AI agent: a salesperson that answers customers on its own

What you'll get: the AI agent answers customers in WhatsApp, Telegram, Instagram and the website chat following your instruction and your knowledge base, creates a card in the CRM, hands the conversation to a person at the right moment, and carries out tasks on its own for the customers where you have allowed it.

Time: 60–90 minutes for the first setup (half of it is the knowledge base), then 10 minutes a day on drafts and learned answers.

Who it's for: the workspace owner and administrator — only they can save the agent's settings. Managers only need steps 9–10 and the cheat sheet.

Where: left menu → “AI agents”. In the AI agent's cabinet, on the left, there is a column of sections grouped under “The agent”, “Conversation” and “Service”.

Where the screen says “AI agent”, the cabinet shows your agent's name if its greeting gives one.

Step zero: 5 decisions before you start

Where it answers. Telegram (you need your own bot), WhatsApp (a business number), Instagram and Messenger, the website chat. Start with one or two channels. If your CRM's bot already answers in a channel, decide who stays: two of them in one channel means two replies to the customer.
When it works. Around the clock by default. If not, set the hours and decide what happens at night: stay silent and save the enquiry, or reply with one line.
What it may promise. Prices, lead times, discounts, free slots — only what is written in the knowledge base and the catalogue. Everything forbidden goes into a separate “Boundaries” document (step 3).
When to hand over to a person. Which customer words mean “get me a manager”, who is on duty and at what hours, and how soon the AI agent comes back into the conversation after a manager has replied.
Tone. Formal or informal address, reply length (we recommend “Short (1–2 sentences)”), language: “Auto — the customer's language” or one main language.

Step 1. Open the AI agent's cabinet

Why: all the agent's settings live in one cabinet.

Click path: “AI agents” → the “My agents” block → the “Your agent” card → “Open cabinet →”. A new workspace does not have this card yet (instead you see “Nobody answers customers yet”) — in that case press “Test the agent” in the steps at the top of the page: the cabinet opens, then work down the column on the left.

What you should see: a status badge in the cabinet header: “○ Switched off”, “Being configured” or “● Answering customers”. A new workspace always starts at “○ Switched off” — that is correct: the platform does not answer customers on behalf of a business it does not know yet. We switch it on in step 8.

The AI agent cabinet: the settings sections on the left, the agent name and its state on top.

Step 2. The instruction and the first line

Why: this is where you decide how the AI agent talks: how it opens, what it asks, where it leads.

Click path: on the left, “💬 What it says”. At the top, the “Agent enabled” checkbox (leave it alone for now) and the language choice. Below:

“Niche preset…” → industry → “Apply preset” — fills in a ready-made instruction and greeting; or paste the template from the appendix to this guide;
“Greeting” — the first line;
“Instruction (agent prompt)” — the conversation rules (see the section “How to write the AI agent's instruction” below);
“✨ AI generate” — rewrites the text in the field; always read the result;
“Save”.

What you should see: the notification “Saved ✓”. Under the field there is a character counter: up to 6,000 the AI agent follows the instruction most precisely; anything beyond 32,000 never reaches it. The preset and the generator only fill in the fields — nothing changes without “Save”.

📷 Screenshot: the “💬 What it says” tab with the character counter, /agents/instruction

Step 3. Knowledge base and catalogue

Why: the AI agent takes the address, hours, prices and availability only from here. If a fact is not in the base, it is not in the reply either.

Click path: “📚 What it knows” (opens as a separate page; back is “← Agent cabinet”):

“Business details”: “Website”, “Address”, “Phone”, “Email”, “Working hours”, “Company / legal details” → “Save details”.
“Knowledge base structure” — eight documents: navigator, boundaries, objections, address and contacts, working hours, prices (day and night in one file), services, timetable. For the missing ones — “Create from template”, then “Open”, replace the [brackets] with your own data → “Save”.
“Company material”: “Upload file” (PDF, DOCX, XLSX, TXT, up to 4 MB), “Link to a page on your site” → “Add page”, or text → “Title” → “Add text”.
Optional — “Keyword triggers”: on words like “price|cost” the AI agent sends the right material → “Add trigger”.

