Most founders I talk to about how to budget AI automation as a percentage of revenue have no AI line item. They have a series of one-off decisions instead: $2k here for a thing a mate recommended, a subscription they forgot to cancel, a quote for a build they sat on for four months because it felt like a lot of money in one lump.
That whole problem disappears when you stop treating automation as a series of isolated purchases. You stop asking “can I afford this?” every time and start asking “am I spending my allocation on the right thing?” Those are completely different questions and only one of them has a good answer.
This post gives you the actual percentage ranges by revenue band, how to split the money across three buckets, what to count and what not to count, and how to review it every 90 days without turning it into a project.
Key Takeaways
- Budget AI automation as a percentage of revenue in the 1 to 3 per cent range for most founder-led businesses, and treat that number as a standing allocation, not an approval each time.
- Under $500k revenue, percentage maths breaks down. Use a fixed floor of roughly one week of your own billable time per quarter instead.
- Split the allocation three ways: about 60 per cent to build, 25 per cent to running costs, 15 per cent to a test budget you are allowed to lose.
- A flat fee anchors you to price. A percentage anchors you to capacity, which is the thing that actually limits your growth.
- Count the automation build, the platform costs, and the training time. Do not count general software you were already paying for.
- Review the allocation every 90 days against hours recovered and tasks automated, not against how impressive the tech looks.
- If you are spending under 1 per cent of revenue on this while spending 8 to 12 per cent on marketing, you are filling a bucket you have not fixed.
Why a Flat Fee Is the Wrong Unit of Measurement
Here is what happens when you think in flat fees. Someone quotes you a number for an automation build. You compare that number to nothing in particular, because you have no reference point for what AI work should cost. So you compare it to other lumps in your life: a new ute, three months of a junior’s salary, the tax bill. It feels big. You say you will think about it. Nine months later, you are still doing the task manually, and you have paid for it many times over in your own hours; you just paid in a currency that never shows up on a P&L.
A percentage does something different. It turns the decision from “is this worth $X” into “this is what I have allocated; what is the best use of it this quarter.” You already do this with marketing. Nobody agonises over whether a specific Facebook ad is worth $400. You have an ad budget, the ad comes out of it, and you judge the budget on results at the end of the month.
The second problem with flat fees is that they do not scale with you. A number that felt reckless at $400k revenue is trivial at $2M, and a founder who set their AI spending mentally at $400k revenue is usually still operating off that anchor two years later. Meanwhile, the business got more complex, more people joined, more things route through your head, and the case for spending has got stronger while the mental budget stayed frozen.
The capacity argument
The real reason to use a percentage is that operational drag scales with revenue. More revenue means more jobs, more quotes, more invoices, more customer messages, more staff questions. The admin load is roughly proportional to turnover. So the budget that keeps that load under control should be proportional too. That is the entire logic. It is not clever; it is just consistent.
If you want the sharper version: your growth is capped by how much of the operation runs through you personally. Spending on removing yourself from the operation is spending on your growth ceiling. A flat fee treats that as an expense. A percentage treats it as what it actually is, which is the cost of having capacity.
The Actual Percentage: What to Budget by Revenue Band
Let me give you real numbers rather than a range so wide it is useless.
Under $500k revenue. Percentages fail here. One per cent of $400k is $4,000 a year, which is not enough to change anything meaningful, and the maths tells you to do nothing when doing nothing is the wrong answer. Use a time-based floor instead. Work out what one week of your own billable capacity is worth and commit to that per quarter. For most owner-operators in this band, that lands somewhere between $3k and $8k a quarter, and it is a number you can actually feel, which matters more than precision.
$500k to $1.5M. Two to three per cent of revenue, weighted toward the top of that range in year one. At $1M, that is $20k to $30k annually. This is the band where the Operator Trap bites hardest, because you have enough volume to be genuinely busy and not enough team to delegate to. Front-load it. Spend more in the first year to get the build done, then the number drops as running costs are lower than build costs.
$1.5M to $5M. One and a half to two and a half per cent. At $3M, that is $45k to $75k a year. You have staff now, which means you have coordination overhead, which is where most of the recoverable time sits. The spending shifts from “automate the founder’s tasks” to “give the team a system so they stop asking the founder.”
Above $5M. One to two per cent, but with a different composition. Less build, more ongoing training and optimisation, because the constraint stops being “we do not have the system” and starts being “the system does not keep pace with how fast the business changes.”
Those ranges are not arbitrary. Cross-check them against what you already spend. Most founder-led businesses run marketing at 7 to 12 per cent of revenue. Software and IT typically land at 3 to 6 per cent. Sitting AI automation at 1 to 3 per cent puts it below both, which is about right for something that is still an emerging line rather than a mature one. It also means the question “can we afford it” is almost always a yes if you are honest about the comparison.
The number most people land on
If you take nothing else: two per cent of revenue, reviewed quarterly. It is high enough to fund real work, low enough that a bad quarter does not sink you, and simple enough that you will actually remember it.
