How do I check if clients need a bot before developing it?

Briefly

• The development is often ruined not by the code, but by the lack of demand: in the latest analysis of CB Insights, a weak entry of the product into the market is indicated by 43% of closed companies.
• The idea of a bot can be tested in a few weeks and with almost no budget: interviews with clients, a "painted door" test, a manual pilot.
• Ask customers about what they have already done, not what they will do. A polite "yes, I would" proves nothing.
• A survey of 100 people gives an error of about ± 10 percentage points, a survey of 400 — about ± 5, and then only if the sample is random.
• If the name, phone number or card number of the club is collected in the questionnaire or bot, a separate consent to the processing of personal data is required from September 1, 2025.

Why test the idea of a bot before development?

In order not to pay for a feature that no one will use. This is the most frequent and most expensive mistake in development, and it does not depend on the quality of programmers.

CB Insights analysts in March 2026 dismantled 431 venture capital companies that have closed since 2023. 70% have run out of money, but the authors themselves call this the last reason, not the root of the problem. The root is more often different: 43% of the product did not hit the market well, 29% did not guess the moment (CB Insights, 2026). In an earlier version of the same study, according to 101 closure stories, the first reason was " a solution in search of a problem" — 42% of cases (CB Insights, The Top 20 Reasons Startups Fail).

This is data about startups, but the mechanism is the same as that of a bot for a fitness club or a thermal spa complex. If it is more convenient for guests to call the administrator, no one will open the bot for recording, no matter how well it is made.

The second group of figures is about functions inside ready—made programs. Pendo analyzed 615 subscriptions of its customers. Conclusion: 80% of the functions in the average software product are rarely or never used, and 12% of the functions account for 80% of daily use (Pendo, Feature Adoption Report 2019). The research is old and done on cloud service data, but it shows well how easy it is to write too much.

The third group is about IT projects in general, and here the source is secondary. The CHAOS reports of Standish Group are paid, we have not seen them, and we cite the figures for the review in PM World Journal (January 2026). According to this retelling, in the Standish data for 2020-2024, 31% of projects were recognized as successful (on time, on budget and in full), 50% with deviations, 19% failed. The share of failures has hardly changed since the mid-2010s. The CHAOS methodology is criticized, so consider these figures a benchmark, not an exact measurement.

What does "test the idea" mean and how is it different from MVP?

To test an idea is to get proof that there is a problem and customers are ready to change their behavior for the sake of solving it. Opinions " it would be great" are not considered proof.

A few concepts that will come up next:

• Custdev (customer development) is Steve Blank's method: first find out if there is a problem and whether they are willing to pay for its solution, and only then build the product. Blank defines a startup as an organization that is looking for a repeatable and scalable business model, and advises testing each hypothesis with customer behavior (Steve Blank, 2010).
• MVP (minimum viable product) is the simplest version of the bot, on which you can already test the main hypothesis on real people. For example, a bot that can only record for one service.
• A pilot is a limited launch of an MVP or a ready—made solution: one club from the network, one shift, one month, pre-agreed metrics.
• Fake door ( "painted door" test) is a button or menu item for which there is no function yet. They count how many people clicked on it (Umbrex, Fake Door Test).
• Smoke test is the same for the entire product: a page or message with an offer, and counting how many people have submitted a request.

The order is usually as follows: conversations with customers, then a demand test, then an MVP and a pilot, and only then a full-fledged development. Each step is cheaper than the next and can show that there is no need to go further.

How to ask customers so that they do not respond out of politeness?

Ask about past experiences, not future intentions, and don't talk about your idea. This approach is described in detail in Rob Fitzpatrick's book "The Mom Test" (2013). The name means that the questions should be such that even a mom who wants to support you will not lie (The Mom Test).

Three rules from the book:

1. Talk about the life of the interlocutor, not about your decision. As soon as "we want to make a bot" sounds, the person starts evaluating and praising you.
2. Ask about specific cases in the past. "When was the last time you rescheduled a workout? How did they do it?" is a good question. " Would you use a bot?" is a bad one.
3. Listen more. Your task is to find out how everything works now, not to sell.

What it looks like in a fitness club. The reception manager does not ask if the guest needs a bot. He asks how he signed up for the last group training session, how many times in a month he couldn't get through and what he did when the place was occupied. If it turns out that the guest is registered in the club's application and does not know any problems, this is also a valuable result.

What is considered a demand signal:

• the person himself describes the problem and how he has already tried to solve it (called, wrote to the messenger, asked a friend);
• does he spend time or money on a workaround;
• he agrees to the next step: leave a contact for the pilot, try the prototype, make an advance payment.

Compliments, "interesting idea" and " probably would have used" are not considered a signal. 10-15 such conversations with different types of clients are enough to see recurring problems. This is not statistics, but a search for hypotheses: they are then checked by a survey or test.

If you're interviewing employees, the rule is the same. The administrator of the thermal spa complex will say "yes, the bot would help ". It is more useful to ask him to mark each recurring question of the guests in one shift. In a week it will be seen what actually takes time.

