Two earlier posts covered how much an AI rollout costs and where the money goes and whether to build or buy the tools. This one is for the owner or finance leader sitting down in November to write next year’s plan. It answers the question in the order a budget gets built: what the lines are, how big each one tends to be, who owns it, how to protect it from the pilots that fail, and when to spend more.
Why does a single “AI” line fail?
Because a single line cannot be defended. When revenue softens in the second quarter and the leadership team looks for cuts, a line that says “AI: $X” has no process behind it, no baseline, and no owner who can say what disappears if it goes. It is the easiest cut in the room.
A line that says “proposal drafting: subscription $A, integration $B, training $C, baseline 9 hours per proposal, target 3” is a different conversation. Cutting it means proposals go back to taking nine hours. Somebody has to argue for that out loud.
The second reason is that a single line hides the real shape of the spend. In most rollouts the software is the smallest piece. People are surprised by that, and the surprise shows up as an overrun in the implementation line that nobody planned for.
What are the five lines, and how big is each one?
The proportions below are a planning shape drawn from how mid-market rollouts tend to go, not a rule. Your own numbers come from the processes you choose.
| Line | What it covers | Typical share of year one |
|---|---|---|
| Software subscriptions | Per-seat licenses, usage-based model costs, the vendor platform for each process | 15 to 25 percent |
| Implementation and integration | Connecting the tools to your CRM, document system or phone system; configuring workflows; outside help | 25 to 35 percent |
| People and training | Time of the internal owner, training sessions for the team, the hours staff spend learning instead of producing | 20 to 30 percent |
| Data and security | Cleaning the data the tools will read, access controls, vendor reviews, the one-page usage policy | 5 to 10 percent |
| Reserve for failed pilots | Money you expect to spend on things you will shut down at 30 days | 15 to 20 percent |
Two things stand out. Software is rarely the biggest line, and the reserve is not optional. A budget with no reserve forces every pilot to succeed on paper, which means weak pilots get kept alive to avoid admitting the loss. The reserve makes it cheap to be honest.
On accounting treatment: most of this is operating expense, but some implementation and custom development work may be capitalized depending on your standards and how the contract is written. Settle that with your accountant before approval so the first invoice is booked the same way as the last.
How do you tie each line to a process?
Start with the processes, not the tools. Pick two or three that are expensive, repetitive and measurable today. Intake, proposal drafting, collections follow-up, first-draft contracts and meeting summaries are common first choices because they already have a number attached: hours per item, days to respond, dollars outstanding.
For each process, write four things on one page:
- The baseline. What it costs now, in hours and dollars, measured over the last quarter. If you do not have the number, spend two weeks collecting it before you spend anything else. This is the same discipline described in how to measure the ROI of AI in a mid-market company.
- The target. The number you expect at 90 days. Be conservative. A target you beat builds the case for year two.
- The five lines for this process. Software, implementation, training, data, reserve.
- The owner. One person who runs the process today and will run the changed version.
Add the pages together and you have the budget. The total is usually smaller than the leadership team feared, because two or three processes is a bounded piece of work. It is a transformation program spread across every department that produces the frightening totals, and a mid-market company should not be funding one of those in year one.
Who should own the budget?
One person. Usually the chief operating officer, or whoever already owns the processes on the list, with the finance leader reviewing. The temptation is to give each department head a slice. That produces four departments buying four tools that do the same thing, nobody accountable for the total, and a year-end review where every slice looks fine and the company cannot say what it got.
The budget owner does not need to be technical. They need to run the scoreboard: the baseline, the target and the actual for each process, reviewed every quarter. When five of us used AI-driven systems to generate $32 million in revenue, the thing that kept the spending disciplined was not technical skill. It was that one person could see every number on one page and shut things down when they did not move.
How do you handle the pilots that fail?
Plan for them. Roughly a fifth of what you try in year one will not produce a result you keep, and the budget should say so in writing. That does three useful things.
