, not the tool: name a task that eats real hours every week, put a payroll number on those hours, and let the spend be a fraction of what the work already costs you. A $30 subscription that saves nothing is expensive; a system that reliably removes a ten-hour weekly job is cheap at several times the price. If you can’t name the job you’re buying back, your correct AI budget, for now, is zero.

Why does nobody answer the AI budget question straight?

This isn’t answered straight because there’s a lot of reasons people aren’t going to name it. Firstly, everyone with an answer is going to try to sell you something. That’s just how the world works nowadays. Whether it’s an AI SaaS tool or all the way down to a full-blown service. Nobody’s going to tell you the real price of setting anything up yourself because they want to sell you something. They’re going to market it as, hey, this solves whatever problem you’re having with AI, or this is how to easily get AI running for you for your business as fast as possible, and telling you the price of what it would actually take to set something up like this isn’t in their best interest of making money. Which you probably understand, people have to make their livings.

The other answer is really “it depends”. I know this is a useless answer and doesn’t really tell anybody anything. A conversation usually dies after this is answered. However, it really does depend on a lot of different variables. Even when we look at it being done by you, the business owner, it can vary a lot as well. It could be as little as 20 bucks a month for a regular Claude subscription to use Claude Code, and it can go up to a few hundred dollars a month depending on how many external services you pay for. It can even go up to thousands a month if you hire somebody to do it for you.

The real problem is that budgeting for “AI” as a category is similar to the previous budgeting for “software”. Budgeting for AI at face value means nothing because the category doesn’t tell you anything about what you’re buying. This could totally mean just the Claude subscription as your AI budget, but it can also be budgeting for different AI SaaS products or tools or different AI services as well.

And so while there isn’t really an answer to the AI budget question, it isn’t really the question you want to be asking. When looking at implementing AI in general, you want to be looking at what you are actually buying back.

What are you actually paying for when you buy AI?

There are three layers for paying for AI. They scale up in pricing, but they also scale downward in the amount of effort you need to actually get results out of them. The first layer of paying for something is the subscription. You might pay 20 bucks a month for a capable model such as Claude or ChatGPT. This will get you access to the coding agents such as Claude Code and Codex to actually get tasks done for you. But this means you actually have to do the research on how to fully capitalize on them. Then you have to build yourself. And you have to maintain everything yourself. If you’re willing to put the work in to get the AI running, this is highly valuable for 20 to 100 bucks a month.

Layer 2 is paying for setup. This is paying somebody to set you up with a specific system and teach you how the system runs on your business. Paying somebody knowledgeable on AI will give you a much better head start in actually running it for yourself. Once they’ve finished their special setup for you, they’ll teach you how everything works. When you’re done training with them, you should feel comfortable enough to build valuable agents and automations for your business. But this still leaves you doing the majority of the work afterwards yourself. Though that work will be much easier to accomplish because you got trained first. However, this still leaves you maintaining it after they leave.

Layer 3 is the most expensive. Usually it’s a monthly retainer, but you’ve hired the person to build and maintain everything for you and you have to worry about nothing. You don’t need to learn AI, you don’t need to keep the automations alive, monitored, and constantly improving, all while you get the benefit of having AI in your business. This is the most expensive because this is the most hands-off. You’re hiring an expert in the AI space to handle everything for you and you need to worry about nothing.

The observation that I’ve made is that most founders only ever budget for layer one, and they dont even get the most out of it. A lot of entrepreneurs have subscriptions to ChatGPT or Claude, around 89% of small businesses, but they’re only using it for the basic back and forth work with a regular chat model. They have no idea about the coding agents and so they never set up anything for themselves to actually get work done on a regular basis. And so, if their spend produces anything, it’s usually very little compared to the output they could get.

Sources: U.S. Chamber of Commerce Foundation — What Small Business Owners Say AI Is Actually Doing at Work

Measure yours in 3 minutes: the free Task Audit scores how much of your recurring work could run without you.

Where does AI spend actually go wrong?