How to check the demand while there is no bot yet?

Offer the service as if it already exists, and count how many people will try to use it. Behavior is more reliable than words, so demand tests are more valuable than questionnaires.

Working options for the service sector:

• The stub button. The "Sign up via messenger" item appears in the club's app or on the website. The person who clicked sees an honest message: " The feature will appear soon, leave a contact — we'll let you know first". Count the percentage of clicks from the number of visitors to the page.
• Mailing with an offer. You send a message to some clients: "Now you can freeze the subscription in the chat ". The answers are processed by a person manually for the first time.
• A manual pilot (a "bot" with a person sitting behind it). The administrator responds in the messenger by script, as a bot would respond. In two or three weeks, you can see how many requests are received, what they are and at what time.

It is important not to deceive the client. The stub should say directly that there is no function yet, and the manual pilot should not impersonate the person for the program if you then ask if the guests liked the "bot".

Check the channel separately. For September 2026, we recommend making MAX and the website the main channel, and Telegram — an additional one. According to Mediascope in the retelling of Habr, in June 2026, the monthly audience of MAX is 86.26 million people, Telegram — 75.69 million, WhatsApp — 63.51 million. Telegram slowdown was officially confirmed by Roskomnadzor on February 10, 2026 (Wikipedia, chronology). In mid-September, users complained about failures, but they did not announce an official full lockdown (Hi-Tech Mail, 09/16/2026). WhatsApp has not been working without a VPN since February 11, 2026 (CNews). In the demand test, give customers a choice of channel — a bot in a messenger (MAX, Telegram) or on a website — and calculate which one they choose more often.

How many clients do I need to survey so that the result can be trusted?

100-200 answers are enough for the first conclusions, 400 or more are better for deciding on large investments. The accuracy depends on the number of answers according to a simple formula.

The error of the fraction with 95 percent confidence is calculated as follows: 1.96 ×(p × (1 − p) / n), where n is the number of answers, and p is the proportion of " yes" answers. The most cautious estimate is obtained at p = 0.5 (Stat Trek).

Answers	Error (95%)	How to read the result "40% are ready to use"
100	±9.8 pp.	in fact, from 30 to 50%
200	±6.9 pp.	from 33 to 47%
400	±4.9 pp.	from 35 to 45%
1000	±3.1 pp.	from 37 to 43%

The error in the table is calculated for the worst case p = 0.5, with a fraction of 40% it is slightly less. To reduce the error by half, you need four times as many answers. Therefore, there is little difference in accuracy between 400 and 1000 answers, and a noticeable difference in cost.

The formula works under several conditions: the sample is random, there are at least 20 times more customers in the database than the respondents, and there are at least ten "yes" and "no" answers (Stat Trek).

How do I find out how much customers are willing to pay for a service through a bot?

Use the Van Westendorp method: four price questions, from the answers to which a range of acceptable prices is constructed. The method was proposed by the Dutch economist Peter van Westendorp at the ESOMAR Congress in 1976 (Van Westendorp's Price Sensitivity Meter).

This is necessary if the bot sells something itself: paid subscription freezing, personal training, additional payment for the VIP zone in the thermal baths. Each respondent is asked four questions:

1. At what price is the service so expensive that you won't buy it?
2. At what price is it so cheap that you doubt the quality?
3. At what price is it already too expensive, but the purchase can still be considered?
4. At what price is this a good purchase?

How to calculate in the table:

1. Write down all the named prices in ascending order.
2. For each price, calculate the proportion of people whose answer is "expensive" is not higher than this price, and separately the same proportion for the answer "too expensive". These are two growing curves.
3. For questions "profitable" and "too cheap", consider the proportion of those whose answer is not lower than this price. These are two falling, that is, inverted, curves.
4. Plot four lines on the same graph and find the intersections.

The intersection of "too cheap" and "too expensive" gives the optimal price. "Expensive" and "profitable" — the price of indifference. "Too cheap" and "expensive" — the lower limit of the acceptable range, "too expensive" and "profitable" — the upper. If you don't want to count manually, there is a ready-made package for the R language (pricesensitivitymeter, CRAN).

Limitation of the method: it shows what people say about the price, not how they buy. He does not give a revenue forecast, so it is better to confirm the result with a pilot with a real payment.

Which survey service should I choose in Russia?

Yandex Forms, a questionnaire or a Testographer are suitable for most tasks. The choice depends on where to get respondents and how many answers you need.