Products, services, properties with prices — “🛍 Catalogue” → “+ Add item” or “⬆ Import Excel / CSV”.

What you should see: the line “6 of 6 filled — the agent can answer every basic question.”, the structure documents created, and “Items: N” in the catalogue.

The knowledge base: the «Business details» block and «Knowledge base structure» below it.

Step 4. Reply style and limits

Why: short, consistent replies and no endless conversations that eat the budget.

Click path: “⚙️ How it works”:

the “🧠 Model & replies” sub-tab: “Agent mode” — “Standard” (included in your plan); “Stability” — “Predictable — same answer to the same question”; “Reply style” — “Short (1–2 sentences)”; “Wait time (message debounce)” — 10 sec, so that three messages in a row from a customer get one reply;
the “⏱ Antispam & hours” sub-tab: “Max messages per dialog”, “Token ceiling per conversation” (leave 120000), “Timezone”, “Stop words (customer → pause agent, split by |)”;
optional — “🎙 Voice & languages”: “🎙 Voice replies — the agent sends customers voice notes” and “Agent languages”;
at the bottom, “Save settings”.

What you should see: “Settings saved ✓”. There is one button for all the sub-tabs and it saves them all at once.

📷 Screenshot: the “🧠 Model & replies” sub-tab with “Standard” selected, /agents/settings?sub=llm

Step 5. Check it in the test chat

Why: to hear what the AI agent will say to a customer before it says it.

Click path: “🧪 Test chat” → expand “🧪 Acceptance checklist · 9 scenarios” → “▶” next to each scenario. Your own questions go in the “Type or speak 🎤…” field → “Send”. “Reset” starts the conversation over.

What you should see: each scenario has three automatic checks — “length: N words (≤ 35)”, “ends with a question or step”, “no forbidden words”. Judge the meaning yourself against the “judge: …” line. The test chat works even with the agent switched off; customers do not receive these messages and the CRM does not change.

📷 Screenshot: the expanded checklist with green and red marks, /agents/test

Step 6. Connect the channels

Why: the AI agent answers where customers write.

Click path: “Settings” → the “Integrations” tab:

Telegram: create a bot with @BotFather → the “Telegram” card → “Connect” → paste the bot token → “Connect”.
WhatsApp: “Connect in one click” → in the Meta window pick the business and the number. Then, on the connected card, “Show phone replies here” — otherwise the AI agent will not know that a manager replied from the app.
Instagram Direct and Facebook Messenger: “Connect in one click” → sign in to your Meta account. Reading Direct additionally waits for Meta itself to approve the app.
Website chat: the “AI chat on your website” card → “Get the install code” → paste the line into your site's HTML. WhatsApp and Telegram buttons in the chat window — “Widget settings” → “WhatsApp number”, “Telegram username (without @)” → “Save widget settings”.

Then “AI agents” → “🔌 Where it answers”: for Telegram, Instagram and Messenger choose “Who answers” — “Our agent”, “The bot in your CRM” or “Nobody — channel off”.

What you should see: in “🔌 Where it answers”, in the “Messaging / Channels” group — “N of M connected”. “Check it now” on a connected card changes nothing; it only asks the provider whether the channel is alive.

«Agent channels»: Telegram, WhatsApp, Instagram and Messenger, each with «Who answers».

Step 7. When it stays silent and who it hands over to

Why: not to answer in your name at night, and not to interrupt when a manager is on it.

Click path: “💬 What it says” → below the instruction, the “When the agent answers and when it stays quiet” block, three expandable panels:

“When the agent works”: “Around the clock” or “On a schedule” → “Time zone”, “Days”, “Hours” → “What to do outside working hours”: “Stay silent, save the enquiry” or “Reply with one line”.
“When the agent stays silent”: “Fall silent once a human has replied”; “The agent comes back after” (do not bring it back / 15m / 30m / 1h / 3h / a day); “…or sooner — as soon as the customer replies”; “Words after which the agent falls silent” and “Words that bring the agent back” → “Save the words”. Here too: “WhatsApp template for returning after 24 hours” → “Submit the template for Meta approval”.
“Who to hand over to”: “+ Manager” → name, contact, days, hours → “Save the rota”.

