Plan on $20 to $30 per person a month for staff who'll use an AI assistant daily, up to about $100 a month for automation tools, plus one-off setup help and paid learning time. Then cap the total: in year one, spend no more than about a third of the value of the hours AI will realistically save.
For a small team that usually lands somewhere in the low thousands of dollars for the first year, and for a sole trader a few hundred. Paid learning time is the line that catches people out: a seven-person firm can easily spend 30 hours of paid time getting people comfortable, which can cost more than a year of the seats themselves. Prices are list USD as of September 2026; check each vendor's page before you commit, because promotions and plan names change often.
The six lines of an AI budget
Most AI budgets that go wrong only had one line in them: subscriptions. The table below is the full set. The first four are cash; the fifth may or may not be; the sixth is almost always your own staff's time.
| Budget line | What it covers | Typical list price | One-off or ongoing |
|---|---|---|---|
| 1. Assistant seats | ChatGPT, Claude, Copilot or Gemini for the people who use them daily | $0 if included in your office suite; ChatGPT Business or Claude Team Standard $20 per user a month annually, $25 monthly; Microsoft 365 Copilot Business $21 annually ($25.20 monthly) | Monthly |
| 2. AI inside software you already use | Add-ons, credits or price rises in your CRM, accounting or quoting tools | Varies; HubSpot's Breeze AI runs on credits at $10 per 1,000, with 500 a month included on Starter | Monthly |
| 3. Automation platform | Connecting apps so tasks run without copying and pasting | Free tiers; Make from about $9 a month; Zapier Professional from $19.99 a month annually ($29.99 monthly) for 750 tasks | Monthly |
| 4. Usage-based AI | Automations that call an AI model directly, billed per token | Often a few dollars to tens of dollars a month at small-business volumes (worked below) | Monthly, variable |
| 5. Setup help | Your own hours, or a consultant or freelancer to build the first workflow | Quote-dependent; ask for a fixed price for a defined workflow | One-off |
| 6. Learning and change time | Training, practice, the slower first weeks, and upkeep | Staff hours; see how much time and money AI staff training takes | Mostly one-off, then an hour or two a month |
Add a contingency of about 10% on the cash lines. Something always turns up: an add-on you try and cancel, a promotion that ends, a month where an automation runs more often than you expected.
Line 4 puts people off because token pricing looks abstract, so here's what it means in practice. Say an automation reads 2,000 supplier and customer emails a month, sending about 1,500 tokens into the model and getting 300 back for each. That's 3 million input tokens and 0.6 million output tokens. On Claude Sonnet 5, at $2 in and $10 out per million, the month costs about $12, and OpenAI's gpt-6-sol has the same list price. Sorting and extracting rarely need a model that size: Claude Haiku 4.5 at $1 and $5 brings the month to around $6, and OpenAI's gpt-6-luna, at $0.10 in and $0.50 out, to about 60 cents. A ChatGPT or Claude subscription pays for none of this; API calls are metered on their own bill.
Line 2 hides usage pricing too. HubSpot's Customer Agent, which answers support chats on its own, needs any Professional or Enterprise hub (usually Service Hub), and each conversation it resolves uses 50 credits. Professional includes about 3,000 credits a month, which covers roughly 60 resolved conversations if nothing else draws on the pool. Say an online shop's agent resolves 200 a month: that's 10,000 credits, 7,000 more than the allowance, and at $10 per 1,000 credits the overage comes to about $70 a month on top of the subscription. Unused credits expire at month end rather than carrying forward. Budget the busy month.
Set the ceiling from the hours, not from the price list
A price list tells you what things cost. It doesn't tell you what you can afford. For that, start with the tasks AI will take on and the hours it will realistically save, and work backwards.
