What AI really costs a small business

Why is the subscription only part of the bill?

We start an AI budget with a job and a baseline. How is the work completed today, who touches it and what would count as a good result? Without that, a cheap subscription can become an additional cost with no task removed.

Separate cash spent from staff time. Software payments leave the bank account. Review time uses capacity, even if salaries stay the same. Time freed for useful work is a capacity benefit; it becomes a cash saving only when an actual expense falls. Do not present those as the same result.

Also separate a pilot from a recurring service. Setup, migration and training happen before the regular month starts. A first month can cost more than later months, while a quiet trial can understate the volume and exceptions of normal operation.

What belongs in the monthly budget?

A line item can be zero if it is genuinely included elsewhere. Mark it as included, so the next reviewer knows it was considered. Keep existing baseline software separate from new incremental spend; otherwise the project gets charged twice for an office suite the team already needs.

How should we budget for seats?

As of October 2026, OpenAI lists ChatGPT Business at $25 per user per month on monthly billing, or $20 with annual billing, with at least two users. Claude Team standard lists the same monthly and annual reference rates. These are published US dollar prices, checked 12 October, before your final local checkout. OpenAI plans and Claude plans.

Do not multiply the lowest advertised annual equivalent by the team size and call that next month's invoice. Check the billing commitment, cash timing, taxes, currency and minimum seats. Keep a renewal date and an owner beside every subscription.

We would give paid access first to the people with a proven repeated use. The benefit is a smaller, observable trial. The tradeoff is that the pilot may not expose wider team permissions or handoffs. Test those deliberately before buying company-wide access.

How do API and automation charges work?

An API is a way for a workflow to call a service automatically. OpenAI's API pricing separates model input and output, with additional pricing for some tools. A chat subscription and an API budget are different purchases; do not assume one funds the other. OpenAI API pricing and subscription pricing.

For a rough model-only estimate, multiply input usage by its rate and output usage by its rate, then add applicable tool, storage and processing charges. Use the provider's recorded billing units. Retries and long source material can change the total. An allowance is a spending plan, not a promise that the service will stop automatically.

Automation has a second meter. Zapier describes successful actions as tasks, with exceptions for some built-in tools. Make uses credits; most actions consume one, while some advanced AI features can consume more. One incoming enquiry can involve several chargeable steps. Zapier task billing and Make credit billing.

Ask the implementer to show a normal run and a failed run in the billing records. Look for repeated polling, duplicate processing, retries and unbounded loops. Set alerts and a tested pause mechanism wherever the service allows them.

Worked example: a small enquiry workflow

Illustrative budget only. This is a fictional planning example, not a DLVX client result, quote or estimate of typical business usage. All quantities, labour rates and non-seat allowances below are assumptions. Replace them with your supplier quote and measured pilot data. Amounts are in USD; tax and currency conversion are excluded.

The seat reference uses the monthly ChatGPT Business rate above. Published seat price. The example assumes the team already has suitable email and business software, so their existing cost is excluded. It also assumes the custom flow's API calls are funded by the separate API allowance.

Suppose the same work currently takes 16 hours each month, valued at the illustrative $25 hourly rate. That baseline is $400. The example leaves $60 of monthly economic benefit after $340 of recurring cost. If setup costs an illustrative $600, simple recovery takes 10 months while those assumptions hold.

Now change review time from 4 to 6 hours. Recurring economic cost becomes $390, leaving only $10. That sensitivity matters more than a small seat discount. If the output fails a hard quality condition, the workflow should stop regardless of the attractive calculation.

What should we measure during the pilot?

Measure accepted work, not generated messages. Keep the input, whether the result passed, the review time, the correction time and the final destination. Include failed and missing outputs in the total. Otherwise the team can rescue a weak system while the dashboard reports success.

Track the reason for each exception. A missing customer detail may need a better intake form. An outdated answer may need a maintained policy source. A formatting problem may need a template. Those fixes can be cheaper than moving every task to a more expensive model.

Use a cost per accepted result when comparing alternatives: all relevant recurring cash and labour costs divided by accepted results. Keep serious failures visible separately. A wrong payment instruction is not merely another inexpensive item that needed editing.

The first-workflow calculator helps estimate human time, including review, exceptions and upkeep. Keep the financial budget beside that calculation so the two tell the same story.

Which cost controls are worth setting?

• Give every subscription and workflow a named budget owner.
• Review paid seats against actual useful work before renewal.
• Record the included allowance, overage behaviour and pause process.
• Limit retries and retain enough evidence to diagnose a failure.
• Use the smallest relevant source pack for each task.
• Recheck the budget after a model, supplier or workflow change.
• Keep a manual fallback for essential work.

Spend controls should protect useful work, not merely make a graph look smaller. A cheaper run that demands a long correction is poor value. Equally, an expensive model should earn its place through observed improvement rather than its position in the product catalogue.

Common mistakes

Counting saved time as immediate cash profit. Forgetting annual commitments. Paying for overlapping seats without a clear job for each. Leaving the business owner to maintain sources for free in the spreadsheet. Treating a budget alert as a guaranteed hard cap. Estimating a production month from only easy trial cases.

Another mistake is hiding setup inside a monthly retainer without explaining what happens at the end. Ask what you own, what remains usable if support stops and which recurring costs are paid directly to suppliers.

Your next step

List your current subscriptions, choose one workflow and build this budget with your real quantities. Record cash spend and staff time separately. Set a review date and a clear keep, revise or stop decision. If the ongoing ownership looks disproportionate to the work, compare an existing product before commissioning a custom system.

Sources and review date

Reviewed 12 October 2026. Prices and plans are as of October 2026. DLVX recommendations are editorial judgements, not a vendor ranking or a promise of results. Recheck the linked terms before buying or connecting business data.

• OpenAI plan pricingChecked 2026-10-12. Monthly and annual ChatGPT Business prices and minimum users API and subscription usage distinction
• Claude pricingChecked 2026-10-12. Team standard seat prices
• OpenAI API pricingChecked 2026-10-12. Input output and tool pricing structure
• Zapier pricingChecked 2026-10-12. Task billing and built-in tool exceptions
• Make pricingChecked 2026-10-12. Credit billing and advanced AI exceptions