Most AI ROI claims are speed claims in disguise. “This drafts a proposal in two minutes instead of an hour” may be true, and it tells you almost nothing about whether the workflow is worth running. Somebody still reads the draft, fixes the one where the model got the client’s name wrong, and pays the subscription in months when volume drops to almost nothing.
This guide measures AI ROI for one specific small-business workflow, the boring way: count what the work costs today, count everything the AI version costs (including the human parts), and see what is left. You get a copyable worksheet and one worked example with the arithmetic shown. The example business is illustrative; none of its numbers are measured results from Viktor’s work or anyone’s client.
Pick one workflow, then count it before you touch any tool
ROI needs a defined unit of work. “Using ChatGPT for marketing” has no baseline; “writing the follow-up summary after each site visit” does. If you are still choosing which process to look at, the guide to choosing a first workflow to automate covers triggers, owners and review points, and the task-first approach to picking AI tools is a good filter before you price anything.
The baseline needs two numbers: how many times the task happens per month, and how long it takes by hand. Estimating both from memory is how people end up with very optimistic spreadsheets. Time five or ten real instances instead, with a phone timer if necessary. A task that takes 20 minutes on a good day and 70 when the client is difficult has an average that matters more than the good day. Take volume from the last three months too, because small-business work is lumpy, and a workflow that pays at 40 runs a month can lose money at 6.
Everything the AI version costs, including people
The cost side is where most estimates go quiet. It includes:
Setup time. Writing the prompt, building the automation, running trial cases and fixing what breaks. Count it at the hourly cost of whoever does it, even if nobody invoices for it.
Subscriptions. Only the incremental ones. Seats you already pay for don’t count again unless the workflow pushes you onto a higher tier; a new automation plan does.
Usage charges. Chat subscriptions and API usage are billed separately, which surprises people who assume their ChatGPT plan covers an automation calling the API. OpenAI states that ChatGPT and the API platform have separate billing systems, and API charges depend on model, tokens and tools used. Anthropic says the same for Claude: the API is billed separately from paid Claude plans, through prepaid usage credits, although Max and Team plans include some monthly API credits. Automation platforms meter too: Zapier counts tasks and Make counts credits, and the rules for what consumes them are covered in the Make vs Zapier billing comparison.
Review time. Every output a customer will see needs a human read. It doesn’t disappear once you “trust the system”; it is the price of putting your name on the output.
Rework. Some share of drafts will need more than a quick edit. Track the rate and the extra minutes.
Failures. Runs where the output is unusable or the automation didn’t fire, and the person does the task manually anyway, plus the minutes it took to notice.
Maintenance. Prompt updates when your services change, a broken connector, a model update that shifts tone. Budget a monthly figure even if it is small.
Current list prices for common pieces
Prices checked 29 October 2026. US list prices, excluding tax; regional pricing varies. Check the linked pages before you rely on them.
| Product | Billing type | Listed price |
|---|---|---|
| ChatGPT Business, standard seat | Chat subscription per user | $25/user/month monthly, $20 billed annually |
| Claude Team, standard seat | Chat subscription per member (min. 2) | $25/member/month monthly, $20 billed annually |
| Zapier Professional | Automation plan, task allowance | From $19.99/month billed annually, 750 tasks |
| Make Core | Automation plan, credit allowance | From $9/month for 10,000 credits |
| OpenAI API, Claude API | Usage-based, separate from chat plans | Per token at the model’s current rate |
If you are deciding between the two chat assistants on seat cost and admin features, the Claude vs ChatGPT comparison for business use goes through plans in more detail.
The worksheet
Copy this into a spreadsheet, one column per workflow you are evaluating. Use minutes for task time and convert at the end; it keeps the arithmetic honest.
| Line | Input | How to get it |
|---|---|---|
| A | Runs per month | Average of the last 3 months |
| B | Manual minutes per run | Timed sample of 5 to 10 runs |
| C | Review minutes per successful AI run | Timed during a trial |
| D | Failed runs per month | Trial log; unusable output or no output |
| E | Minutes per failed run (noticing + doing it by hand) | Usually B plus a few minutes |
| F | Runs needing rework per month | Trial log |
| G | Extra minutes per rework | Timed |
| H | Maintenance minutes per month | Estimate, then replace with actuals |
| I | Hourly cost of the people involved | Wage plus employer costs, or your own rate |
| J | Incremental subscriptions per month | Only new spend |
| K | Usage charges per month | API or platform usage at real volume |
| L | Setup hours | Build plus testing |
| M | Months to spread setup over | Be conservative; 12 is reasonable |
| Formulas | ||
| N | Manual minutes per month | A × B |
| O | AI-path minutes per month | (A − D) × C + D × E + F × G + H |
| P | Hours freed per month | (N − O) ÷ 60 |
| Q | Value of freed capacity | P × I |
| R | Setup cost per month | L × I ÷ M |
| S | Net monthly capacity value | Q − J − K − R |
| T | Actual cash change per month | Spend that stops − J − K |
Line T is the one people skip, and the next section explains why it deserves its own row.
Time saved is capacity, not cash
If you pay someone a salary and an AI workflow frees ten hours of their month, your payroll expense next month is exactly the same. You have bought capacity. It is worth something only if it goes somewhere: billable work, sales calls, hiring the next person later. It can also quietly evaporate into longer email threads.
