Someone on your board, or your CFO, or your own leadership team asked the question: what's our AI budget for next year?
And you realized you had no defensible way to answer it.
Not because you're behind. Because the question, asked first, has no correct answer. It's the equivalent of asking "what's our construction budget?" before anyone has said what building.
You already know what happens when a number gets pulled from the air. It becomes a mandate. Someone has to spend it. And spending without a scoped play is precisely how companies end up funding science projects that demo beautifully in March and are quietly abandoned by September.
The discipline here is simple and it will feel familiar, because it's the same discipline you already apply to capital expenditure: budget follows a scoped play. Never the reverse. Pick the work first. Size it second. That's the entire method, and everything below is just the mechanics of doing it well.
What are you actually paying for?
Four buckets — and most budgets only account for two of them.
Discovery and strategy
This is the cost of deciding what to build. It is the smallest line item on the page and the one that de-risks every dollar after it.
Discovery is where you find out that the thing everyone was excited about saves eleven hours a month, while the boring thing nobody mentioned saves four hundred. That reordering is worth more than any build.
Build and integration
The one-time cost of standing the thing up and wiring it into the systems you already run — your ERP, your CRM, your ticketing system, the spreadsheet three people depend on.
Integration is usually the larger half. The intelligence is increasingly cheap; connecting it to your actual data and your actual workflow is the work. Worth knowing: AI has meaningfully lowered the cost of building custom software, which has changed the buy-versus-build math for a lot of mid-market companies — see our breakdown of when to own software versus rent it.
Run
The ongoing cost people forget: licenses, usage-based model fees, hosting, monitoring, and the maintenance that keeps it working when a vendor changes an API.
Run cost is never zero. Any proposal that shows a build number and no monthly number is incomplete, and you should send it back.
People and change
Training, the internal owner's time, and the unglamorous adoption work of getting humans to actually use the thing.
This is the most under-budgeted bucket in the mid-market, and it's the one that decides whether the other three produce a return. A tool nobody adopts has the same ROI as a tool nobody built, at considerably higher cost. Name an internal owner, fund a real share of their time, and treat that time as part of the budget rather than something absorbed "on top of the day job."
What does year one actually look like?
A responsible first year has a recognizable shape: a small discovery spend, one bounded build, a modest run cost, and real money for enablement.
Ignore anyone quoting you an industry average as a percentage of revenue. Those numbers blend a bank's data science org with a distributor that bought some licenses. They cannot tell you what your next play costs, and they give a board false confidence in a number nobody can defend.
Here is the shape, in plain terms.
- Discovery is small and knowable. Ours is published: a 2-Hour AI Deep Dive at $2,500, an Advanced AI Strategy & Implementation Plan at $7,500, and a Full AI Transformation Engagement at $13,000. Whatever partner you choose, this end of the budget should be four to low-five figures — not six.
- The first build is bounded on purpose. One workflow, one owner, one measurable outcome. If the scope needs a diagram to explain, it's too big for play number one.
- Run is monthly and modeled from day one. Estimated before you sign, not discovered in month five.
- Change gets a real line. If it isn't on the page, it isn't funded, and it won't happen.
The governing rule: size the first play so it pays back inside twelve months. Not because faster payback is inherently virtuous, but because a first play that pays back inside a year earns you the credibility to fund the second one. A three-year payback is a bet on your own patience, and boards are rarely patient with something they don't yet understand.
If you want to pressure-test a candidate play before you take it anywhere, our ROI estimator will let you run your own hours, rates, and volumes. And the framework for turning that into something a board will actually approve is in how to build an AI business case.
Why do run costs surprise people?
Because usage-based pricing means success increases the bill. That is the whole surprise, and it catches finance teams trained on flat per-seat software.
When adoption doubles, consumption roughly doubles. A CFO looking at a line that grew 140% in a quarter reasonably assumes something is broken. Usually nothing is broken — the tool is being used.
Model it the way you'd model any variable cost. Estimate volume at low, expected, and high adoption. Multiply through. Put all three in the budget, and tell your CFO in advance that you expect to land near the high case if the rollout goes well.
Then tie the variable cost to the variable return. If run cost rises with usage and usage is what generates the savings, a rising bill is evidence the thing is working. That reframe only survives contact with a finance committee if you instrumented it — which is why you decide which metrics you'll measure before launch, not after someone asks.
What should you stop paying for?
Fund a meaningful share of this from consolidation, not from new money.
Nearly every mid-market company we work with is already paying for the budget it needs. It's just distributed across:
- Redundant SaaS. Two tools doing one job because two departments bought separately.
- Licenses nobody logs into. Pull last quarter's active-user counts against seats paid. The gap is rarely small.
- Manual workarounds you're paying people to perform. The re-keying, the reconciling, the copying between systems that don't talk. That's payroll spent on integration you never bought.
We wrote about this pattern at length in the real cost of tool sprawl, because it changes the conversation entirely. "We need $X of new AI spend" is a hard ask. "We're redirecting existing spend from tools we don't use into one system that does the work" is a story your board already knows how to say yes to.
How should you phase the budget?
Stage-gate it. Three asks, each unlocked by the last, never one lump sum.
Gate one: fund discovery. A small, defined amount to identify and size the candidate plays. The deliverable is a ranked list with estimated costs and returns. Cheap to approve, and it's the only spend that doesn't require you to already know the answer.
Gate two: fund one build, against a defined payback. You name the number before you spend. "This costs $X to build and $Y a month to run, and it returns $Z by month nine — measured this specific way." The gate is the payback commitment, not the dollar amount.
Gate three: fund expansion out of proven returns. Play two gets funded when play one has demonstrably paid. Now you're not asking for budget on a promise. You're reinvesting a measured return, which is a fundamentally different conversation.
Boards approve staged asks far more readily than lump sums, because each gate is small and each one is reversible. You're not asking for faith. You're asking for the next increment, with evidence attached. If you want to know what they'll press you on, we mapped it in the questions your board actually wants answered.
The bottom line
There is no correct AI budget for a $10M company or a $250M one. There is only a correct first play, honestly sized across all four buckets — discovery, build, run, and change — and funded in stages against a payback you commit to out loud.
Do that, and the budget question answers itself. You won't be defending a number pulled from the air. You'll be presenting a scoped play with a cost, a return, and a date — which is the same thing you'd bring to the board for a new facility or a new hire.
And if the honest answer this year is "one bounded play and nothing else," that is a perfectly defensible budget. Far better than a large number with nowhere to go. If you're still forming a view on what the plays even are, start with our plain-English explainer on what AI agents actually do in a mid-market company.
When you're ready to size your first play with real numbers instead of estimates, tell us about your business here and we'll walk you through what a defensible year one looks like for your company.