November and December are when most small businesses sit down with a spreadsheet and decide what next year costs. AI is now a line in that spreadsheet — and for a lot of owners it is the one line nobody can explain. "Six AI subscriptions, no coherent workflow" is how one owner described it. The money leaves the account every month, the value is assumed rather than measured, and nobody has ever added the whole thing up.
This is a planning guide, not a price list. If you want benchmark prices for individual tools, our breakdown of how much AI actually costs a small business covers the going rates. What follows is how to turn those prices into a budget you can defend in a planning meeting: how much to allocate in total, how to split it, how to handle usage-based billing that swings month to month, and what to hold back for tools that do not exist yet.
Why AI spending gets out of hand
Almost no small business decided to spend what it currently spends on AI. The number accumulated. The pattern is consistent enough to be predictable:
- Per-seat creep. One person expenses a €22 plan. Three months later four people have it. Nobody re-approved anything; the invoice just got bigger.
- Trials that quietly convert. A 14-day trial for a tool used twice becomes a €39 monthly charge that nobody notices on a card statement with 40 other lines.
- Usage-based surprises. API and credit-based pricing means one enthusiastic afternoon of bulk work can triple a month's bill.
- Features you are paying for twice. Your CRM, your helpdesk and your office suite all added AI features in the last 18 months, often included. You may be paying separately for something you already own.
- No owner. When AI spend belongs to everyone, it belongs to nobody, and nobody cancels anything.
A 12-person firm paying €20–30 a seat across three tools is spending €700–1,000 a month — €8,000–12,000 a year — without a single deliberate decision having been made. That is a junior hire's worth of budget allocated by accident.
Start with what you already spend
You cannot budget forwards until you know the baseline, and almost every owner underestimates it by 30–50%. This is a 30-minute exercise, and it is the single highest-value thing in this article.
Pull the last three months of card and bank statements and search the descriptions for: AI, GPT, OpenAI, Anthropic, Claude, Copilot, Gemini, Perplexity, Otter, Jasper, Midjourney, Zapier, Make. Then ask each team member, in writing, what AI tools they personally use for work — including ones on personal cards and free tiers. The free ones matter too, because they are tomorrow's paid ones and today's data-protection exposure.
Record four things per tool: monthly cost, number of seats, who owns the renewal, and what it is actually used for. If nobody can name a specific recurring task in one sentence, that is your first cut. Our guide to auditing your AI tool stack walks through the full version of this exercise, including how to decide between consolidating and cancelling.
If you cannot name the recurring task a tool performs, you are not buying capability — you are buying the feeling of keeping up.
Set a total before you set the line items
Decide the ceiling first, then fill it. Working the other way round — adding up tools people want and calling the total a budget — is how the number doubles every year.
Two sanity checks that hold up well for SMBs:
- Per head. €30–60 per employee per month in software for a business doing knowledge work. Below €30 you are probably on free tiers with real data risk; above €60 you should be able to point at specific revenue or specific hours saved.
- As a share of revenue. 0.5–1.5% of revenue for a service business. A €900,000 agency landing at €9,000–13,500 a year is in normal territory. A €300,000 consultancy spending €12,000 is not.
The 60/30/10 split
Once you have a total, divide it three ways. The ratio matters more than the absolute figures, and it keeps the two mistakes — spending everything on subscriptions, or spending everything on experiments — off the table.
60% — Core tools
The two or three tools your team genuinely uses most days: usually one general assistant on a team plan, plus whatever is embedded in the software you already run. This is boring, predictable, per-seat spend, and it should be the majority of the budget. Buy the team or business tier rather than individual plans — the admin controls, shared prompt libraries and data-handling terms are worth the uplift.
30% — Workflow and integration
This is the part most SMBs underfund, and it is where the actual return lives. It covers an automation platform to connect tools to each other, any paid connectors, and one properly resourced pilot per quarter. A chat assistant nobody has wired into your CRM or inbox saves minutes; a workflow that drafts and files something automatically saves hours. If this bucket is empty, you are buying capability and never installing it.
10% — Experiments and training
A standing allowance for trying things, with two rules: a per-experiment cap (€100–200 is plenty for most SMBs) and a 60-day verdict — keep it, cut it, or decide properly. Training belongs here too. An hour of structured practice per person per quarter does more for output than another subscription, and it is the cheapest line in the whole budget.
