Founders ask me how to budget for AI more than they ask me almost anything else. It usually arrives mid-conversation, phrased as an aside, like they are slightly embarrassed not to already know the answer.
Nobody knows the answer. The category is barely three years old and most of the available advice is either vendor marketing or confident guessing.
So here is what I actually tell them. Thirty percent of your AI budget goes to building internally. Fifty percent goes to buying best of breed tools. Twenty percent stays unallocated, reserved for work you cannot justify yet.
Then I tell them why that split might be wrong for their company. That second part is the one worth reading.
Say the quiet part first: this is a heuristic, not a benchmark
The 30/50/20 split is my rule of thumb. It is not an industry standard, and I am not going to dress it up as one.
Before writing this I went looking for real data on how companies divide AI budgets across building, buying and experimenting. There is no credible public source that gives a three way split for companies under about five hundred people. Not one. The research that does exist measures adjacent things, uses incompatible definitions, and in a couple of places points in the opposite direction from my own advice.
I would rather show you that than pretend the number fell out of a dataset.
The fifty percent you buy is the easy part
Buying is the default, and it should be. Menlo Ventures surveyed 495 US enterprise AI decision makers in November 2025 and found that 76 percent of AI use cases were purchased rather than built internally. The year before, the split was almost even at 53 percent purchased and 47 percent built.
That is a fast reversal, and the reason is not mysterious. In 2024 a lot of companies built because the tools they wanted did not exist yet. By 2025 they did.
There is a second finding in that survey I think about more than the headline. Menlo found that 47 percent of AI deals reach production, against 25 percent for traditional software. Bought AI tools actually get used at roughly double the rate of bought software generally. Whatever your instinct about shelfware, the evidence is that AI purchases stick better than the average SaaS purchase.
So why not buy everything?
The thirty percent you build is where your edge lives
Because the economics of building changed underneath all of us.
McKinsey's 2026 State of AI survey, which drew 1,719 responses across 97 countries and weighted the results by each country's share of global GDP, found that 32 percent of organizations have decided against buying at least one software product because they could build it internally with agentic coding tools. Among the companies McKinsey classifies as AI high performers, it is close to half.
Read those two findings together and they look contradictory. Menlo says buying won. McKinsey says building is coming back. They are not actually in conflict, because they measure different things. Menlo asked what share of use cases get bought. McKinsey asked whether you ever declined a purchase. A company could buy ninety percent of its AI stack and still answer yes to McKinsey.
I should give you the strongest argument against my own position here. a16z's survey of 100 enterprise CIOs found that companies are moving from build to buy specifically because internally developed tools are hard to maintain and frequently do not deliver a business advantage. That is a real pattern and I have watched it play out. Most internal builds are worse than the thing you could have bought, and they keep costing you after launch.
I hold thirty percent for building anyway, and the reason has nothing to do with cost. It is that a vendor can only sell you what it can sell everyone else. If the thing you are automating is how your company specifically wins, buying it means buying parity. Build where the workflow is the edge. Buy everywhere else, which is most places. The a16z finding is a warning about building the wrong things, not an argument for building nothing.
The twenty percent is the part I get argued with about
This is where the evidence pushes back hardest on me.
The same a16z survey found that innovation budgets, the money set aside for open ended experimentation, fell from 25 percent of large language model spending to just 7 percent in a single year. The market moved decisively away from what I am recommending.
I still think twenty percent is right for the companies I talk to, and here is my argument.
Seven percent is what you spend on experimentation once you already know what works. Those CIOs run large organizations that have been through two or three cycles of this. They found their use cases, they funded them, and the experimentation line shrank because it had done its job.
If you are a fifty to three hundred person company, you have not been through those cycles. Your experimentation line is not a luxury you trim once you are disciplined. It is how you find the thing you will eventually put the other eighty percent behind.
There is a finding in Deloitte's latest State of AI survey, which reached 3,500 director to C-suite leaders across 25 countries and six industries between May and June of 2026, that makes the case better than I can. Eighty one percent said AI had improved the speed of their decision making and execution. Only 25 percent said they were using AI to deeply transform the business, meaning new products, reinvented core processes, or a genuinely changed business model.
