The 5-step framework for evaluating any AI tool.
Before you buy another subscription, run the tool through five plain questions. You do not need a technical background. You need a pen and about twenty minutes with the vendor website.
Most tools fail at question two. We will tell you why, and what a good answer sounds like.
Five questions. No jargon. Walk away if the answers are mushy.
- 1. Training. Is my data used to train your models?
- 2. Access and exit. Who can see my data, and can I get it back?
- 3. Mistakes. What happens when the AI gets something wrong?
- 4. Portability. Can I take my work with me if I leave?
- 5. Outgrowing. What does it cost when I outgrow the free tier?
Know your own rules first.
You cannot judge a vendor answer until you know what you would accept. Take ten minutes before you read a single pricing page and write down three things.
What data must never leave the building?
Customer lists, contracts, health or payment details, anything you promised someone you would keep private. Be specific. "Sensitive stuff" is not a rule.
Who approves a new tool?
One named person who can say no, and the no sticks. Tools that arrive through a free sign-up by one enthusiastic employee are how data ends up in places nobody chose.
What do you promise your customers?
In writing or by habit. If you tell clients their files stay in-house, the tool has to fit that promise, not the other way round.
Not sure where you stand? The free readiness self-check has a policies section for exactly this. It runs in your browser and nothing you tick is sent anywhere.
Ask these before you pay.
Put every question to every tool, including ours. A good vendor answers each one in plain words, in writing, without a sales call. A vendor who dodges a question has answered it.
1. "Is my data used to train your models?"
The big worry, and a fair one. What you want to hear: no, not on the plan you are buying, and here is the page that says so. Watch for the quiet version of yes, such as "we may use content to improve our services." That sentence can mean training. Ask which plan changes the answer, because many tools treat free and paid accounts differently.
2. "Who can see my data, and can I get it back?"
This is the question most tools fail. It has two halves, and a vendor has to pass both. Who sees it: named roles, not "authorized personnel." Get it back: a real export you can open, and a real delete you can confirm. The next section explains why this one trips people up.
3. "What happens when the AI gets something wrong?"
The question almost nobody asks, and the most important one. AI tools get things wrong with total confidence. Ask whether a person checks the output before it reaches your customer, whether the tool keeps a record of what it did, and what the vendor does when a mistake slips through. "Our model is very accurate" is not an answer. A trail you can read is.
4. "Can I take my work with me?"
Over a year, the prompts you tune, the templates you build and the notes you store become the real asset. Ask whether you can export them in a format you own, and whether they still work on a different AI model. A tool that keeps your work behind its own login has quietly become your landlord. Good work is portable: the instructions, the checking rules and the examples live in files you hold, so the engine underneath can change.
5. "What does it cost when I outgrow the free tier?"
Free is a trial, not a plan. Ask what the limit is (messages, seats, files, history), what happens at the limit, and what the next step costs for your real usage, not the demo. Then ask the sneaky follow-up: does anything you get today, like export or longer history, move to a higher plan? Do the math for a team of your size in a busy month before you build a habit around it.
Why most tools fail question two.
Not because vendors are villains. Because the honest answer is usually inconvenient, and a vague answer is cheaper than a precise one. Here is the pattern we see most often.
"Who sees it" gets a fog answer
You hear "our team" or "authorized personnel" and never a list of roles. Ask which humans can open your content, under what conditions, and whether outside providers touch it too.
"Get it back" means a screen view
Many tools let you view your history but not export it in a form another tool can read. If you cannot download it, you do not really have it.
"Delete" can mean "hide"
Removing a chat from your screen is not the same as removing it from the vendor systems. Ask what is actually erased, how long it takes, and whether you get a confirmation.
It is buried, so nobody asks
The answer, when it exists, sits in a policy page nobody reads. That is why you ask in writing and keep the reply.
A passing answer is short: the named roles that can see your data, a copy of everything on request, and a deletion you can confirm. We hold ourselves to the same bar, and our own plain-words answers are on the AI safety basics page.
How to read the answers.
Give each question one of three marks: clear, fuzzy or missing. You do not need a spreadsheet.
Clear on all five
Safe to start small. Pick one low-risk task, run it for two weeks and check the output by hand.
Fuzzy on one or two
Ask again in writing. A good vendor sharpens the answer. A weak one repeats the brochure.
Missing on question two, or on any three
Walk away, or keep sensitive data out of it entirely. A cheaper tool that holds your data hostage is not cheaper.
Questions four and five echo how we work. We are vendor-agnostic and pick the engine that fits your problem, and the structure we help you build is meant to be portable. See the method for how portable prompts and evaluation rubrics fit in.
Have a tool you are unsure about?
Send us its name and the answers you could find. We will tell you where it is fuzzy and what to ask next. No pitch attached.