AI is worth it for a small business when it saves or earns more than its full cost, including the time spent checking its work, within about three months. A quick test: multiply the hours of repetitive reading and writing it would remove each month by what an hour of that person's time is worth, then compare with the monthly cost.
For most firms that test says yes to a general AI assistant on a business plan, and "not yet" to bigger automation projects until one job has proved itself. The answer flips to no when checking eats the gain: a quote or contract summary that takes as long to verify as it would have taken to write saves nothing, however quickly it appears on screen.
One thing to rule out first: you may already be paying for some AI. Microsoft 365 business plans include Copilot Chat at no extra cost, and Gemini is built into Google Workspace business plans. Check what your existing subscriptions cover before buying anything new; if you use Microsoft 365, what you get free with Copilot Chat sets out the difference from the paid add-on.
The three-number test
You need three figures, rough is fine:
- H: hours saved per month. Hours of repetitive reading, writing, sorting or summarising the AI would remove, after subtracting the time spent checking its output.
- V: value of an hour. For staff, their pay plus the other costs of employing them (use your payroll figures, not the hourly wage alone). For the owner, what you'd earn or protect by spending that hour on something else.
- C: monthly cost. Subscriptions, automation-platform charges, and the time someone spends looking after it.
AI is worth it when H × V is comfortably bigger than C. Aim for at least twice C, because first estimates of time saved are nearly always optimistic.
V and C are where rough figures go most wrong, so work one through. Say a receptionist is paid $16 an hour, and your payroll shows that pension contributions, holiday pay and employer charges add roughly a quarter on top. Her hour is worth about $20, not $16. On the cost side, two business seats at $25 a month look like C = $50. But if the office manager spends two hours a month updating the price list the AI drafts from, answering questions and checking the bill, and her hour is worth $25, that's another $50. The real C is $100, and the break-even doubles with it. Most "it obviously pays" sums skip that second $50.
The table shows break-even: how many hours a month the AI must save just to cover its cost, at different costs and hourly values. Double each figure for the safety margin.
| Monthly cost | At $15 an hour | At $25 an hour | At $40 an hour | At $60 an hour |
|---|---|---|---|---|
| $20 (one paid assistant seat) | 1.3 hours | 0.8 hours | 0.5 hours | 0.3 hours |
| $50 (two business seats) | 3.3 hours | 2 hours | 1.3 hours | 0.8 hours |
| $150 (a few seats plus automation) | 10 hours | 6 hours | 3.8 hours | 2.5 hours |
| $400 (a team setup with tools) | 26.7 hours | 16 hours | 10 hours | 6.7 hours |
Two things jump out. A single $20 seat needs to save barely an hour a month, which almost any regular user manages, so the question at that level is really whether someone will use it. And the bar rises quickly: a $400 monthly setup needs a team saving several hours a week between them. For a fuller calculation across a year, including setup costs, see how to calculate AI ROI with a worked example. For current prices across the main tools, what AI costs a small business in 2026 has the figures.
What the quick sum leaves out
The three-number test flatters AI unless you adjust for these. Deal with each honestly and your answer will hold up.
- Checking time. Every draft needs reading. Until you've measured it, assume checking takes about half the time the job took before, then correct the figure after your trial.
- The learning dip. The first two to four weeks are often slower than the old way while people learn what to ask and what to trust.
- Time saved that evaporates. Five minutes saved here and there often disappears into the day. Count only time that turns into something: fewer overtime hours, the owner's evenings back, more bookings handled. Why AI saves time but not money explains how to capture it.
- The cost of mistakes. A wrong price or promise sent to a customer costs goodwill and sometimes money. Where mistakes are expensive, the checking step has to be thorough, which cuts H.
- Subscription creep. Tools bought for one job keep billing after the job changes. Put a review date on every subscription.
- Upkeep. Price lists, opening hours and policies the AI drafts from must be kept current, which is a small but permanent task.
Checking time is the adjustment that most often turns a yes into a no. Take an illustrative three-site café group whose manager reads new supplier contracts for price-review and termination clauses. Reading one properly takes about 30 minutes. An AI summary appears in under a minute, but the summary is only useful if it's right, so the manager checks every clause it mentions against the contract and skims for anything it left out: about 25 minutes. At three contracts a month, the saving is around 15 minutes. That's a no for this job, even though the tool is fast. The same seat might still pay on a different job where a wrong answer is cheap and easy to spot.
Evaporating time needs a destination named in advance. A two-person bookkeeping practice that saves about 20 minutes a day on drafting client emails will, left alone, spend it on whatever turns up next. Their fix was to point it at one job: the saved time goes on chasing clients for missing receipts, and the test is whether the "waiting on client" list is shorter at month end. If it is, the time turned into something real. If it isn't, the saving was probably smaller than it felt.
Subscription creep is easiest to see on a bank statement. One illustrative events company found a meeting-notes tool at about $17 a month still billing six months after the weekly meeting it was bought for had stopped: roughly $100 for nothing, found in five minutes by searching the statement for recurring payments.
Where the answer is usually yes, and usually no
Usually yes:
- Owner-run businesses where the owner's time is the bottleneck, because the owner's hour is usually the most valuable one in the firm.
- Firms with a lot of written customer contact: enquiries, quotes, follow-ups, reviews.
- Businesses with large product or service catalogues that need descriptions and updates.
- Anyone whose staff already copy information between documents, emails and spreadsheets for hours a week.
Usually no, or not yet:
- Work that is almost entirely hands-on, with little admin or writing.
- Very low volumes: a task done twice a month rarely repays setup and learning time.
- Chaotic processes, where the real fix is agreeing one way of doing things first.
- Jobs involving data you can't put into any tool you're able to secure properly.