Read that as a budgeting statement. The money that buys speed is working. Four fifths of companies got exactly what they paid for. The money that changes what the business actually does is barely being spent.
Speed is what the buy bucket gives you. Transformation is what the build and experiment buckets might give you, and almost nobody is funding them seriously enough to find out.
Cut the twenty percent and you have not saved money. You have opted into being one of the 81 percent who got faster at the work they were already doing.
What the split is really for
Here is the thing I believe more strongly than the specific numbers.
The ratio matters less than the act of classifying. Most companies do not have an AI budgeting problem. They have an AI categorization problem. Spend accumulates across a dozen tools, a few API keys and someone's personal subscription, and nobody has ever asked which bucket any of it belongs in.
Three questions sort almost everything:
Does this encode something only we know? Build it.
Is this a real capability that a dozen vendors already do well? Buy it, and stop deliberating.
Do we genuinely not know yet whether this matters? Fund it small, time box it, and put it in the twenty percent where it is allowed to die quietly.
If your mix comes out at 20/60/20 or 40/40/20 after you run that exercise honestly, you have done the work and my numbers were never the point.
What this actually costs
The spread here is wider than most people expect.
Ramp's AI Index, which is built from real card and bill payment data rather than survey recall, put the median firm that pays for AI at $16.32 per employee per month in October 2026. The top ten percent spent $800.84. The top one percent spent $10,199.20.
Two caveats before you measure yourself against that. Ramp is reporting on its own customers, which skews toward US technology companies. And the denominator is every employee, not just the people using AI, so a thousand person company with twenty seats shows up as a tiny per employee number. It is not a seat price.
What it does tell you is that there is no normal. A 625 times gap between the median and the top one percent is not a distribution with a sensible middle.
One more number worth having. Menlo's 2026 survey found that 55 percent of AI spending is net new money rather than reallocated from existing software or labor budgets. That is the one I would flag to a CFO. More than half of this is additive. It is not paying for itself by displacing something yet.
The line item nobody budgets for
Which brings me to the part that actually determines whether any of this works.
Deloitte asked those same 3,500 leaders about cost visibility. Seventy one percent track their total AI inference and API spend. Only 51 percent track token volume. Only 44 percent track per unit inference cost.
Sit with the gap between the first number and the last. Most companies know what AI costs them in total. Fewer than half can tell you what any individual piece of it costs, which means they cannot tell you what any individual piece of it returned. Deloitte's global generative AI lead, Costi Perricos, put it plainly: that gap is where ROI can get lost.
McKinsey found something related in its 2026 survey: 20 percent of respondents said AI operating costs, including token costs, had actually constrained their use of AI. That is a real budget problem arriving for companies that did not model consumption pricing.
You have probably seen the statistic that 95 percent of AI initiatives fail. It comes from a preliminary MIT working paper, it is widely misquoted, and what the paper actually said is that 95 percent of organizations were getting zero measurable return. That is a different claim, and honestly a more damning one. Not that the projects broke. That nobody could show they were worth anything.
A budget split will not fix that. Measurement will.
What I would do on Monday
Pull every AI related charge from the last ninety days, including the ones on personal cards. Most companies are surprised by this step alone.
Sort each line into build, buy or experiment. Argue about the ambiguous ones. The argument is the valuable part.
Set the ratio you want for next quarter. Start at 30/50/20 if you have nothing better, then move it based on what you learn.
Decide now how you will tell whether any of it worked, before you spend the money. This is the step everyone skips, and it is why only 44 percent of those Deloitte respondents can see their own unit costs.
That last step is the one we built Hatch for. Your headcount sits in an HR system. Your agent spend sits in a cloud bill. Nothing connects what you spent to what your team actually produced, which is why most companies can tell you their AI budget to the dollar and cannot tell you whether it bought them anything.
Budget however you like. Just make sure you will be able to answer that question in ninety days.
.png)