1. Realistic hours saved per week (named tasks only, after checking time)
2. x 46 working weeks
3. x loaded hourly cost (wage plus employer on-costs)
= annual value of the time
4. Year-one budget ceiling = annual value / 3
Why a third? It's my rule of thumb, and there are three reasons for the margin. Early estimates of time saved usually run high. The first month or two with any new tool is slower, not faster. And saved time only becomes money if you use it for paid work or avoid a cost, which doesn't happen automatically. A ceiling at a third of the value leaves room for all three and still pays back within the year if the savings are half what you hoped. If you want to test the numbers more formally, calculating AI ROI with a worked example takes it further.
The formula works at any size, including one person. A self-employed garden designer expects AI to save about two hours a week on client proposals and follow-up emails. At a loaded $40 an hour, that's 2 × 46 × $40 = $3,680 a year, so the ceiling is about $1,230. An individual plan such as ChatGPT Plus or Claude Pro at $20 a month costs $240 a year, well inside it. The two-seat minimum on ChatGPT Business and Claude Team would double that for no benefit, so a sole trader is usually better on an individual plan with the model-training switch turned off in its privacy settings.
Two cautions. The ceiling is a maximum, not a target; if you can get the savings for less, spend less. And if you can't name the tasks in step 1, you're not ready to set a budget, only a small pilot allowance.
That second caution comes from a mistake worth describing. A 12-person recruitment firm's owner reckoned AI would "save everyone an hour a day": 12 people × 5 hours × 46 weeks × $30 is about $82,800, a ceiling near $27,600, and that felt like permission to buy 12 annual Microsoft 365 Copilot Business seats. At the 90-day review, the Copilot usage report in the Microsoft 365 admin centre showed four people using it regularly. The other eight had tried it once or twice, because no one had named a task for them. Those eight seats cost 8 × $21 × 12, about $2,000 for the year, and annual seats generally can't be reduced until renewal. Built from named tasks instead (candidate summaries and job adverts for the four consultants who write them), the estimate would have been about 10 hours a week and four seats.
Three budgets at different sizes
These are illustrative first-year budgets built from the list prices above. Setup help is left out of the totals because it depends entirely on whether you build things yourself or pay someone.
| Line (year one) | Owner plus part-time admin | 7 people, 3 daily AI users | 20 people, 12 daily AI users |
|---|---|---|---|
| Assistant seats | $240 (one $20 a month plan) | $720 to $900 (3 seats at $20 to $25) | $4,800 to $6,000 (10 standard seats plus 2 premium seats at $100 to $125) |
| AI in existing software | $0 | $0 to $300 | $500 to $1,500 allowance for CRM credits and add-ons |
| Automation platform | $0 to $108 (free tier or Make) | $240 (Zapier Professional, annual) | $828 (Zapier Team at $69 a month, annual) |
| Usage-based AI | $0 | $0 to $240 | $300 to $1,200 |
| Contingency (10%) | About $25 to $35 | About $95 to $170 | About $640 to $950 |
| Year-one cash, before setup help | About $265 to $385 | About $1,050 to $1,850 | About $7,100 to $10,500 |
| Staff learning time | About 10 hours of the owner's time | About 35 to 40 hours in total | About 120 to 150 hours in total |
If you're on Microsoft 365 or Google Workspace, check line 1 carefully before paying for a separate assistant. Copilot Chat comes with every Microsoft 365 business plan, and Gemini is now part of Google Workspace's business tiers. On Microsoft, the Business Standard plus Copilot bundle at $23.50 per user a month (annual) costs less than Business Standard at $14 plus the $21 Copilot add-on bought separately. Not everyone needs a paid seat, either; who in your team actually needs a paid AI licence helps you trim the count.