How to Split the Allocation: The Three Buckets
A percentage on its own is not a budget. It is a pool. What stops the pool from being spent badly is a split, and the split I use is roughly 60 / 25 / 15.
Build: 60 per cent
This is the money that buys new capability. A lead response system, a reactivation campaign against your dormant database, a phone that gets answered when nobody is free to answer it. Build spending is lumpy by nature. You will not spend it evenly across the year, and you should not try. One quarter you spend most of it, the next quarter almost none, then a big chunk again when the next bottleneck surfaces.
The discipline here is doing one thing at a time. Founders with a fresh budget want to buy three builds at once. Three half-adopted systems produce less than one system everybody uses. Sequence them. Get one working and used, then start the next.
Run: 25 per cent
Platform costs, phone numbers, message credits, the CRM the whole thing sits on. This bucket is boring and predictable and it is the one people forget to budget for, which is how you end up with a build you cannot afford to operate. Assume it grows a little each year as more of the business runs on the system.
Worth noting: running costs for AI are dropping, not rising. The model pricing curve has gone one direction for three years. Budget for what it costs today and expect to be pleasantly surprised.
Test: 15 per cent
This is the money you are allowed to lose. Every quarter, pick one thing you are not sure about and try it. Half of these will not work. That is the point. The 15 per cent exists so that the failure of an experiment does not become an argument about whether AI was a waste of money, because you allocated it as an experiment up front.
Without a test bucket, every attempt has to succeed to justify itself, which means you only ever try safe things, which means you learn nothing. I have written more about how to handle this when a previous attempt went badly in what to do about AI automation after a bad vendor.
What Counts and What Does Not
Budgets get useless when the definition is loose. Here is where I draw the line.
Counts:
- The build cost of any automation, whether done by a partner or built in-house
- Platform and usage costs for anything the automation runs on: the CRM layer, voice minutes, SMS credits, model usage
- Time spent by you or your team on training the system and reviewing its output, valued at what that time is genuinely worth
- Strategy and diagnostic work that decides what to build
Does not count:
- Software you were already paying for that added an AI feature. Your accounting package bolting on a chatbot is not AI automation spending; it is your accounting package.
- Individual ChatGPT or Claude subscriptions for staff. That is a tool, not a system. It belongs in software.
- Hardware.
- Marketing automation that is really just email scheduling with a new label on it.
The reason to be strict is that the loose version lets you tell yourself you are spending 3 per cent when you are actually spending 0.4 per cent on anything that changes how the business runs. I have seen founders count a $ 200-a-month tool subscription and a couple of ChatGPT seats and conclude they were well covered. They were not. Nothing in that list removes a task from anybody’s day.
The hidden line item: your own time
The one people always leave out is founder time spent on adoption. A system nobody uses returns nothing, and adoption takes hours, particularly in the first month. If your allocation assumes zero hours from you, it is wrong by a wide margin. Development is a fraction of the work. Getting the thing actually used is most of it, and that lands on you.
Budget your own hours into the number. If two per cent of revenue buys a build but you have no capacity to adopt it this quarter, you do not have a budget problem; you have a sequencing problem. Move it a quarter.
Reviewing It: The 90-Day Rhythm
An allocation you never review becomes a subscription you resent. Every 90 days, sit down for half an hour and answer four questions.
One: How many hours came back? Not estimated, counted. Which specific recurring tasks stopped landing on a human this quarter, and how long did each take? If you automated quote follow-up and that was 4 hours a week, that is 52 hours a quarter. This is the number that justifies the whole line.
Two: what percentage of recurring tasks now run without a person? List your recurring tasks once, properly, and then track the fraction handled by the system. Most businesses start near zero. Twenty to thirty per cent is where you first genuinely feel it. Sixty to seventy per cent within six months is achievable if you sequence well.
Three: did the spend land where the allocation said? If build was supposed to be 60 per cent and you spent 95 per cent of the pool on running costs, you did not build anything this quarter. Worth knowing.
Four: what is the one bottleneck for next quarter? Not five. One. The next build comes out of this answer.
That is the whole review. It is short on purpose, because a review that takes a day gets skipped. I go deeper on the review structure in the 90-day AI automation checkpoint.
What good looks like in numbers
A finance broker I worked with had 319 dormant contacts his team had written off entirely. Multi-touch AI reactivation across SMS and email reopened those conversations and recovered $49,000. Set that against a percentage-of-revenue allocation and the maths stops being a debate. That was one build, out of one quarter’s build bucket, against a database he already owned and had already paid to acquire.
The point is not the specific number. The point is that when you have a standing allocation, you spend it on the highest-value bottleneck rather than on whatever you can talk yourself into affording this month. The dormant database is almost always sitting there. So is the phone that goes unanswered at 4:45 pm on a Friday.
When the Percentage Should Go Up or Down
Three situations where you move off the default.
