Budgeting for AI in a Small Business: Counting Cost per Task, Not per Subscription

A marketing agency of nine people discovers, when the owner finally adds up the card statement, that it pays for eleven AI subscriptions: three writing assistants, two image tools, a meeting transcriber on six seats, a chatbot builder nobody finished setting up, and usage charges from an automation that ran in a loop for a weekend. The monthly total equals a junior salary. Nobody can say what it produced. Each purchase was small and sensible. The AI budget was never a budget, only an accumulation.

How AI is priced

Three models, often mixed.

• Per seat: a fixed monthly fee per user. Predictable. Wasteful when seats go unused.
• Per credit or action: a bundle of actions, with more to buy when it runs out. Predictable until the bundle is exceeded.
• Per usage: charged by the amount of text, images, minutes or calls processed. Scales with use, and can spike.

Many products combine a subscription with an allowance and charges beyond it. Read what happens at the limit: does the tool stop, slow down or keep charging?

Count by task

The useful unit is a completed task, not a subscription.

For each task you give to AI:

• Volume: how many per month?
• AI cost per task: the tool's charge divided by tasks done, or the usage charge per task.
• Review time per task: minutes a person spends checking and correcting, multiplied by what that time costs.
• Total cost per acceptable result.

Then compare with:

• Cost of doing it by hand: minutes per task multiplied by the cost of that time.

Example. A team answers 1,200 customer enquiries a month. By hand, a reply takes four minutes. With AI drafting, a person reviews and sends in one minute. The saving is three minutes per reply, 3,600 minutes or sixty hours a month. If the tool costs less than sixty hours of that staff member's time, and quality holds, it pays. If review takes three and a half minutes because drafts are poor, it does not.

The review time is the figure most people leave out, and it decides the result.

Where the hidden costs are

• Review and correction.
• Setup and maintenance: writing instructions, keeping price lists and policies current.
• Mistakes: a wrong price sent to a customer has a cost.
• Unused seats.
• Overlapping tools doing the same job.
• Retries and loops in automations.
• Long inputs. Sending a whole document when a page would do multiplies usage.
• Training time for staff.

Set a monthly cap

• One owner for all AI spending.
• One payment method, so the total is visible.
• A cap per tool and an overall cap.
• Usage limits and alerts switched on in every tool that offers them.
• A rule for new tools: a named task, an expected saving, a trial period, and something else cancelled if it overlaps.

A cap is not about spending little. It makes sure spending is a decision.

Choose the level of capability per task

Not every task needs the strongest and most expensive option.

• Simple, high-volume tasks: sorting messages, extracting an order number, short standard replies. A lighter, cheaper option usually suffices.
• Difficult or high-stakes tasks: a nuanced customer reply, a long document summary, anything customer-facing with money involved. Quality matters more than unit cost.

Test both on twenty real examples and compare the cost per acceptable result. The cheapest per unit is often not the cheapest per result.

Reduce consumption without losing quality

• Send only what is needed. The relevant part of a document, not the whole archive.
• Shorter instructions that are still clear.
• Reuse: save good answers to frequent questions as templates, so AI is not asked the same thing a hundred times.
• Stop loops. Every automation has a maximum number of attempts and a person to alert.
• Draft where it helps, skip where it does not. A two-word confirmation needs no AI.
• Batch non-urgent work where the tool charges less for it.

The quarterly review

One hour.

1. List every AI tool and its cost over three months.
2. For each, name the task it performs and the volume.
3. Estimate the saving against doing it by hand, including review time.
4. Mark each: keep, reduce seats, replace, cancel.
5. Look for overlap. Two tools for the same task means one goes.
6. Check usage spikes and their cause.
7. Ask the team which tools they would miss.

Tools nobody would miss are the easiest saving in the business.

Value beyond time saved

Some benefits are real and harder to count.

• Faster replies, which win more enquiries.
• Coverage outside working hours.
• Consistency of tone and information.
• Work that would not be done at all, such as following up every old quotation.

Count them where you can: enquiries answered within five minutes, follow-ups sent, orders from those follow-ups. Do not use vague benefits to justify a tool that fails on measured ones.

Before you sign

• What exactly is counted as a unit of usage?
• What happens at the limit?
• Can limits and alerts be set?
• Is there a usage report by user and by task?
• How easy is cancellation, and can the data be exported?
• Does the price depend on a model or supplier that may change?
• Where does the data go, and is it used for training?

Common mistakes

• Subscribing first, finding a task later.
• Ignoring review time.
• Seats for everyone when three people use it.
• Several tools for one job.
• No usage limits.
• Automations without a maximum number of attempts.
• Choosing by price per unit and not per acceptable result.
• Never cancelling.
• No single owner of the spend.

Fewer tools, one workspace

Part of AI cost is duplication across separate products. Polynusa includes AI drafting and agents inside the same workspace as the inbox, customer records and bookkeeping, with usage counted in actions, so the spend sits next to the work it supports. See the pricing page for current plans and allowances.

Summary

Budget for AI by task, not by subscription. For each task, add the tool's cost and the review time, and compare the cost per acceptable result with doing it by hand. Put one person in charge, set caps and alerts, match the level of capability to the difficulty of the task, cut waste such as loops, long inputs and unused seats, and review every quarter to keep, reduce or cancel.

FAQ

How much should a small business spend on AI?

There is no universal figure. Start from tasks: choose one or two that are frequent and time-consuming, estimate the hours they take now and what that time is worth, then compare with the cost of having AI handle them, including review time. Spend where the saving is clear, cap the monthly total, and expand only after a measured result.

Why does the cost of AI vary so much from month to month?

Because many tools charge by usage: per message, per action, per document or per unit of text processed. A busy month, longer conversations, larger documents or a workflow that retries on errors all increase consumption. Usage limits, alerts and a look at which tasks consume the most keep it predictable.

Is a cheaper AI model always the better choice?

Not necessarily. A cheaper model that produces drafts needing heavy correction costs more in staff time than it saves. Compare the total cost per acceptable result, including the minutes a person spends reviewing and fixing. For simple, high-volume tasks a lighter model is often enough, and for difficult ones a stronger model can be cheaper overall.