Worked example: a seven-person kitchen fitter's first year
Take an illustrative kitchen fitting firm: the owner, an office manager, a designer and four fitters. The owner expects AI to save about seven hours a week across quote writing, customer emails and supplier orders once it's running. At a loaded cost of $32 an hour, that's 7 × 46 × $32, about $10,300 a year, so the year-one ceiling is roughly $3,400.
| Months | What they paid for | Cash |
|---|---|---|
| 1 to 2 | Pilot: 2 ChatGPT Business seats, monthly, at $25 (the plan's minimum is two seats) | $100 |
| 3 to 4 | Designer added: 3 seats monthly at $25; Zapier Professional monthly at $29.99; a quoting add-on trial at $79 a month, cancelled after two months | $368 |
| 5 to 12 | After the 90-day review: 3 seats moved to annual billing at $20; Zapier moved to annual at $19.99 | $640 |
| 6 to 12 | A small automation that sorts supplier emails, calling a model through the API | About $70 |
| Year-one cash | About $1,180 | |
On top of the cash, the three office staff spent about 8 hours each learning and practising, and the fitters about 90 minutes each on the phone app for job photos and notes: roughly 30 hours, or around $940 at their hourly costs. The table spreads the annual plans across the months they cover; in practice annual plans are usually billed upfront, so that cash leaves in month five. The owner built the Zapier automation over about 12 hours of evenings.
All in, cash plus staff time came to roughly $2,100 before the owner's own evenings, comfortably under the $3,400 ceiling. By month six they were saving around six hours a week, a little under the estimate, which is exactly why the ceiling has margin. Two things they'd do differently: move to annual billing after the review, as they did, but not sign the annual Zapier plan until the automation had run cleanly for a full month, because they changed its design twice.
Budget lines the pricing pages leave out
- The monthly premium. Monthly billing costs 20 to 25% more than annual on ChatGPT Business, Claude Team and Microsoft 365 Copilot Business. Worth paying while you test; wasteful once you're sure.
- Promotions that end. Microsoft 365 Copilot Business has a promotional $18 per user a month on annual billing from 1 July to 31 December 2026, after which it's $21. Budget the full price.
- Price rises at renewal. Microsoft 365 Business Basic went from $6 to $7 and Business Standard from $12.50 to $14, applying from 1 July 2026 at each customer's next renewal.
- Minimum seats. ChatGPT Business and Claude Team both need at least two seats, so a solo buyer pays for two or uses an individual plan.
- Usage that counts differently from how you'd guess. Zapier bills each successful action step as a task, while triggers and filters don't count; a Zap with five action steps running 200 times a month uses 1,000 tasks. HubSpot's included credits don't roll over.
- Upkeep. Automations break when a form, inbox or app changes. Budget an hour or two a month of someone's time per live automation.
There's a longer list in hidden costs of AI implementation, including the ones that only show up in year two.
Spend it in three steps, not all at once
Releasing the whole budget on day one is how businesses end up with annual seats for a tool nobody uses. Split it:
- Pilot (months 1 to 3): about 10 to 15% of the ceiling. A couple of monthly seats, free or entry tiers of any automation tool, one task. Measure hours saved.
- Expand (months 4 to 6): up to half the ceiling. Add seats for the people whose tasks proved out, move them to annual billing, build the second automation.
- Review (month 6 and month 12). Compare actual hours saved with the estimate in the ceiling formula. Cut what isn't used before renewal dates, not after.
For anything billed by usage, set a hard monthly limit and an alert at around 75% of it from day one. Setting spending limits and alerts on pay-as-you-go AI tools shows where those settings live.
If the number comes out too high
Start with what you already own before you cut ambitions. Use the AI included in your office suite for staff who only need occasional drafting. Buy paid seats only for daily users. Keep automation on free or entry tiers until volume forces an upgrade. Point simple sorting and extraction jobs at a cheaper model; on Anthropic's API, the Batch option halves input and output prices for work that doesn't need an instant answer. And if the ceiling formula itself gives a small number, that's useful information too: it may mean AI suits one or two of your tasks well rather than the whole business, which is a perfectly good place to start.