What you should see: the panel headers show “around the clock” or the hours, “comes back by itself · Nm” or “you bring it back by hand”, “N on the rota”. The night line goes out once per night; the enquiry is always saved. These rules apply at once in every channel — this is not a draft.

📷 Screenshot: the expanded “When the agent stays silent” panel, /agents/instruction

Step 8. Switch the AI agent on

Why: until this checkbox, the AI agent answers nobody, however many channels are connected.

Click path: “💬 What it says” → the “Agent enabled” checkbox → “Save”.

What you should see: the cabinet header shows “● Answering customers”, and the “The agent is switched off” badge has gone from “🔌 Where it answers”. Write to a connected channel from someone else's phone: the AI agent replies, a card appears in the CRM, and the conversation appears in “Dialogues”. The “Make active” button in “My agents” changes which agent answers but does not switch it on — only this checkbox does.

📷 Screenshot: the “Agent enabled” checkbox and the “● Answering customers” badge, /agents/instruction

Step 9. Who answers in a specific conversation

Why: to take an important customer yourself and hand them back to the AI agent when you are done.

Click path:

Lead card (CRM → card): next to the channel there is an “AI agent” (or your agent's name) or “Operator” tag. The “•••” menu → “Pause the agent” or “Bring the agent back now” (owner or administrator).
“Dialogues” → a conversation → on the right, “Answering” → “Hand to operator” or “Return to the agent”.
All stopped ones at once: “🔌 Where it answers” → “Paused conversations” → “Resume all”.

What you should see: after “Hand to operator” or a pause from “•••”, the AI agent stays silent in that conversation until it is brought back — a customer message does not bring it back. If a manager simply replied themselves, the AI agent falls silent and comes back according to the rules of step 7.

«Dialogues»: the queue by channel on the left, the conversation with the customer on the right.

Step 10. The AI agent's tasks for a single customer

Why: the AI agent itself does what is in the tasks: writes, calls, reminds.

Click path: CRM → lead card → the “Next step” block → above the task list:

“🤖 AI agent completes the tasks” — when the due time comes, the agent writes the customer one message based on the task's title and note and closes the task; a call task goes into its call queue (from 9 to 21 in the customer's time);
“Applies:” — permanently / 1 day / 3 days / 7 days / 30 days; the period applies to the position of the checkbox (on for a week or off for three days), after which the general setting decides again (step 11);
“➡️ Sets the next task itself” — after a completed task, it sets the next one based on the script and the conversation; one open task at a time, no more than six tasks set by the agent per customer in two weeks;
“⏰ Meeting reminders: a day and an hour before”.

A task is set as usual: “+ Create task” → “What happens next?”, “Due” → “Save task”.

What you should see: a completed task moves to “Done” with a note that the AI agent did it and wrote to the customer (or queued a call). The check does not run the same second — give it 15–20 minutes after the due time.

📷 Screenshot: the “Next step” block with the AI agent's three checkboxes, /crm/leads/…/workspace

Step 11. Autonomy for the whole workspace

Why: so you don't have to tick the checkbox in every card.

Click path: “⚙️ How it works” → the “🔀 Agents per channel” sub-tab → at the bottom, the “Agent autonomy” block: “Fully autonomous” or “Custom”, and the checkboxes “Runs tasks itself for every customer”, “Sets the next task itself”, “Calls on call tasks”, “Reminds about meetings a day and an hour before”, “Books meetings in the calendar”. The “Next-steps script (optional)” field — for example “after the proposal — check the decision in 2 days; silent for 3 days — call”; if it is empty, the AI agent uses the main instruction. Then the “📈 Stage moves” sub-tab: which events move the deal and where to, “Hands off: the agent never moves a deal out of these stages”. At the bottom — “Save settings”.