When looking at Enterprise pilots specifically, it was shown in an MIT finding that around 95% of generative AI pilots delivered no measurable P&L impact. If we flip this and look at small business version of it, it looks like the following: Five different subscriptions across the team, everyone is prompting from scratch, there’s no shared business context or database they use, and everything only runs when somebody asks for it.

You can see that there isn’t a system being used, everyone’s doing their own thing, nobody’s working together, and everything is just unorganized. There isn’t a central knowledge base that AI can use to really get good answers, Where a knowledge base can be the defaults that the AI is using to get real business relevant information from. And that last part of only running when somebody asks for it is the biggest one there when every one of the people on the team can be setting up scheduled jobs that actually get real business tasks done for them. We have this universal adoption, again the 89%, but also a near universal waste.

AI isn’t actually fixing any of their problems. It might help them get an idea of how to and make it slightly faster. But in terms of actually fixing pain points and doing tasks for them, there is no actual AI impact here. We saw the 95% of pilots failing. However, when we look at the 5%, the ones that actually succeeded bought actual focus solutions for specific workflows. Nothing broad. So, we know it can work. The issue is just actually focusing on what it’s supposed to solve.

Sources: Yahoo Finance — MIT report: 95% of generative AI pilots at companies are failing

How do you size the number?

If you pay for any tool on the market, it’ll tell you exactly what it costs. Could be 50 bucks a month, for example. But only your payroll can tell you what it’s worth. It’s the price of the job that it’s fulfilling, not the software.

I’m going to lay out a scenario for you. Let’s take a look at the cost of hiring a manager. Their base salary is around $65,000 a year. But you also have to spend on payroll taxes, benefits, insurance and compliance, equipment and tooling, and some overhead. You’re paying them 65k a year, but the overall cost runs at around 1.25 to 1.4 times their base salary. So in reality, it’s going to cost you anywhere from $81k to $91k for that employee per year. The manager’s job consists of around 16 hours a week of coordination work on status meetings, re-explaining, and searching, and the rest is spent on actual judgment work.

You are stressed with the amount of work you’re doing. You’re doing all of the coordination on top of all of the judgment work and all of the other things in your business. So you need to decide whether or not it’s a good idea to actually hire a manager to handle some of your work. You do an audit of your work and you find that on average you take around eight hours in meetings, four hours re-explaining things to VAs and employees, and three hours searching for where things stand. So you yourself are losing those same hours to coordination work as well. You don’t really have a problem with the judgment work. It’s more of the coordination of your team. So even part-time, maybe you hire somebody for 16 hours a week rather than a full hire at $25 to $30 an hour. It’s still $20,800 per year.

But you’re also considering AI. A do-it-yourself system could definitely be set up here on a $20/month Subscription and a handful of small service fees that you could call $50 to $150 per month at around $600 to $1800 per year all in. You’d have to maintain this yourself, but for a small team, this shouldn’t be too hard. But even if we go up to the layer three version of this, a managed system. This could cost you a few thousand per month. maybe $1500/month. And even at its base layer, $1,500 per month is still $18K instead of the $20K per year. If you have your managed system remove more than just the coordination jobs, say saving 30 hours a week across the team, it’s even more valuable at a few thousand per month.

At the highest level, this is saving you $40k, but even at the lowest level of just a part-time manager, this is saving you probably around $2k for a base system managed for you. The money here isn’t just what we need to focus on though. It’s the actual named job that we want to look at. With the manager example, we looked at the coordination as the actual valuable thing to look at. We were able to name the job and so we were able to give it budget and perceived value.

Sources: In Parallel — Coordination Tax Index 2026, TimeClick — The Real Cost of Hiring an Employee in 2026

When is the right AI budget zero?

You should not be spending on AI if you cannot name the job that it’s helping you solve. Using AI broadly does not actually solve anything and it doesn’t help you in any way and you shouldn’t be spending on it if this is the case. You should not be spending on “marketing”. Marketing is not a task. It’s not specific. And so immediately just saying, hey, I want AI for marketing won’t get you results. You can, however, spend on marketing when you have a specific task. Say you have the need to put out ads on different platforms. And one of your tasks is actually going and creating ad creatives to go send out. You now have a specific task that AI can help with. AI can go search up the best ad creatives in your niche and help you create similar ones branded to your business. The specificity of the task actually allows you to understand what AI can and should do for you, and so the budget is easily justified.