So keep three numbers separate. Capacity is hours freed, valued at what those hours cost you. Cash savings is spending that actually stops: a freelancer you no longer book, overtime you no longer pay, a transcription service you cancel. Realised revenue is extra income you can trace to the freed time, such as additional billed hours. A workflow can be strongly positive on capacity and slightly negative on cash at the same time. Presenting the capacity figure as “savings” to a business partner is how AI projects get a bad reputation six months later.
The same caution applies to your business assumptions generally. If the workflow sits inside a bigger plan (a new service line, a hiring decision), the prompts for pressure-testing business plan assumptions are a useful second pass on the numbers you feed in here.
A worked example (illustrative)
Saltmarsh Interiors is a fictional three-person interior design studio. After every site visit, a designer writes a follow-up summary: measurements, decisions, open questions. The proposed workflow: the designer records a voice memo, an automation transcribes it and drafts the summary with an AI model, and the designer reviews it before it goes out. Every figure below is an invented assumption, not a measurement.
| Line | Input | Illustrative value |
|---|---|---|
| A | Runs per month | 40 visits |
| B | Manual minutes per run | 35 |
| C | Review minutes per successful run | 12 |
| D, E | Failed runs; minutes each | 2; 40 |
| F, G | Rework runs; extra minutes each | 6; 20 |
| H | Maintenance per month | 90 minutes |
| I | Hourly cost | $45 |
| J | New subscription | $19.99 (Zapier Professional, annual) |
| K | Usage charges | $6.00 (assumed API spend, not a quoted rate) |
| L, M | Setup hours; months | 14; 12 |
The arithmetic, step by step:
- Manual minutes: 40 × 35 = 1,400 minutes.
- AI-path minutes: (38 × 12) + (2 × 40) + (6 × 20) + 90 = 456 + 80 + 120 + 90 = 746 minutes.
- Hours freed: (1,400 − 746) ÷ 60 = 654 ÷ 60 = 10.9 hours.
- Capacity value: 10.9 × $45 = $490.50.
- Setup cost per month: 14 × $45 ÷ 12 = $630 ÷ 12 = $52.50.
- Net monthly capacity value: $490.50 − $19.99 − $6.00 − $52.50 = $412.01.
On capacity, that is a clear yes. Now line T. Saltmarsh’s designers are salaried, and nothing they paid for before stops. The cash change is −$19.99 − $6.00 = −$25.99 a month. The workflow only pays in cash if the studio uses those ten-odd hours for something that bills, for example one more paid consultation a month. The spreadsheet can’t make that decision for them.
Break-even volume
Average AI-path time per run, excluding maintenance, is (456 + 80 + 120) ÷ 40 = 16.4 minutes. Each run saves 35 − 16.4 = 18.6 minutes, worth 18.6 ÷ 60 × $45 = $13.95. Monthly fixed costs are maintenance (1.5 hours × $45 = $67.50) plus $25.99 in subscription and usage, so $93.49. Break-even is $93.49 ÷ $13.95 = 6.7, so about 7 visits a month. Add the $52.50 of spread setup cost and it becomes $145.99 ÷ $13.95 = 10.5, so 11 visits. Below that, the studio is paying to feel modern. (Usage is treated as fixed for simplicity.)
Sensitivity to errors
The error rate matters more than the drafting speed. Here is the same workflow with worse reliability, everything else unchanged.
| Scenario | Failed runs | Rework runs | AI-path minutes | Hours freed | Net monthly capacity value |
|---|---|---|---|---|---|
| As assumed | 2 | 6 | 746 | 10.9 | $412.01 |
| Shaky prompt | 6 | 12 | 978 | 7.0 | $238.01 |
| Poor fit | 12 | 14 | 1,186 | 3.6 | $82.01 |
(Shaky prompt: 34 × 12 + 6 × 40 + 12 × 20 + 90 = 978. Poor fit: 28 × 12 + 12 × 40 + 14 × 20 + 90 = 1,186. Hours freed are rounded; the dollar values use exact minutes.)
At a 30% failure rate the workflow still looks positive on paper, but it saves under four hours a month while adding interruptions, and none of this counts a wrong measurement reaching a client. One bad summary that leads to a mis-ordered custom sofa could cost more than many months of the “poor fit” net value. That risk doesn’t fit in a cell, so write it next to the table.
Running the measurement in practice
Do a short trial before the spreadsheet decides anything. Two to four weeks of real runs with a simple log (succeeded or failed, review minutes, rework minutes, what went wrong) gives you inputs C to H from evidence instead of hope. The log also shows which mistake keeps coming back, which makes it your best prompt-improvement tool.
Run the numbers again after three months. Maintenance estimates usually start too low, and review time often drops as the prompt improves. If a workflow has gone negative on capacity for two quarters in a row, switch it off. Nobody gives out prizes for keeping an automation alive.
This is a decision tool, not financial advice, and it doesn’t predict returns. It also ignores benefits that are real but hard to price, like faster client replies. Note those in words next to the numbers instead of inventing a dollar value.
Direktorium keeps adding practical guides for small teams working out where AI earns its place. If you would rather have someone build and measure a workflow with you, BBDirector does that as a done-for-you service.
