Before buying any AI tool, check ten things: the job it must do, what that job costs you now, whether your existing software already does it, what data it touches, how it handles ten real examples, total time per job including checking, whether users like it, the data terms, the exit terms, and who owns it.
Most of these checks take minutes. Together they add up to roughly a day of effort spread over two or three weeks. The two that people skip, pricing the current job and reading the exit terms, are the two that decide whether you're still paying for an unused tool this time next year. Each check below comes with a pass condition, so you can tick it honestly rather than hopefully.
Checks 1 to 4: before you look at any vendor
These four happen at your desk with a notepad. They stop you shopping for "an AI tool" in general, which is how businesses end up with three overlapping subscriptions.
1. Write the job down in one sentence, with a number in it
"Help with admin" isn't a job. "Turn site-survey notes into a written kitchen quote, about 30 times a month" is. The number matters because it tells you how much value is at stake and whether per-seat or per-use pricing suits you.
Three rewrites show the difference (all illustrative):
| Business | Vague | One sentence with a number |
|---|---|---|
| Independent optician | "Use AI for patient communications" | "Turn the day's appointment notes into recall letters, about 120 a month, in our standard wording" |
| Bookkeeping practice | "Automate receipts" | "Read photographed receipts into the right expense category in the accounts, about 900 a month across 25 clients" |
| Wedding florist | "Help with marketing" | "Write a caption and a one-line email subject for each finished wedding, about 6 a month, from the photos and the order sheet" |
The florist's sentence, written down, shows the job is six captions a month. That's worth a better prompt in a tool already on the laptop, not a new subscription, and check 2 confirms it in about ten minutes.
Pass if: someone outside the business could read your sentence and know what the tool must produce, from what input, how often.
2. Price the job as it's done today
Time five or six real instances of the job, average them, multiply by the monthly volume and by a realistic hourly cost (wage plus employer on-costs such as payroll taxes and pension contributions, or your own charge-out rate if you do it yourself). This is your ceiling: no tool is worth more than the time it can plausibly free, and most will free only part of it. If you want a fuller method, setting a baseline before you introduce AI walks through it.
Pass if: you have a monthly figure in hours and in money, written down before any demo.
3. Check the software you've already got
Plenty of small businesses hold AI they've never switched on. Every Microsoft 365 business plan comes with Copilot Chat, with no add-on to buy. Gemini now ships inside Google Workspace's business plans too: on Business Starter it's limited to Gmail and the Gemini app, while Business Standard and upwards spread it across Docs, Sheets, Meet and the rest. Your CRM, accounting package or quoting software may have added AI features in its last update, sometimes with a price rise attached.
Pass if: you've tried the job in at least one tool you already own and noted how close it got.
4. Decide which data the tool will touch
Sort the inputs into four buckets: public (your website text), internal (price lists, templates), client-confidential (drawings, contracts) and personal data (names, addresses, phone numbers). The bucket decides the plan you need. Anything beyond public and internal should go only into a business plan that doesn't train on your content by default, with an agreement covering how the vendor processes personal data. Classifying your business data before using AI tools gives a fuller scheme.
Pass if: you can say, in one line, the most sensitive thing the tool will ever see.
Checks 5 to 7: while you test
Only now do you look at vendors. Shortlist two at most, plus the "already own it" option from check 3, and put all of them through the same test.
5. Test on ten real examples, including two awkward ones
Pull ten past instances of the job from your files. Eight should be typical; two should be the ones that caused trouble (an odd layout, a missing measurement, a client who changed their mind twice). Run every candidate on the same ten. Demos use the vendor's cleanest data; your awkward cases are where the difference shows. Running a two-week AI tool trial before you commit covers how to structure the test period itself.
Choosing the awkward two takes some thought. The bookkeeping practice from check 1, testing a receipt-reading tool, picked a fuel receipt photographed at an angle in a dark car and a credit note from a supplier. The tool read the angled receipt correctly. It entered the credit note as a $212 expense rather than a $212 refund, and the preview screen gave no hint anything was wrong; the error only showed when the test client's supplier balance came out $424 adrift. Had the practice tested on ten clean invoices, it would have bought a tool that quietly reverses every credit note, and found out at a client's year end.
Pass if: you have a score for each example: usable as is, usable after light edits, or had to redo.
6. Time the whole job, not the AI's part
A draft that appears in 20 seconds and then needs 40 minutes of fixing hasn't saved anything. Time from "I start the job" to "it's ready to send", including pasting inputs in, reading the output, correcting it and reformatting. Compare that with the figure from check 2.
Pass if: the average end-to-end time is clearly lower than today's, and no example took longer than doing it by hand.
7. Put it in front of the person who'll actually use it
Owners often test tools alone and then hand them to staff who find them awkward. Sit the real user down for an hour with the shortlisted tool and three of the test examples. Watch where they hesitate. Ask afterwards: would you choose to use this on a busy Tuesday?