Here's that trimming applied to the 20-person budget in the table above, which topped out at about $10,500. Twelve paid seats become six, for the people whose named tasks save the most, at $20 each on annual billing: $1,440 instead of up to $6,000. The other six use the Copilot Chat or Gemini already included in their office plan. The two premium seats wait until someone actually hits the standard limits. Automation drops from Zapier Team to Professional on annual billing ($240) while volume stays under 750 tasks a month. Usage-based AI stays near $300 by sending sorting jobs to a small model. With the software allowance and contingency, year-one cash comes to roughly $2,700 to $3,800, and seats can be added at the six-month review for anyone the included tools don't serve. For a wider view of what everything costs this year, see how much AI costs a small business in 2026.
Budgeting questions that come up next
Should my AI budget be a percentage of revenue?
I wouldn't use one. There's no reliable benchmark for small businesses, and a percentage of revenue ignores whether AI suits your work at all. A decorating firm and a bookkeeping practice with the same turnover could sensibly spend very different amounts. Base the figure on the hours AI can realistically save in named tasks, cap it at a fraction of their value, and adjust after a 90-day review.
Do ChatGPT or Claude subscriptions cover automations built on their models?
No. ChatGPT Plus and Pro and Claude Pro and Max cover use of the chat apps, not API calls. Automations that call a model directly, for example through Zapier, Make or custom code, are charged by the token through a separate developer account. Budget that as its own line, set a spending limit on the account, and check actual usage after the first month.
When should I switch from monthly to annual billing?
After a seat has been used for real work for 60 to 90 days and you're confident you'll keep it for a year. Annual billing saves roughly 15 to 20% on the main business assistant plans, but on ChatGPT Business, Google Workspace annual plans and Microsoft annual subscriptions you generally can't reduce seats until renewal, so an unused annual seat costs more than the saving.
Further reads
- AI Subscription Prices Compared for Small Businesses (2026) — Current prices for every plan named in your budget.
- What Is the Cheapest Way to Start Using AI in a Small Business? — The lowest-cost starting point if the budget is tight.
- How Soon Should AI Pay for Itself? Payback Periods by Project — Payback periods to set against your first-year spend.
- How to Write a One-Page AI Business Case for a Small Business — Put the budget and the expected return on one page.
- AI Maintenance Costs: What You Pay After an Automation Goes Live — What automations cost to keep running in year two.
- How to Find AI Savings Line by Line in Your Profit and Loss — Find the cost lines AI can realistically reduce.
- How to Write a One-Page AI Strategy for Your Business — The seven boxes a one-page AI strategy needs, a filled-in garden centre example, and five tests that show whether your page will guide real decisions.
- Pros and Cons of Using AI in Your Business, With Real Costs — Every pro and con of business AI with a price attached, a first-year ledger for a small law firm, and the costs that never reach an invoice.
- Barriers to AI Adoption in Small Businesses and How to Clear Them — Eight practical barriers that keep small firms from adopting AI, the cheapest way past each, and the few barriers you should respect rather than clear.
- How Much Does AI Cost a Small Restaurant Each Month? — A line-by-line monthly AI budget for a small restaurant, three priced set-ups, and the staff hours that cost as much as the subscriptions.
- How Much Should You Spend on AI-Run Ads Before Judging Results? — A budget formula for testing Meta and Google AI campaigns, three spend scenarios, a butcher's four-week test and stop rules to write first.
- Is Open-Source AI Really Free? The True Cost for a Small Business — What running an open-source AI model really costs a small business, itemised: hardware, rented GPUs, hosted APIs, setup time and the break-even point.
- Do You Need a Copilot+ PC or AI Laptop for Your Business? — What a Copilot+ PC adds, which AI works on any computer, and when it's worth choosing one as you replace old Windows 10 machines.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI ChatGPT Business, Anthropic Claude and Microsoft 365 Copilot pricing pages; Zapier and Make pricing pages; Anthropic and OpenAI API pricing pages; HubSpot credits pricing. Prices are list USD, checked 27 September 2026.