You should also not be spending on AI if the process still only lives in your head. Fix up documentation first because it’s free and it’s the raw material any system or even any hire would need anyway. If your process isn’t set and documented, there’s a good chance that there isn’t a right way of doing it that you’ve found yet as well, and so even before building, you should have something to work off of so AI has the correct way of doing things and that budget isn’t wasted on a failed build.

You shouldn’t be spending on AI if you’re driven by the fear of missing out. It’s totally fine to not have AI on everything yet. You can put AI on pretty much anything digital nowadays, but again, if your process doesn’t work, or it’s not ready yet, or whatever other reason, that means you probably shouldn’t be jumping to AI too “fix” it. Putting AI on your processes does not solve them. It just elevates the issues. And AI on good and working processes elevates them to make them even better.

The first step that you should be taking that costs you zero is to write down how the one task actually works end to end before AI even touches it. And if you’re already using AI, you can use AI in this case where you can have AI create that document for you by interviewing you on how that task works. You can have AI help you create SOPs for every process and have it ensure that it’s asking you enough follow-up questions to get every edge case down and prepare it for turning it into an AI process.

What should your first real spend prove?

Your first real spend should prove that the process that you are giving to AI works and that you don’t need to be the one doing it. And this should be for one job only. Pick the most annoying, most repeatable weekly task in your business. If you have a team, this can be something with real payroll on it to actually see the monetary value behind it, but it doesn’t need to be. Size it against the time and what the cost of hiring for that time would be rather than the actual software budget. How much would it cost to hire somebody to do this versus how much would it cost for AI to run this? Write down the documented process somewhere and start working with AI to automate it. Once you’ve got it fully automated and working, the task should actually leave someone’s plate and it shouldn’t be a dashboard or a demo. The job should actually be gone and handled by AI.

Once you get one down, you get some sort of budget back in your business. The cost of AI should be much smaller than the cost of hiring, So the next automation you decide to make gets funded by the proof of the first one. And so you can see this will continue to compound until your time as the founder is only spent doing the things that really need you. And if you go back to the last issue, this is the fifth layer of an AI operating system, which is the build layer. This is the layer that allows you to work on the business instead of in it. And this is where you get your lifestyle freedom from.

Frequently asked questions

Why does nobody answer the AI budget question straight? Everyone with an answer is selling something, and “it depends” kills the conversation. Budgeting for AI as a category is like budgeting for “software”: the category tells you nothing about what you’re buying. The better question is what work you’re actually buying back.

What are you actually paying for when you buy AI? Three layers: a subscription ($20 to $100 a month, where you research, build, and maintain everything yourself), paid setup and training (someone installs a system on your business and teaches you to run it), and a monthly retainer (everything built and maintained for you). Most founders only ever budget for the first layer, and most get very little out of it.

Where does AI spend actually go wrong? An MIT finding showed around 95% of generative AI pilots delivered no measurable P&L impact. The small business version looks like five subscriptions across the team, no shared business context, and nothing running unless somebody asks. The 5% that succeeded bought focused solutions for specific workflows.

How do you size the number? Price the job, not the software. A part-time hire covering 16 hours a week of coordination runs about $20,800 a year. A do-it-yourself AI system covering the same job runs $600 to $1,800 a year all in, and a managed one starts around the cost of that part-time hire while removing more jobs across the team. Once you can name the job, the budget justifies itself.

When is the right AI budget zero? When you can’t name the specific task, when the process still lives only in your head, or when you’re buying out of fear of missing out. Write down how the task works end to end first. Documentation is free, and it’s the raw material any system or any hire needs anyway.

What should your first real AI spend prove? That one documented process works without you doing it. Pick the most annoying, most repeatable weekly task, size it against what hiring for those hours would cost, and automate it. The job should actually leave someone’s plate, and the next automation gets funded by the proof of the first.

Want to put real numbers on the jobs in your own week? Book a free discovery call and we’ll price them together.

Featured image: Photo by Jakub Żerdzicki on Unsplash