What you should see: “Settings saved ✓”. The checkbox in a card beats the general setting while its period lasts.

📷 Screenshot: the “Agent autonomy” block with the script field, /agents/settings?sub=channels

Step 12. Follow-ups

Why: to win back those who went quiet.

How it works: the follow-up ladder writes a short message to a customer who went quiet, in their language — one question or an offer of the next step. It stops when the customer replies. At night in the customer's time (22:00–08:00) the step waits for morning. In WhatsApp, 24 hours after the customer's last message, only a Meta-approved re-engagement template goes out. The follow-up will never send a message like “I've sent you the link” or “you're booked in” — it cannot do those things.

Click path — what you do: the re-engagement template (step 7), stop words (steps 4 and 7), “📈 Stage moves” → “No reply to the first follow-up” — where to move the deal. The ladder's steps and switching it on are set up by the Boardroom team: this section is open only in service mode. Write to support saying which touches you need — for example the ready ladder: after 15 minutes, 3 hours and 20 hours, then as a template on day 3 and day 10.

What you should see: follow-up messages appear in the conversation as the AI agent's replies; the line “[re-engagement template …]” means the re-engagement template went out.

Why it works this way — Meta's rules. Within 24 hours of the customer's last message WhatsApp carries plain text — that is free, so the ready ladder makes three touches inside the day: after 15 minutes, 3 hours and 20 hours. After that only an approved template can go out: it is paid, and Meta counts it as marketing. Meta limits marketing to each person and lowers the number's quality rating if you write too often, so after the first day at most 2 templates go out, a week apart (day 3 and day 10), then the ladder stops. If Meta did not deliver marketing to a person (its limit, code 131049), templates to them pause for 7 days; US numbers wait until they write themselves — Meta does not deliver marketing there. Anyone who opted out of marketing (code 131050) hears nothing until they write. Win-back messages and broadcasts go only to people who wrote to you in the last 90 days. Meeting reminders and replies within the day are not limited by these rules. All of this is explained in the «Why it works this way» block under the ladder, with the three numbers (a week, the pause, 90 days) and how many people were skipped in 7 days. Steps more frequent than the rules get a yellow warning, not a block: the system simply does not send the extra template. On the customer's card such a step shows as «skipped: Meta limits marketing to this person».

Plain text to someone who never wrote on WhatsApp. If a customer came from the website or by phone and never wrote to you on WhatsApp, plain text does not go: you see «WhatsApp does not carry plain text: the customer has not written to you on WhatsApp in the last 24 hours. Send a template» and a picker of approved templates under the box.

📷 Screenshot: a conversation with a follow-up message, /conversations

Step 13. Drafts and learned answers — 10 minutes a day

Why: not to lose a reply that did not go out, and to teach the AI agent what you have already answered yourselves.

Click path:

“Dialogues” → the “Drafts” tab: conversations where the AI agent's reply did not reach the customer. In the conversation it starts with “🕶 Draft (not sent):”. Read it and answer the customer yourself.
“📚 What it knows” → at the bottom, “Learned answers — awaiting your decision”. When there are new ones, the notification bell shows the line “Learned answers awaiting your decision”. A manager's answer to a customer's question becomes a candidate: “Approve” — it goes into the knowledge base, and from then on the AI agent answers the same way on its own; “Reject” — it does not.

What you should see: “Approved — in the knowledge base.” The system will not accept an answer containing a phone number or email — remove the personal data and approve again. Nothing gets into the base without your decision.