Push it higher when you are about to hire. If you are one signature away from adding a person to handle admin or coordination, spend the equivalent of two or three months of that salary on the system first. Not instead of the hire necessarily, but before it. A hire into a business with nothing documented takes three to six months to become useful, and the knowledge leaves when they do. Put the system in first and the same hire is productive in days because the context already exists. This is the single highest-return timing decision in the whole exercise.
Push it higher in year one, then let it settle. Build costs are front-loaded. Three per cent in year one dropping to one and a half per cent in year three is a normal shape, and if your number is not dropping, it usually means you keep starting new builds without finishing old ones.
Pull it back when adoption is behind. If last quarter’s build is not being used, do not fund another one. Spend the quarter on training and adoption instead and let the allocation roll. Under-spending on purpose is a legitimate move. Buying a second system to fix the fact that nobody uses the first one is not.
One more, which is less a rule than a reality check. If your marketing spending is at 10 per cent of revenue and your automation spending is at 0.3 per cent, you are paying to generate leads into an operation that cannot answer the phone. Research on lead response has consistently found that the business that responds first wins the overwhelming majority of the time, and most businesses take hours. Every dollar of that marketing budget is running through a process nobody has funded. Fix the ratio before you scale the top of it.
Conclusion
The decision to budget AI automation as a percentage of revenue is not a finance trick. It is a way of making a decision once instead of forty times. Pick your band, land on something around two per cent, split it 60 / 25 / 15 across build, run and test, be honest about what counts, and review it every 90 days against hours recovered rather than against how clever the technology looks.
The deeper thing this fixes is the one you already know about. Your business works because you work it. Every question, every exception, every decision routes through your head, and no amount of working harder changes that because the constraint is structural. A standing budget is how you fund the way out of it, quarter by quarter, instead of waiting for the year you finally have time to sort it all at once. That year does not arrive. The allocation does, every quarter, whether you use it well or not.
Book a Discovery Call
If you want a second opinion on where your allocation should sit and what the first build should be, book a 30-minute discovery call. No deck, no pitch. We map where your time actually goes, name the bottleneck that is costing you the most, and you leave knowing what to spend the first quarter of your budget on. If that turns out to be nothing, I will tell you that too.
Frequently Asked Questions
What percentage of revenue should a small business spend on AI automation?
For most founder-led businesses turning over $500k to $5M, one to three per cent of revenue is the working range, with two per cent as a sensible default. Weight it higher in the first year because build costs are front-loaded, then let it settle as running costs take over. Below $500k revenue, use a fixed quarterly amount instead, because percentage maths on small revenue produces numbers too small to fund anything useful.
How much does AI automation cost for a small business?
It varies enormously with scope, which is why a percentage is more useful than a price. A single automation like lead response or database reactivation sits in the low four figures to set up, plus a small monthly running cost. A full system covering several functions is a bigger commitment spread across a year. The better question is what percentage of revenue you can commit consistently, then work backwards to scope.
Should AI automation come out of the marketing budget or operations?
Operations, in almost every case. Marketing budgets buy leads. Automation buys capacity to handle them, plus everything else the business does that has nothing to do with lead generation. Mixing them means automation gets cut first whenever marketing wants more room. Keep it as its own line so it survives a bad quarter, and so you can see the ratio between what you spend attracting work and what you spend handling it.
How do I know if my AI automation budget is working?
Track two numbers quarterly. First, hours recovered: which specific recurring tasks stopped landing on a person, and how long did each take? Second, the percentage of recurring tasks now handled by the system. Twenty to thirty per cent is where the difference becomes obvious. If neither number has moved after 90 days, the problem is usually adoption rather than the tools, and more spending will not fix it.
Is it cheaper to hire someone or automate the work?
Depends on the work, but the sequencing matters more than the comparison. A hire into a business with no documented process takes three to six months to become useful, needs managing, and takes the knowledge with them when they leave. Put the system in first, and the same person is productive in days because the context already exists. The answer is usually not one or the other; it is system first, then hire.
What should I automate first with a limited budget?
The task that is both high-frequency and low-judgment. For most businesses, that is lead response or follow-up, because the infrastructure already exists and the loss is measurable. Second most common is anything happening to a database you already own but have not worked. Avoid starting with complex judgment calls, which is where most first attempts fail and put founders off automation for a year.
How often should I review my AI automation spend?
Every 90 days. Half an hour, four questions: how many hours came back, what percentage of recurring tasks now run without a person, did the money land in the buckets you planned, and what is the single biggest bottleneck for next quarter. Annual reviews are too slow to correct a bad allocation, and monthly reviews turn into admin nobody keeps up.
About Octavius
Titus Mulquiney is the founder of Octavius AI, where he builds AI brains and AI workforces for founder-led businesses stuck running everything out of their own head. Twenty years in marketing, ex-Sony product manager, ex-GM Zeal NZ. Based in Auckland, working with operators across NZ, Australia, and the US. Connect on LinkedIn.