- Buying a specialist AI product before trying a general assistant on the same job.
Some jobs also shouldn't go to AI however cheap it is; when not to use AI in your business lists nine to keep human.
Three ten-minute worth-it sums you can copy
These are illustrative sums of the kind you can do for your own business in ten minutes.
A barber working alone
He spends about three hours a month writing social captions and replying to booking messages. A $20 assistant could cut that by about two hours, net of checking. Say his hour is worth $35, because the time he saves can go on an extra cut: 2 × $35 = $70 against $20, so a clear yes. He also considered an AI phone-answering service; say it was quoted at $100 a month. He gets few calls, most are from regulars, and he can reply between clients; saving perhaps an hour a month, it would return $35 against $100. That's a no for now.
One detail changes his sums if customer messages go into the assistant. Business plans such as ChatGPT Business and Claude Team need at least two seats, so a one-person business pays for two, about $40 to $50 a month, or stays on an individual plan such as ChatGPT Plus or Claude Pro and switches off the model-training setting in its privacy settings. At $40 his yes still holds ($70 against $40), but with less room, which is why the safety margin matters for sole traders most.
A dry cleaner with two branches
The manager handles roughly 150 emails a month about collections, stain questions and alteration quotes, plus a daily scramble to match tickets to queries. AI sorting and drafted replies could save about eight hours a month after checking. At $25 an hour that's $200 of value. Costs: two business seats at $25 a month billed monthly plus an entry automation plan at $19.99 a month billed annually, about $70. That's nearly three times the cost, which clears the safety margin. Yes.
A pet shop adding new lines to its web shop
About 40 new products a month each take 15 minutes to describe. With AI drafting from the supplier's spec sheet, each takes about 5 minutes to check and correct. That's 400 minutes, or about 6.7 hours, saved a month. At $20 an hour, around $133 of value against one $25 seat. Yes, and that's before counting any sales from getting products listed a week or two sooner.
When the value is extra sales, not saved time
Some AI jobs pay back through revenue rather than hours: replying to enquiries within minutes instead of a day, following up every quote, listing products sooner. That value can be larger than the time saved, but it's also the easiest to overestimate, because sales go up and down for many reasons.
Count it only if you can measure it against a baseline. For enquiry replies, that means knowing what share of enquiries became bookings before the change, then comparing the same months afterwards, not a quiet February against a busy May. If the share rises from, say, 30% to 36% on 100 enquiries a month, that's six extra bookings you can put a value on. Even then, count only half of the uplift in your first calculation, since some of it would probably have happened anyway. If the maths only works with the revenue included, treat the answer as "probably" rather than "yes", and give it a firm review date.
Settle it with a two-week test
Estimates only go so far. Before committing to anything beyond one seat, run a short test on the job that looked most promising:
- Days 1-3: time ten instances of the job done the normal way. Note the start and finish.
- Days 4-5: set up the AI with clear instructions and your reference material, such as the price list or product spec sheets. Use a business plan or a trial of one for anything involving customer information.
- Days 6-14: do the next 15 to 20 instances with AI, timing them including all checking and corrections. Note any mistake that would have mattered.
- Day 14: work out the real minutes saved per instance, multiply by your monthly volume, and redo the three-number test with the measured figure.
Filled in, a test sheet for an illustrative tutoring agency that writes end-of-term progress notes to parents might end up like this:
| Item | Result |
|---|---|
| Job | Progress note to parents from the tutor's session notes |
| Baseline (10 notes, normal way) | 18 minutes average |
| With AI (18 notes, including checking) | 9 minutes average |
| Mistakes that would have mattered | 1: a note used the previous pupil's name, because the tutor reused the same chat |
| Monthly volume | About 60 notes |
| H | 9 minutes × 60 = 9 hours |
| V | $22 an hour (tutor time, payroll figure) |
| C | Two business seats ($50) + 1 hour of upkeep ($22) = $72 |
| H × V against C | $198 against $72: about 2.75 times |
That clears the twice-C bar, so the answer is yes, with one change made before rollout: a fresh chat, or a saved project with no previous pupil in view, for every note. The wrong-name mistake is exactly the kind a two-week test should surface, because it would have reached a parent within a month of normal use.
If the measured saving is less than an hour a month, the answer for now is no. That's a perfectly good outcome: you've spent two weeks and very little money finding out, rather than a year paying for something nobody uses. For realistic ranges to compare your results against, see how much time AI can really save each week.
Buy, wait or test another job: the rule after two weeks
AI is worth it for your business if, after a two-week test, the hours it saves multiplied by their value come to at least twice its full monthly cost, and nothing in the job would cause serious harm if the AI got something wrong and nobody noticed. If both hold, buy it for the people who tested it and review in three months. If either fails, don't buy yet; pick a different job and test again, or wait until your processes are ready for it.
Further reads
- Do I Need AI in My Business? A Decision Guide for Owners — A wider decision guide if cost isn't your only question.
- Pros and Cons of Using AI in Your Business, With Real Costs — The non-financial trade-offs set out side by side.
- Who in Your Team Actually Needs a Paid AI Licence? — Work out which staff actually need a paid seat.
- Is Paying for AI Help Worth It? Your Time vs an Expert's Fee — When outside help pays compared with doing it yourself.
- AI Myths That Stop Small Businesses Getting Started — Clears up assumptions that skew the maths either way.
- How to Set a Baseline Before You Introduce AI — Measure the job properly before your two-week test.
- Is AI Worth Paying For in a Small Tutoring Business? — Which single AI subscription a tutoring business should pay for, the jobs that earn it back, and the checking time that decides it.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: ChatGPT, Claude, Zapier, Microsoft 365 and Google Workspace pricing and plan pages (checked September 2026).