How much should a small business spend on AI? Budget per outcome, not per seat

Owners ask what they should spend on AI and want a number back. There isn't one.

There is a way to decide, and it starts by refusing to treat AI as a budget line at all. Two failures come from asking the wrong question. One is paying for forty seats when six people log in. The other is spending nothing for a year while waiting for a case that will never feel complete.

There is no right number, and asking for one is the problem¶

A percentage of revenue is not an answer. It is a way to feel decisive without deciding anything.

The number depends entirely on what you are trying to remove, and the word "AI" does not tell you that. Ask a different question. Which hour of work do you want back? That has a price attached to it. "How much for AI?" does not.

Stop budgeting for AI as a category¶

A category budget invites spend to fill it. Assign $2,000 a month to AI and someone will find $2,000 of tools. Nobody has to prove any of it worked, because the budget was the goal.

Budget per outcome instead. Each spend attaches to one job, one owner, and one measure. Tools without a named outcome do not get funded.

An AI line item buys tools. An outcome line item buys results.

Start from the hours or errors you want to remove¶

Pick one thing that is costing you now, and be specific:

• Six hours a week of quoting rework
• Invoices going out three days late
• Two client emails a month with the wrong figures

Put a rough cost on it. Six hours a week at a $40 loaded rate is about $1,250 a month. That figure is your context for every decision that follows. If the fix costs more than the problem, you have your answer before you start.

Put a ceiling on what proving it is worth¶

Proving something works is a separate budget from running it. Cap the proof.

A workable default: spend no more than one month of the problem's cost, and give it no more than four weeks. Write the cap down before you buy anything. Write down what "it worked" looks like too, in a sentence someone else could check. Then stop when you hit either limit.

A pilot with no ceiling is not a pilot. It is a subscription with optimism attached.

The costs owners miss are rarely the subscription¶

The $30 a month is almost never the expensive part. Three costs get left out:

• Setup time. Someone writes the prompt, connects the data, and tests it. Count those hours at their real rate.
• Review time. Most AI work adds a check step. Ten minutes a day of someone reading output is roughly 40 hours a year.
• Maintenance. Prompts break when a form changes. Somebody owns that, or the thing quietly rots.

Add all three before you compare against doing nothing. Plenty of tools lose on the second line alone.

Cost per outcome is the number that matters¶

Cost per seat tells you what you are billed. Cost per outcome tells you what you bought.

Divide total cost, including review and maintenance, by the thing you removed. Four hundred dollars a month to remove six hours a week is about $15 for each hour returned. That is a decision you can actually make. Track it with a weekly AI scorecard so the number stays honest after the first month.

Per-seat licensing turns into sprawl when nobody owns the renewal. Seats get added for a trial and never removed. Assign one person to each tool's renewal date and to its cost-per-outcome number. If neither has a name next to it, cancel it.

Keep exploring¶

To see what you are already paying for, start with a 20-minute AI tool audit. For a full picture of where spending would actually pay off, start the AI Readiness Audit or contact FIT.