An AI sales manager can be tested during a seven-day trial without a programmer, provided you don't drag the setup out over the entire week. The first two days are spent on the knowledge base, the third on an exam using your own questions, the fourth and fifth on the channel and a test order in the CRM. The last two days the agent works in the evenings and at night, and you tally the results. Below is a day-by-day plan using an example that is easy to adapt to your own niche. Imagine an online skincare store: sales via Instagram and a website with several hundred items, one female manager until 19:00, and orders living in a CRM.
What to prepare for the AI sales manager test
The trial begins at the moment of registration, so gather everything you can beforehand. A week spent looking for a price list and passwords will say nothing about how the AI sales manager handles your customers. With Marichka AI, a free 7-day period is provided on every tariff plan, and every single one of them should be spent on testing.
The owner of a cosmetics shop won't need much:
- link to the website or current price list in PDF or Excel;
- store rules in a single document: delivery, payment, exchanges, minimum order amount for free delivery;
- 30 real customer questions from Direct over the last month, copied verbatim, complete with errors and abbreviations;
- access to the Instagram account and the CRM with permission to create orders;
- one person who spends half an hour every morning reading dialogues.
The final point is most frequently underestimated. An agent will not learn from its own mistakes if no-one notices them. In a shop with a single manager, this could be the owner herself: half an hour with a coffee before opening.
Days 1–2: knowledge base and no-code setup
The technical part takes less time than it seems. In Marichka AI, the no-code setup is described in an eight-step video guide: registration, choosing a template for the business scenario, uploading documents, connecting Instagram, Telegram or a website, text-based configuration, reviewing the dashboard, placing a test order via a Telegram bot, and viewing the leads page. The service estimates that uploading a catalogue—whether as a price list, a file or a website link—takes less than 10 minutes. Prices, availability and product descriptions are then updated daily.
The remaining two days will go on the rules. The catalogue tells the agent what is being sold, and the rules explain exactly how you sell. They are written in plain text, like instructions for a new manager on her first day at work.
For a cosmetics shop, the rules look something like this. When selecting a product, the agent first asks the client's skin type and what she has already tried, and then suggests no more than two items and explains the difference between them. They quote the ingredients from the product card and add nothing of their own. Prohibitions are listed in a separate paragraph. Any question about breakouts or allergies is not commented on by the agent. The same goes for pregnancy and medications: such a question is immediately forwarded to the manager along with the chat history. The agent does not promise any discount beyond what is on the website.
It is also useful to note down the tone. Whether to address the client formally or informally, whether emojis are allowed, how to sign off. Female clients who are used to your manager will notice the difference on the very first evening.
Day 3: Exam on your own questions
Thirty questions you gathered before the start are now becoming an exam. Open the test chat and paste them one by one, verbatim, with the exact same mistakes. The benchmark for each answer is the same: would your manager have answered that way.
| Question type | Example for a cosmetics shop | What is considered a pass |
|---|---|---|
| Availability and price | “Do you have vitamin C serum? How much is it?” | price and availability match the website as of today |
| Selection | “Dry skin, something lightweight for summer under makeup” | clarifies one or two questions, suggests up to two items from the catalogue |
| Out of stock | “Do you have that cream the influencer has?” | it honestly says that there isn't any, and offers something similar from what is available |
| Prohibited subject | “I broke out after using the cream, what should I do?” | not advising, passing it on to the manager with context |
| Complex order | “Two sets, one a gift, to Lviv by Friday” | clarifies details and brings to completion or hands over to a person |
Analyse every failed response down to the root cause. For the most part, it is visible immediately: the rule is missing from the database, the product card has an outdated description, or the rule is formulated ambiguously. Update it, run the same question through again. By the end of the third day, out of thirty questions, not a single one should remain with a fabricated price or composition. If any such questions remain, it is too early to switch on the customer-facing agent.
Days 4–5: one channel and a test order in the CRM
One channel is enough for the test. Take the one with the most messages after closing. For our cosmetics store, this is Instagram Direct. The limitations that each messenger imposes on bots are described in detail in the review. rules for bots in various messengers.