Budgeting for usage-based pricing
Per-seat pricing is easy to forecast. Usage-based pricing — API calls, credits, tokens, minutes of transcription — is where budgets break, because the cost scales with enthusiasm rather than headcount.
Three habits keep it survivable:
- Set a hard cap in the vendor console on day one. Almost every usage-priced platform lets you set a monthly spend limit and an email alert at a threshold. Set the limit at 1.5× your estimate and the alert at 70%. Do it before the first real month, not after the first surprise.
- Forecast from one measured month, then add 30%. Run normally for 30 days, take the invoice, add a third for growth and the occasional bulk job. Guessing from a per-token price sheet is a waste of an afternoon.
- Budget automations separately from people. A scheduled job that processes every inbound enquiry has a cost curve tied to your sales volume, not your staff count. Give it its own line so a good month does not look like an overspend.
One prompt for the monthly review:
"Here are my last three months of AI tool invoices, pasted below. For each tool, tell me the monthly trend, flag anything that grew more than 25% month over month, and list any two tools whose descriptions suggest overlapping functionality."
The costs that never make the spreadsheet
Subscription fees are the visible part. The budget that survives contact with reality includes the rest:
- Learning time. Four hours per person to get genuinely useful with a new tool. At a loaded cost of €35 an hour that is €140 a head, one-off — real money, and real money well spent. Skipping it is how you end up paying for seats nobody uses.
- Checking time. Someone reviews the output. On client-facing work that review is not optional, and for the first few months it can eat a meaningful share of the time the tool saved. It shrinks as people learn what to trust, but it is never zero.
- Setup and integration. Connecting AI to your existing systems takes a few days of somebody's attention, internal or paid. Budget it as a one-off project cost rather than pretending it is free.
- Review and compliance. Reading a data-processing agreement, or having someone read it for you, is an hour you should plan for rather than discover.
A reasonable rule: for every €1,000 of annual subscription spend, allow a further €400–600 of one-off implementation cost in year one. If that makes the business case look thin, the business case was always thin — our walkthrough of how to calculate the ROI of an AI implementation shows how to test it honestly before you commit.
Build a renewal calendar
The single cheapest control you can add is a one-page list of every AI tool with its renewal date, billing term, notice period and named owner — plus a calendar reminder 30 days before each renewal with one question attached: keep, cut, or consolidate?
On billing terms, be deliberate. Annual commitments typically save 15–20%, and that is real. But this market reprices and re-features itself every few months, and a 12-month lock-in on a tool you might replace in April is a false economy. A workable default: monthly for anything under €100 a month or less than six months old in your stack, annual only for the one or two tools you would genuinely fight to keep.
Assign one person to own the whole list. Not a committee — a name. Spend without an owner is spend that only ever goes up.
A worked example: a 12-person agency
A 12-person marketing agency, roughly €1.1m revenue, budgeting for the year ahead. Total ceiling set at €1,000 a month — about €12,000 a year, near 1.1% of revenue and €83 per head.
- Core (€600/month). A team plan on one general assistant for all 12 staff at roughly €30 a seat, plus the AI features already bundled in their project management and design tools at no extra cost.
- Workflow (€300/month). An automation platform at €80, usage-based API spend of around €120 for an automated client-reporting job, and €100 towards the current quarter's pilot — this quarter, automated first-draft campaign briefs.
- Experiments and training (€100/month). One or two tools on trial at any time with a €150 cap each, and a quarterly team practice session.
Plus a one-off year-one implementation allowance of about €5,000: two days of integration work, 48 hours of staff learning time across the team, and a legal review of two data-processing agreements.
The bottom line
An AI budget is not a tool list with prices next to it. It is a ceiling you chose, split deliberately between what you run every day, what connects it together, and what you are still learning — with the implementation costs written down instead of absorbed. Do the 30-minute audit, set a per-head or percentage-of-revenue ceiling, divide it 60/30/10, cap anything usage-based on day one, and put one name against every renewal date.
Owners who get this right next year will not be the ones who spent the most. They will be the ones who can say, in one sentence, what each line is for.
Where does your business stand on AI?
Take the free 3-minute AI Readiness Quiz and get a personalised score with your next steps.
Take the Free Quiz →