How to Test a New Service Idea With AI Before You Launch It

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Test a New Service Idea With AI Before You Launch It.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Test a New Service Idea With AI Before You Launch It.

Use AI to sharpen the idea and draft the tests, then get the evidence from real people. Write the idea as a priced offer, list the assumptions that must hold, run ten short conversations about how customers handle the problem now, and publish a page with a real price and a sign-up button. Set the pass mark first.

The trap is asking an assistant "would customers want this?" It will usually say yes, with reasons, because it's reasoning from general text and tends to agree with the person asking. That answer feels like research and isn't. Evidence has a pecking order: compliments are worth almost nothing, an email address is worth a little, and money down, even a refundable deposit, is the only signal that reliably predicts sales.

Follow me on Instagram@sagnikteaches

Turn the idea into an offer someone could buy today

Most ideas are too vague to test. "Something for puppies" can't fail, so it can't pass either. A testable version names who it's for, the problem, what they get, the price and what they'd do instead. This fill-in line forces all five:

Connect on LinkedInSagnik Bhattacharya
For [who], who [problem in their words], we offer [what exactly],
for [price], so they [result], instead of [what they do now].

Consider an illustrative case: a two-person dog grooming salon whose owner has an idea for young puppies. Her first attempt read: "For new puppy owners, we offer a puppy grooming programme." She then gave an assistant her notes and asked for three sharper versions with different formats and prices:

Subscribe on YouTube@codingliquids
Here is a rough idea for a new service at my dog grooming salon: [notes].
Write three versions of it as a one-line offer using this pattern:
For [who], who [problem], we offer [what], for [price], so they [result],
instead of [what they do now].
Vary the format and price between versions. Keep each under 45 words.
Don't claim benefits I haven't mentioned in my notes.

What came back (illustrative):

1. For owners of puppies aged 12-20 weeks, whose puppy panics at the dryer and clippers, we offer three short, calm sessions over six weeks for $95, so their dog's first full groom isn't a battle, instead of waiting until the coat is matted.

2. For new puppy owners, a single 45-minute "meet the salon" visit with a bath and nail trim for $35.

3. For busy families, a monthly puppy membership: one short visit a month until 12 months old, $30 a month.

Version 1 was the strongest because the problem is specific and observable. Version 3 had slipped in "busy families", an audience her notes never mentioned, so she deleted it. She kept 1 as the offer to test and 2 as a fallback if $95 proved too high.

List the assumptions that would sink it

Every service idea rests on a few beliefs that have to be true. A quick pre-mortem, imagining the launch has flopped and asking why, surfaces them. This is a job AI does well, because it's good at generating possibilities you'd rather not think about:

Imagine it's six months after launch and this service has failed.
Offer: [your one-line offer].
List the 10 most likely reasons it failed, from the customer's point of
view and from the business's point of view. For each, name the
assumption that turned out to be wrong.

From the ten it produced, the groomer kept five and scored each on two things: how badly it would hurt if wrong, and how sure she already was. The ones that are both damaging and uncertain are what the test has to answer.

AssumptionDamage if wrongHow sure nowTest
Owners see puppy nerves as a problem worth paying to fixHighLowInterviews
They'll pay around $95 for three sessionsHighLowPriced smoke test
Enough puppy owners find the salonMediumMediumSmoke-test traffic
Short sessions fit around existing bookingsMediumHighCheck the diary
Puppies can safely visit at 12 weeksHighMediumAsk the vet practice she works with

The last row is a reminder that some assumptions aren't market questions at all. An assistant can't tell you when puppies are safe to be around other dogs; a vet can. A pre-mortem is also useful before any project, not just a new service; running a pre-mortem before an AI project uses the same idea.

Desk research AI can speed up, and where it stops

An hour of desk research with an assistant that can search the web is worth doing before you talk to anyone. Ask it to find:

  • Similar offers elsewhere: do other groomers sell puppy introduction sessions, and how do they describe them? Ask for links so you can read the originals.
  • What people already do instead: videos, home brushing, waiting until the first full groom.
  • The questions owners ask online in forums and comments, which tell you the words to use.
  • What nearby rivals offer, from their own websites. A one-afternoon competitor review covers doing this properly.

Where it stops: any figure for "market size" or "demand around your salon" is a guess unless it comes with a checkable source. An assistant asked how many puppy owners live within five miles will produce a number; there is no reliable way it could know. Treat desk research as a map of what exists, not proof anyone will buy yours.