Here are two technical conditions that stop the test more often than anything else. According to Meta developer documentation, messaging via the Instagram API is only available to professional accounts. If you have a personal profile, switch it to a business or creator account before starting the test. The second condition relates to Telegram. If the service connects a separate store bot, the token for it is issued by @BotFather after the /newbot command. Telegram documentation asks you to treat the token like a password and never share it with anyone, so paste it only into the service cabinet and nowhere else.
Then place a test order on behalf of the female customer: select the product, specify the city and branch, and confirm the amount. Open the CRM and check the card. The contact, product, quantity, amount and comment must be in their respective fields. If something has ended up in the general comment, the female manager will have to transfer the data manually, and the benefit of night orders will decrease.
Days 6–7: Night shift and a fair count
For the past two days, the agent has been responding to clients from 19:00 to 09:00, when the manager is no longer on shift. During the day, everything remains as it was. That way, you will see real questions from female clients without risking the daytime flow.
Every morning, read all the overnight chats. Pay attention to the moment when the female client stops replying: right after the price, or when the agent suggested a different product instead of the one that was asked about. Conversations that break off in the same place indicate a problem with a rule or a product card – this becomes noticeable by the second night. Also reread the chats that the female manager took over in the morning. Check whether a human was genuinely needed there, or if the agent was simply lacking a single line in the knowledge base.
At the end of the week, you need four numbers. First: how many dialogues the agent handled alone. Second: how many of those ended in an order in the CRM. Third: how many conversations they handed over to the female manager, and whether those handovers were appropriate. Fourth, most importantly: how many responses you had to correct.
Make the decision on the AI manager after the trial by the fourth. If the corrections concerned tone or wording, the agent is ready, and you will fine-tune the knowledge base as you go. If it invented a price, inventory level or delivery condition even once and you couldn't find which rule was missing, continue the test in night mode until such responses disappear. The second figure shows money: orders placed at night used to wait until morning. How to convert them into subscription ROI is explained in the article about, how to choose an AI agent and calculate ROI.
When the test should show “no”
It happens that a week honestly shows: now is not the time. This is a normal test result, and it is better to get it on a trial than after three months of payment.
First situation: over seven days during non-working hours, several messages arrive, and the manager calmly replies to all of them in the morning. An AI sales manager will add nothing here, at least until the flow increases. Second: the catalogue exists only in the owner's head, there are no prices or stock levels on the website, and it couldn't be compiled into a spreadsheet in two days. First, the data needs to be put in order. Third: the sale requires a long consultation, such as selecting professional care for a salon with a call and samples. The agent can take the request and gather the initial data, but a human will close the deal.
Що б не показав тест, почніть з одного кроку ще до реєстрації: відкрийте директ і скопіюйте в документ тридцять останніх питань клієнтів дослівно, поруч позначте час кожного. Цей документ стане іспитом на третій день і покаже, скільки роботи припадає на години, коли вам зараз ніхто не відповідає.
Frequently asked questions
Is a programmer needed to test an AI manager?
No, if the service connects channels and catalogue via the dashboard. You provide the website or price list file link, store terms as text, and access to the messenger account yourself. A programmer will only be needed for a custom CRM integration that is not on the list of ready-made ones.
Can the AI manager only be tested at night?
Right, and for the first week that's the safest option. During the day a female manager responds, in the evening and at night an agent, and in the morning you read all the night-time dialogues. That way you see real customer queries and don't risk the daytime flow.
What should I do if an agent has given a customer the wrong answer?
First find the cause in the knowledge base: most of the time, a rule is missing there or an out-of-date price is listed. Add the rule, repeat the same question in the test chat, and write to the client on behalf of the female manager with the corrected response.
What to count after seven days of testing?
Головний показник – скільки відповідей агента вам довелося виправити вручну: якщо це були лише дрібниці тону, тест пройдено, а базу знань можна доточувати вже в роботі. Якщо агент хоча б раз вигадав ціну, склад засобу чи умову доставки і причину знайти не вдалося, тестовий тиждень варто продовжити в нічному режимі. Окремо гляньте, чи з’явилися в CRM замовлення саме в ті години, коли раніше ніхто не відповідав, – це і є гроші, які тест мав показати.