- It makes the 30-day decision easy. Every pilot gets a checkpoint at 30 days. If the baseline number has not moved, the pilot stops and the remaining money returns to the reserve. No budget fight, because the loss was already planned.
- It protects the pilots that work. Money that would have been trapped in a struggling pilot goes to the one that is beating its target.
- It tells the team the truth. People try harder on a pilot when they know shutting it down is a normal outcome and not a personal failure.
Record what each failed pilot taught you in one paragraph. That paragraph is the cheapest research you will buy all year, and it shapes what goes on the list for year two.
How often should the budget be reviewed?
Approve the annual envelope once. Release it quarterly. Tool prices change, model capabilities change, and your own results will tell you within 90 days which processes respond and which do not. A twelve-month plan written in November cannot know any of that.
The quarterly review is short. The budget owner brings the scoreboard: for each process, baseline, target, actual, spend to date. The leadership team decides three things: which pilots continue, which stop, and whether the next quarter’s release goes up or down. Twenty minutes, four times a year, is enough to keep an AI budget honest.
When is it time to spend more?
When the scoreboard says so, and not before. Three signals in particular:
- A process beat its target by a wide margin. Fund the next process that looks like it. Collections responded, so try invoicing follow-up.
- The team is asking for more. When staff who were skeptical in January are proposing their own uses in June, the training line has done its job and the bottleneck has moved to the implementation line.
- The reserve went unspent. If few pilots failed, your list was conservative. Add a harder process next year.
The opposite signals matter too. If the implementation line overran by half, the processes were less ready than they looked, and the data line was too small. If training was underspent, adoption is probably lower than reported. Spend on the lines the numbers point to, and resist the pull of a new tool that nobody has tied to a process.
If you want the processes chosen, the baselines measured and the budget built as a 90-day plan before you commit the money, the engagements and FAQ page explains how the AI Clarity Sprint works. The prompts for turning a vendor quote into a five-line budget page are in ChatGPT for Profit, Volumes 1 and 2.
Frequently asked questions
- What percentage of revenue should a mid-market company spend on AI?
There is no reliable percentage, and a company that budgets from one is budgeting backwards. Start from the processes you intend to change, estimate the cost of changing each one across the five budget lines, and add them up. The total is usually far smaller than a percentage-of-revenue rule would suggest, because a mid-market company should be funding two or three specific projects, not a transformation program. This is general education, not financial advice.
- Should AI spending be capitalized or expensed?
Most of it is expensed, because subscriptions, training and outside help are operating costs. Some implementation and custom development work may qualify for capitalization depending on the accounting standards you follow and how the project is structured. Ask your accountant before the budget is approved, not after the invoices arrive, so the treatment is consistent from the first month.
- Who should own the AI budget in a mid-market company?
One person, usually the chief operating officer or whoever owns the processes being changed, with the finance leader as the reviewer. Spreading the budget across every department head means nobody is accountable for the total and every department buys its own tool. A single owner with a single scoreboard keeps the spend tied to results.
- How much should be set aside for AI pilots that fail?
Plan for roughly a fifth of the total to produce nothing you keep. That is not waste. It is the price of finding out which processes respond to the tools and which do not. The reserve lets a leader shut down a weak pilot at the 30-day mark without a budget fight, which is what keeps the money flowing to the pilots that are working.
- Should a mid-market company budget AI annually or quarterly?
Approve the envelope annually and release it quarterly. Tool prices, model capabilities and your own results change too fast for a twelve-month plan to hold. Each quarter, the budget owner shows what the previous quarter’s spend changed on the baseline numbers, and the next quarter’s release is adjusted up or down from there.
Want next year’s AI budget built on measured baselines?
The AI Clarity Sprint gives mid-market companies, law firms, wealth managers and insurance agencies an AI audit, deployment recommendations with ROI models, and a 90-day plan in three weeks.