Pass if: the daily user says yes without being steered, and can do the job without you in the room.
Checks 8 to 10: before you pay
8. Read the data terms for the plan you'll buy
Check four things on the vendor's own pages: whether your content is used to train models on that plan, how long it's kept, whether you can delete it, and where the setting lives if training is on by default. On a consumer ChatGPT or Claude plan, each user has to opt out of model training themselves, in the privacy settings. The business tiers (ChatGPT Business, Claude Team, Microsoft 365 Copilot, Gemini in Workspace) keep company content out of training as standard. For smaller vendors, the tutorial on what to check in an AI tool's privacy policy and terms lists the clauses to find.
Pass if: you've read the actual terms for your plan, not the marketing page, and they match the data bucket from check 4.
9. Check the exit before the entrance
Find out four things: the minimum commitment, whether you can reduce seats mid-term, the notice period, and what you can take with you when you leave. The seat rules catch people out. ChatGPT Business seat reductions take effect at the start of the next billing cycle. On a Google Workspace annual plan you can't reduce licences until renewal. Microsoft lets many business customers remove licences only within seven days of buying or renewing. Meanwhile monthly billing costs more per seat: $25 against $20 for ChatGPT Business or Claude Team Standard, and $25.20 against $21 for Microsoft 365 Copilot Business. For export, ask whether your templates, prompts, generated documents and settings can leave in a usable format.
Vendors often answer exit questions in a sentence that sounds complete and isn't. An illustrative reply from a small AI scheduling vendor, and the follow-ups it needs:
VENDOR: "You own your data and can export it at any time. We're
flexible on terms for growing businesses."
FOLLOW-UP QUESTIONS
1. Export in what format? CSV of bookings, or also our reminder
templates, custom fields and message history?
2. Is export available after we give notice, and for how long?
3. How many days after cancellation is our data deleted, and will
you confirm deletion in writing?
4. "Flexible on terms": is the 12-month minimum in the order form
negotiable to monthly, and can we drop seats mid-term?Get the answers by email, not on a call, and file them with the pre-purchase record below. A vendor that won't put "monthly, cancel with 30 days' notice" in writing has told you what the exit will be like.
Pass if: you know the date you could next cancel or shrink, and what you'd lose on that date.
10. Name an owner, a review date and a kill rule
One person owns the tool: they manage seats, answer questions and run the review. Put a review in the diary for 60 to 90 days out, and write the rule that would make you cancel before you've grown attached. "Cancel if average quote time isn't under 35 minutes by the second month" is a kill rule. "See how it goes" is not.
Pass if: the owner's name, the review date and the kill rule are written on the same page as the price.
Worked example: a kitchen fitter weighing an AI quoting add-on
Here's how the list plays out for an illustrative six-person kitchen fitting firm: the owner, an office manager, a designer and three fitters. The designer turns survey notes and measurements into written quotes. Their quoting software now offers an AI add-on; for the example, say it's $79 a month billed monthly, or less on an annual contract.
| Check | What they found |
|---|---|
| 1. The job | Draft a written quote from survey notes and measurements, about 30 a month. |
| 2. Current cost | 90 minutes per quote, so 45 hours a month. At a $32 loaded hourly cost, about $1,440 a month. |
| 3. Already owned | Microsoft 365 Business Standard, so Copilot Chat was already available. With a pasted template it produced usable drafts for 7 of 10 examples, at about 40 minutes each end to end. |
| 4. Data | Client names, addresses and floor plans: personal and client-confidential, so business-grade terms only. |
| 5. Ten examples | The add-on gave usable drafts for 8 of 10. Both failures were awkward cases: it left out end panels on an L-shaped run and priced the wrong worktop edge profile. |
| 6. End-to-end time | About 25 minutes per quote with the add-on, including checking. |
| 7. Real user | The designer preferred the add-on because it pulled unit codes straight from the price list. |
| 8. Data terms | Customer data stays inside the quoting system they already use; no model training on customer content, per the vendor's terms. |
| 9. Exit | Monthly option available at the higher price. Quotes export as PDF only, so templates would need rebuilding if they left. |
| 10. Owner and rule | The office manager owns it. Review after 90 days. Cancel if average quote time isn't below 35 minutes by month two, or if more than one quote a month goes out with a pricing error. |
The arithmetic: the add-on saves about 65 minutes a quote, roughly 32 hours a month, against Copilot Chat's saving of about 25 hours for nothing extra. The extra seven and a half hours are worth about $240 a month at their hourly cost, so the add-on wins, by roughly $160 a month after its own price: a far smaller margin than comparing it with quoting by hand would suggest. They bought monthly, added a rule that the designer checks every end panel and edge profile by hand, and moved to annual billing only after the 90-day review came back clean. Without check 3, they'd never have known the free option was close; without check 5, the edge-profile error would have reached a customer.
A pre-purchase record you can copy
Keep one of these per tool, in the same shared folder. It takes 15 minutes to fill in and becomes the starting point for the review in check 10.