Service	What's Important to Know (September 2026)	When to choose
Yandex Forms	Surveys, tests, feedback forms; there are APIs and integrations, corporate functions in Yandex 360 for business (Yandex help). Tariff limits are not specified in the certificate.	Own customer base, simple questionnaire, need integration
Questionnaire specialist	Free of charge — 3 questionnaires, 10 questions, 50 answers. Paid rates from 2,990 rubles per month. Own panel of 336,000+ respondents, the answer is from 90 rubles (rates of the Questionnaire).	Need answers from people outside your base
Testograph	More than 30 types of questions, API. There are up to 10 responses in demo access. According to the statement of the service, it is included in the register of Russian software and stores data in the Russian Federation (Testograph).	Complex logic of the questionnaire, detailed analytics

The survey can also be conducted in the messenger itself, by a separate bot. This is convenient for guests, but the answers there are linked to the account, so this survey is not anonymous. The next section is about this.

Do I need the client's consent to the processing of personal data during the survey?

If it is possible to identify a person based on the answers, consent is needed, and from September 1, 2025, it will be issued separately from other documents. A completely anonymous questionnaire does not require consent.

Article 9 of the Law on Personal Data now explicitly requires that consent be " formalized separately from other information and (or) documents". This part was changed by the law of 06/24/2025 No. 156-FZ (152-FZ, Article 9, ConsultantPlus; Guarantor). The check mark "I agree" cannot be hidden in the rules of the promotion or in the text of the questionnaire.

For processing data without consent or with an incorrectly executed consent, the company faces a fine of 300 to 700 thousand rubles under part 2 of Article 13.11 of the Administrative Code. Fines have been increased by Law 420-FZ, increased amounts are effective from May 30, 2025 (Administrative Code, Article 13.11).

The questionnaire is not anonymous if it contains a phone number, email, name, club card number, or if the answer is linked to an account in messenger. A practical solution is to split the questionnaire: collect answers without contacts, and those who want to participate in the pilot should be offered to leave contact separately with separate consent.

How to conduct a bot pilot and what to measure in it?

A pilot is a limited launch with metrics that have been agreed upon in advance. Without them, the result will turn into a dispute of opinions.

Procedure of actions:

1. Choose one task: registration for group classes, answers to frequently asked questions from the guests of the term, freezing the subscription.
2. Limit the scale: one club from the network, one service, fixed term.
3. Remove the initial figures before the launch: how many calls per day, how long they take the administrator, how many missed calls.
4. Name the success criterion in advance: what proportion of requests should go to the bot and what part of them should be resolved without a person.
5. Collect feedback from guests and staff. It is especially important in which cases the bot transmitted the conversation to the person.
6. Make a decision: scale up, refine or stop. Stopping after the pilot is a normal outcome, not a failure.

It is convenient to entrust the development of the MVP and the pilot to one team, which will then support the solution. This is how we work with chatbots and AI agents, and the order of stages is described on the "How we work" page.

Checklist: the idea of the bot is verified if…

• 10-15 conversations were held with clients about past experiences, and the problem is repeated for many.
• It is known in which channel customers are ready to communicate with the bot: MAX, Telegram or website.
• A demand test (button, mailing list, or manual pilot) showed real actions, not just words.
• The margin of error is clear for the survey, and the sample does not consist only of the most loyal guests.
• If a bot sells a service, an acceptable price range is known.
• Consent to the processing of personal data is issued separately.
• The term, scale, initial figures and success criteria are set for the pilot.

If some point is not fulfilled, it is better to postpone the development. For fitness club chains, we have a separate page with typical automation tasks.

Sources

1. CB Insights. Why Startups Fail: Top Reasons — 05.03.2026.
2. CB Insights. The Top 20 Reasons Startups Fail — early version of the study, 101 closing stories.
3. Pendo. The 2019 Feature Adoption Report — 2019.
4. G. Arcidiacono. Comparative research on IT project failure rates: a 2025 update. PM World Journal — January 2026 (secondary review of Standish Group CHAOS data; the report itself is paid).
5. Steve Blank. What's A Startup? First Principles — 25.01.2010.
6. Rob Fitzpatrick. The Mom Test is a 2013 book.
7. Umbrex. Fake Door Test — reference material, tested in September 2026.
8. Wikipedia. Telegram blocking in Russia (2026) — verified on 09/17/2026.
9. Hi-Tech Mail. What is happening with Telegram — 09/16/2026.
10. MAX bypassed Telegram by monthly audience (Mediascope data for June 2026) — Habr, 07/22/2026.
11. The authorities finally nailed WhatsApp in Russia — CNews, 02/11/2026.
12. Stat Trek. Confidence Interval: Proportion — reference material.
13. Van Westendorp's Price Sensitivity Meter is the primary source of the method: a report at the ESOMAR Congress, 1976.
14. pricesensitivitymeter package for R, CRAN — package documentation.
15. Yandex Forms Help — verified in September 2026.
16. The questionnaire specialist. Tariffs — verified in September 2026.
17. Testograph — tested in September 2026.
18. Federal Law No. 152-FZ of 27.07.2006, Article 9 (as amended on 26.07.2026) — ConsultantPlus.
19. The guarantor. Consent to the processing of personal data from September 1, 2025 — 2025.
20. Administrative Code of the Russian Federation, Article 13.11 (as amended law of 11/30/2024 No. 420-FZ) — ConsultantPlus.