Why "would customers want this?" is the wrong prompt

Here's the difference in practice. The groomer first asked: "Would dog owners want a puppy grooming programme?" The answer ran to 300 words of encouragement: growing pet ownership, owners who treat dogs like family, the value of early socialisation. None of it was about her salon or her customers, and none of it could be wrong.

Then she asked this instead:

Play a sceptical new puppy owner who has just seen this offer: [offer].
Give me the 8 most likely reasons you would NOT book, in your own words.
Then list the questions you'd want answered before paying.

That produced objections she could actually test: "Can't I just do this at home with treats?", "What if my puppy isn't fully vaccinated yet?", "Three sessions sounds like a lot of driving." Simulated customers are good for exactly this, generating objections to prepare for. They are poor at telling you how many real people will pay, and an agreeable simulated panel can lull you into launching something nobody wants. For how far to trust AI-made customer profiles generally, see whether AI-generated personas are accurate.

Ten conversations about the past, not the future

People are bad at predicting what they'll buy and very good at describing what they did. So the interviews ask about the last time the problem happened, not whether they'd like your idea. This is the approach Rob Fitzpatrick sets out in his book The Mom Test, and AI can draft the guide in a minute:

Write a 10-minute interview guide for new puppy owners. I want to learn
how they handle grooming and handling nerves NOW. Rules: ask about past
behaviour and specific recent events, never ask if they'd buy something,
don't mention my offer until the last question. Max 7 questions.

The guide she used, after editing:

  1. Tell me about the last time you tried to brush, bath or trim your puppy. What happened?
  2. What have you tried so far to get them used to being handled?
  3. Where did you look for help or advice? What did you find?
  4. Have you paid for anything to help with it, such as classes, a trainer or products? How much?
  5. When are you planning the first full groom? What worries you about it?
  6. Who else in the household deals with this?
  7. (Last) If a salon offered short, calm sessions to get puppies used to grooming, what would you want to know?

She found interviewees through existing customers with new puppies, the vet practice's puppy evening, and a post on the salon's Instagram. Eleven conversations took about three hours spread over a fortnight. With permission she recorded them on her phone, transcribed them, and asked for analysis:

Below are transcripts of 11 interviews with puppy owners.
1. List every problem mentioned, how many people mentioned it, and quotes.
2. List what they currently do and spend.
3. Separate what people DID from what they SAID they'd do.
Don't summarise beyond what's in the transcripts. Give interview numbers.

A trimmed version of the output (illustrative): "Dryer panic: 7 of 11 (interviews 1, 2, 4, 5, 8, 9, 11). Nail trims a struggle: 8 of 11. Already paid for help: 4 of 11, mostly puppy classes at $80-$120 for a course. Said they'd book a salon session: 9 of 11; did anything towards it: 0." That last split is the one that matters. Nine people said yes, which means little; four had already spent around $100 on a related problem, which means a lot. The count for nail trims was also one too high: interview 6 mentioned a previous dog, not the puppy, which she caught by reading the quotes. For larger volumes of feedback, analysing customer feedback surveys with AI covers the method at scale.

A smoke test with a real price and a real button

A smoke test puts the offer in front of people as if it's about to launch, with a real price, and measures how many take a concrete step. The honest version says "launching on [date], register interest" or "reserve a place with a refundable $20 deposit". Choose one or two channels you already have:

  • A one-page web page with the offer, price, what's included, the launch date and a form. Building a landing page with AI covers the page itself.
  • A social post with the offer and a link or a "comment to register" prompt.
  • A counter card with a QR code to the same page, for existing customers.
  • A small paid test, say $50-$100 of ads to people nearby, if you need more traffic than your own channels bring.

AI can draft the page copy from your offer line and interview quotes. The prompt that worked for the groomer: "Write a 150-word page for this offer, using the owners' own phrases from these quotes. State the price, the three sessions, the ages it suits, the launch date, and that deposits are refundable. No exclamation marks, no claims I haven't given you." The first draft promised the sessions "eliminate grooming anxiety"; she changed it to "make the first full groom calmer for most puppies", which she could stand behind.

Set the pass mark before the results arrive

Decide in advance what counts as a pass, a rethink and a stop, and write it down. Otherwise any result feels encouraging after the fact. Work it back from break-even: the groomer calculated that eight programme sales a month would justify the time, so her thresholds for a four-week test were:

Signal over four weeksGoRethinkStop
Refundable $20 deposits6 or more2-50-1
Registrations of interest (no money)25 or more10-24Under 10
Visitors to the page who register5% or more2-5%Under 2%

Deposits outrank registrations in the table: if deposits hit the go mark, a low registration count doesn't matter.