AI TOOL PRE-PURCHASE RECORD
Tool and plan:
Date: Filled in by:
1. The job (one sentence, with volume):
2. Current cost: ___ min per job x ___ per month = ___ hours = $___ /month
3. Tried in software we already own? Tool: ___ Result: ___
4. Most sensitive data it will see: public / internal / client-confidential / personal
5. Ten-example test: usable ___ light edits ___ redo ___
Awkward cases and what went wrong:
6. Average end-to-end time per job: ___ min (was ___ min)
7. Daily user's verdict (their words):
8. Trains on our content? Y / N Retention: ___ Setting location: ___
9. Minimum term: ___ Seat reductions allowed: ___ Notice: ___
What we can export on exit:
10. Owner: ___ Review date: ___
Kill rule: cancel if ___
Decision: buy monthly / buy annual / don't buy / retest after ___Signs you're buying for the wrong reason
Some purchases pass a checklist on paper and still go wrong because the reason behind them was weak. Watch for these:
- A deadline on the discount. "Price rises on Friday" pushes you past checks 5 to 7. A tool worth buying on Thursday is still worth buying at the higher price, and if it isn't, it wasn't worth it.
- The demo was the test. If nobody ran your own examples, you've evaluated the vendor's sales material.
- A lifetime deal from a young company. A one-off payment is only cheap if the company is still running in three years and the product still does your job.
- It solves a problem you don't have in volume. A tool that saves 20 minutes on a task you do twice a month is a hobby, whatever the price.
- Nobody could name the owner. Unowned tools drift: seats stay assigned to people who've left, and nobody notices the renewal.
When a member of staff asks you to buy something
The same ten checks work as an approval form. Ask the person requesting the tool to fill in checks 1, 2, 5, 6 and 7 themselves, since they know the job, and do 3, 4, 8, 9 and 10 yourself or with whoever handles IT. That split keeps enthusiasm and scrutiny in the right hands. If requests are becoming frequent, a lightweight AI tool approval process turns this into a routine rather than a debate each time. And when you get to the vendor conversation, the questions to ask an AI vendor before you sign pick up where check 9 leaves off.
Buying questions owners ask at this stage
Is a vendor's free trial enough to judge an AI tool?
Only if you spend it on your own work. A trial spent clicking through demo content tells you very little. Book two or three hours inside the trial window, run your ten real examples, time the whole job including checking, and let the person who will use it daily try it. If the trial runs out before you finish, ask for an extension rather than paying to finish testing.
Should I take the annual discount to save money?
Not for a tool nobody has used for a month of real work. Annual billing is cheaper per seat (ChatGPT Business is $20 per user a month billed annually against $25 monthly), but ChatGPT Business, Google Workspace annual plans and Microsoft annual subscriptions generally don't let you cut seats until renewal. Start monthly, confirm usage for 60 to 90 days, then move only the seats you're sure of to annual.
What if staff already use a free AI tool for this job?
Treat it as evidence, not a problem to stamp out. Ask what they use it for, how often, and what they paste in. That tells you the real job and the data involved, which are checks one and four on this list. Then decide whether to move that use onto a business plan that doesn't train on your content, and say clearly what may go into the free tool in the meantime.
Further reads
- Annual or Monthly Billing for AI Tools: Which Saves You More? — Work out when annual billing actually saves money.
- AI Features Already in Your Software: What to Switch On First — Switch on the AI in your current software before adding a tool.
- AI Washing: How to Spot Software That Oversells Its AI — Tell real AI features from relabelled old ones.
- Is That AI Tool a ChatGPT Wrapper? How to Check Before You Pay — Check whether you're paying extra for a thin layer over ChatGPT.
- What If Your AI Vendor Shuts Down? Checks Before You Commit — Plan for a small vendor disappearing mid-contract.
- How to Review an AI Tool After 90 Days: Keep, Fix or Cancel — The keep, fix or cancel review to book on day one.
- How Many AI Tools Does a Small Business Actually Need? — Why most small firms need two or three AI tools, how to spot paying twice for one job, and three businesses' stacks with monthly costs.
- What Is an API? Why It Matters When You Buy Software — A plain-English explanation of APIs for people buying software, with the vendor questions and pricing traps that decide whether tools can connect.
- What to Ask Before Buying Any AI Tool for a Medical Practice — A grouped checklist of questions for any AI vendor selling to a medical practice, what a good answer looks like, a scored example and the red-flag answers.
- AI Estimating Software for Small Builders: Six Options Compared — Six AI estimating options for small building firms compared on what the AI does, list price, best fit and the catch, plus a way to test them on a real job.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI ChatGPT Business pricing and seat-management help pages; Microsoft Learn, Buy or remove licenses for a Microsoft business subscription; Google Workspace Help, Compare Flexible and Annual/Fixed-Term payment plans; Microsoft 365 Copilot Business pricing.