The situation
Note: The business described below is an illustrative example created to show a common pattern. It is not a real client or a real company.
Picture a four-person bookkeeping firm based in Brisbane. The owner runs the business day-to-day, two staff handle client accounts, and one part-timer manages admin. Solid, busy, and always looking for ways to work smarter.
After sitting through a couple of vendor demos in early 2024, the owner signed up for two AI tools: one for drafting client emails and reports, another for summarising documents and pulling out figures. The expectation was straightforward — a few dollars a month, a bit of time saved, done.
Twelve months later, the combined cost of running those tools was sitting well above what the owner had budgeted. More troubling, it was not obvious where the money was going or whether the tools were actually saving more time than they were consuming.
This situation is not unusual. The cost of AI for small business is rarely just the subscription price shown on the sign-up page. There are at least three layers underneath it, and most small business owners only discover them after the bill arrives.
What was going wrong
When the owner sat down to work out the real numbers, three cost layers emerged that had never been counted.
1. The wrong subscription tier
Both tools offered a free tier, a mid-tier plan, and a higher business plan. The owner had signed up for the mid-tier on both, assuming the free tier would be too limited. That assumption was only half right.
For one of the tools, the mid-tier unlocked features — advanced integrations, priority processing, extended history — that the team never once used in a typical month. The free tier lacked one specific feature they did rely on, but a lower paid tier would have covered it at roughly half the mid-tier price.
(Illustrative figures: mid-tier plan at AU$65/month versus a lower plan at AU$32/month — a difference of AU$33/month, or AU$396 over a year, for features the team was not using.)
2. Token and usage overages
A token is simply a small chunk of text — roughly three-quarters of a word — that an AI model reads or writes each time you use it. Every prompt you send and every response you receive is counted in tokens, and most plans include a monthly allowance before extra charges kick in.
The problem for this firm was document summarisation. A single 10-page client report fed into the AI tool can consume several thousand tokens in one go. Across a team of four, each running two or three summaries a day, the monthly token count climbed far beyond the plan allowance without anyone noticing — because no one was watching the usage dashboard.
(Illustrative figures: the plan included 500,000 tokens per month. Actual usage was tracking at roughly 820,000 tokens, triggering overage charges of approximately AU$28/month.)
3. The hidden staff-time cost
This was the biggest cost layer, and the least visible. The owner had a habit of reviewing every AI-generated email and report before it went to a client. That is sensible practice. But no one had ever timed it.
When the owner finally tracked it, the review-and-correction process — reading the output, fixing the tone, adjusting figures, re-prompting when the first result was off — was taking roughly 40 minutes a day. At a conservative hourly rate for a working owner, that adds up quickly.
(Illustrative figures: 40 minutes/day × 5 days × 4 weeks = approximately 13 hours/month. At AU$80/hour, that is AU$1,040/month in owner time — never once counted against the tool's supposed savings.)
What they changed
The owner did not cancel anything straight away. Instead, three specific changes were made over about two weeks.
Step 1: Audit actual feature usage and downgrade
The owner logged into the account settings of each tool and pulled up the usage history for the past 30 days. Most platforms show this under a section labelled Usage, Activity, or Billing. The goal was simple: list every feature used at least once in that period.
For one tool, the list was short. The team had used the core text generation feature and nothing else. The owner compared that list against what each pricing tier actually included, then downgraded to the plan that covered those features and nothing more.
Step 2: Set a token budget and build a shared prompt library
Using the provider's usage dashboard — found under Settings > Usage on most platforms — the owner set a soft monthly token limit and turned on usage alerts at 70% and 90% of the allowance.
The second part of this step tackled waste at the source. Each staff member had been writing their own prompts from scratch every time, often producing longer, less efficient instructions than necessary. The owner created a shared document — a simple table in Google Docs — listing the ten most common tasks and the shortest prompt that reliably produced a usable result for each one. Staff stopped re-writing prompts, and token use per task dropped noticeably.
Step 3: Time the workflow and narrow where AI is used
For one full week, the owner used a basic timer — a phone stopwatch works fine — to record the end-to-end time for each AI-assisted task: prompt written, output received, output reviewed, corrections made, final version approved. That total time was then compared against a rough estimate of how long the same task took before the AI tool existed.
Two task types came out clearly ahead: drafting routine client update emails and formatting standard monthly summaries. These needed minimal correction and saved meaningful time. Two other task types — interpreting ambiguous figures and drafting advice-style content — required so much correction that the AI was adding time, not saving it. The owner restricted AI use to the two tasks where it genuinely helped and stopped using it for the other two.
The result
All figures below are illustrative. They are based on a plausible scenario and are intended to show the shape of the outcome, not to represent a specific real business.
Here is the before-and-after comparison across the three cost layers:
- Combined subscriptions: AU$130/month before; AU$80/month after downgrading one plan. Saving: AU$50/month.
- Token overages: approximately AU$28/month before; AU$0 after setting a budget and using the shared prompt library. Saving: AU$28/month.
- Owner review time: approximately 13 hours/month before; approximately 5 hours/month after narrowing AI use to two task types. Saving: roughly 8 hours/month, or AU$640/month at AU$80/hour.
The combined illustrative saving is roughly AU$718/month. The largest share of that — nearly 90% — came not from switching tools or finding a cheaper provider, but simply from narrowing where the tools were used.
It is worth being honest about what did not improve. The token overage alerts helped, but the team still occasionally exceeded the budget in busy months when a large batch of documents came in. That is a workflow problem, not a technology problem, and it has not been fully solved. The shared prompt library also required a short training session before staff used it consistently — that took time upfront.
How to apply this to your business
You can run a version of this audit yourself in a few hours spread across one week. There are three parts.
Part 1: Check your subscriptions against actual usage
Write down every AI tool your business currently pays for. For each one, log into the account and find the usage or activity report for the past 30 days. List every feature you actually used. Then compare that list to what your current plan tier includes. If you are paying for features you have not touched, check whether a lower tier covers what you do use.
Part 2: Pull your usage or token report
Most AI tools include a usage or billing dashboard. Find it — usually under Settings or Account — and look at which tasks or workflows generated the most token use or the highest activity. If you are seeing overages, trace them back to specific task types. Long document inputs and open-ended generation tasks are the most common culprits.
Part 3: Time one AI-assisted workflow end-to-end
Pick one task you regularly use AI for. Time the whole process: writing the prompt, waiting for the output, reading it, correcting it, and finishing the task. Write that number down. Then estimate how long the same task took before you had the AI tool. If the AI version is not clearly faster once correction time is included, that task may not be a good fit for AI right now.
The decision rule is straightforward: if the corrected time cost plus the subscription cost exceeds the time saved, the tool is not yet earning its place. That does not necessarily mean cancelling it — it may mean using it differently. But the number needs to be honest.
If you would like help running this audit on your own setup, the team at Rapid Ready AI is happy to have a straightforward conversation about what you are currently running and whether it is working as hard as it should be. No pressure — just a practical look at the numbers.