One mistake worth avoiding, from an earlier attempt at the same test: the first version of her page left the price off, on the theory that people would register and ask. Forty people registered in a week, and when she emailed them the $95 price, two replied. A registration without a price measures curiosity, not demand, so every smoke-test page states the price in the first two lines.

The groomer's six weeks, end to end

WeekWhat happenedTime spent
1Offer line, pre-mortem, desk research, interview guide; asked the vet about minimum age4 hours
2-3Eleven interviews; AI analysis of transcripts4 hours
3Page written and published; counter card printed; Instagram post3 hours plus $25 printing
3-6Smoke test runs; $60 of local ads in week 430 minutes a week
6Results read against the pass marks1 hour

The results (illustrative): 380 page visits, 27 registrations (7.1%), and 9 deposits. All three signals hit the go column. Six of the nine deposits came from existing customers via the counter card, which told her something else useful: the first launch should be marketed to her own customer list, and the paid ads, which brought 140 visits and one deposit, weren't worth repeating. In total the test cost about 13 hours and $85, against the cost of equipping and marketing a service that might have sat empty.

Reading results that aren't a clean yes or no

Most tests come back mixed. Three patterns and what they usually mean:

  • Plenty of interest, few deposits. People like the idea but not enough at this price or in this format. Rerun with the cheaper fallback offer, keeping everything else the same.
  • Interest from a different group than expected. If half the registrations are for nervous older rescue dogs, not puppies, that may be the better service. Interview three of them before changing course.
  • Warm comments, no action. "What a lovely idea" under a post, with no clicks, is politeness. Treat it as a stop unless the other signals disagree.

Change one variable per rerun and give it two weeks. If you change the price, the audience and the wording at once, a better result won't tell you which change worked.

Four other businesses, four cheap smoke tests

The steps don't change much by trade; the smoke test does. Four illustrations:

  • A bakery testing a cake-decorating evening class. The cheapest test is selling eight tickets for one date at $45 each through the existing newsletter. If it sells out in a week, schedule a second; if four sell, the price or the evening is wrong. AI drafts the class description and a list of questions people will ask about allergies and what to bring.
  • An independent bookshop testing a monthly book subscription. Pre-sell three months to a limited first group of 20 at a stated price, with the first box shipping on a fixed date. Interview five regulars first about how they choose books now; the answers shape whether the box should be "staff picks" or "chosen for you after a questionnaire".
  • A bicycle repair shop testing collection and delivery. Add a tick box to the online booking form, "collect and return my bike, $15", for a month before hiring a van or rearranging staff. The tick rate is the test. An assistant can scan a year of booking notes for customers who mentioned having no way to get the bike in, though the shop should confirm the pattern by reading a sample itself.
  • A café testing a quiet co-working membership. Offer ten "founder" memberships at a fixed monthly price for weekday mornings, sold at the counter. The pre-mortem matters here: the likeliest failure is annoying regular customers who can't get a table, so the test also tracks complaints.

In every case AI does the drafting, sorting and devil's advocacy, and customers do the voting with money or at least with an email address. That split keeps the speed of AI without trusting it on the one question it can't answer.

Testing a service idea: questions owners ask

Can AI-generated customer personas replace real interviews?

No. Simulated customers are useful for brainstorming objections and questions you hadn't thought of, but they can't tell you whether your real customers will pay. They reflect patterns in general text, and they tend to be agreeable. Use them to prepare for interviews and to draft copy, then get the actual evidence from conversations, sign-ups and deposits.

Is it honest to take deposits before a service exists?

Yes, if you're clear about it. Say the service is launching on a stated date, that places are limited, and that deposits are fully refundable if it doesn't go ahead. What isn't honest is implying the service already runs, or keeping money if you cancel. Write the refund promise on the page and keep it.

How many sign-ups count as enough demand?

There's no universal number. Work it back from your break-even: how many paying customers a month make the service worth your time, then set the pass mark at a fraction of that from a small test. If ten customers a month would make it pay, six deposits from a single post and a counter card is a strong early signal.

What if the test is inconclusive?

Change one thing and rerun it for two weeks, rather than launching anyway or giving up. Usually the variable to change is the price, the format or the audience. Keep the rest of the test identical so you can tell what made the difference, and set the new pass mark before you start.

Further reads

Sources: general AI assistant documentation; Rob Fitzpatrick, The Mom Test (interviewing approach); no vendor pricing was needed for this tutorial.

Want help designing the test for your idea?

On a 1:1 call we'll turn your idea into a testable offer, pick the two riskiest assumptions, and set up the smoke test and pass marks so the results actually answer the question.

Book a 1:1 call with me