📷 Screenshot: the “Learned answers” queue with the “Approve” / “Reject” buttons, /agents/knowledge

How to write the AI agent's instruction

One thought and one question per message. No five-question questionnaires.
No longer than two short sentences. Back this up with the “Short (1–2 sentences)” setting; the checklist rejects replies longer than 35 words.
Facts only from the knowledge base and the catalogue. Do not make up prices, free slots or specialists' names. If it is not in the base — “I'll check with a manager” and hand the conversation over.
Every message ends with a question or the next step.
Collect details one at a time: name → phone → convenient time, each in its own message.
Repeat the phone number and the time and ask “is that right?” — before treating the booking as done.
Do not announce what has not been done: “I've sent you the link”, “you're booked in” — only if it really happened.
Do not greet again in a conversation that is already going.
Prohibitions go into the “Boundaries” document of the knowledge base, with the word “forbidden”, and next to each one — what to say instead.
No “we guarantee”, “the best”, “100%” — the checklist catches them.
Keep the instruction under 6,000 characters. The price list, service descriptions and FAQs go into the knowledge base: from there they are pulled in according to the meaning of the question.

Pitfalls

Symptom	Cause	What to do
Customers write, no reply; “🔌 Where it answers” shows the “The agent is switched off” badge	A new workspace is created with the agent switched off. Messages and cards are still saved.	“💬 What it says” → “Agent enabled” → “Save”.
Pressed “Make active” — the AI agent is still silent	The button changes which agent answers but does not switch it on.	Same fix: the “Agent enabled” checkbox.
The AI agent went silent in one conversation	A manager replied themselves, the conversation was handed to an operator, or the customer wrote a stop word.	“Return to the agent” in “Dialogues” or “•••” → “Bring the agent back now”; all at once — “Resume all”.
No replies at night, a batch of follow-ups in the morning	“On a schedule” is on; the follow-up waits 22:00–08:00 in the customer's time.	Working as intended. If you want a night line — “Reply with one line”.
The conversation shows “🕶 Draft (not sent): …”	The channel did not accept the reply: sending is not enabled in it yet, or the provider refused.	Answer the customer yourself; if drafts keep coming — “Integrations” → “Check every connection”.
WhatsApp: after a day neither the follow-up nor the task went out	The 24-hour window is closed and there is no approved re-engagement template.	Step 7: “Submit the template for Meta approval”, enter the name of the approved template.
A manager replied from the phone, but the AI agent keeps going	We cannot see replies from the WhatsApp app without the echo subscription.	“Integrations” → WhatsApp → “Show phone replies here”.
The customer got two replies in Telegram or Instagram	Both the AI agent and your CRM's bot answer in the channel.	“🔌 Where it answers” → “Who answers” → keep one.
The “🤖 AI agent completes the tasks” checkbox is ticked, but the task is stuck	Tasks are picked up only in WhatsApp (including WhatsApp via GREEN-API), Telegram, VK and MAX conversations not stopped by a stop word, with a due date no older than a week; a call only with a phone number and no later than two hours after the due time. Sending tasks and follow-ups is switched on for the platform by the Boardroom team.	Check the channel and the due date. In Instagram and on the website the task stays with the manager.
The checkbox in the card is empty, but the AI agent still writes to the customer	An empty checkbox means “as the workspace decides”, and there “Runs tasks itself for every customer” is on.	To switch it off for one customer, tick the checkbox and untick it — an explicit “no” is saved.
Switched off for 3 days — the AI agent is handling tasks again	After the “Applies:” period, the general setting decides.	For good — “Applies:” → “permanently”.
Turned the checkbox on — the AI agent immediately wrote about an old task	Tasks with a due date within the last 7 days are picked up.	Before turning it on, close or reschedule overdue tasks.
The AI agent handed the customer to a manager mid-conversation	The conversation hit the “Token ceiling per conversation” or “Max messages per dialog”.	Raise the limit in “⏱ Antispam & hours”.
Can't find the “Agent autonomy” block	The block sits in the “🔀 Agents per channel” sub-tab.	Step 11.

Check: 9 scenarios

Acceptance checklist. “🧪 Test chat” → “🧪 Acceptance checklist · 9 scenarios” → “▶” on all of them → the header shows “9/9 passed auto-checks”; the meaning of each reply matches the “judge: …” line.
A fact from the base. In the test chat: “Where are you located and until what time are you open?” → the address and hours word for word from “Business details”. A question about a price that is not in the base or the catalogue → it does not name a price and offers to check with a manager.
Live channel. From someone else's number, “Hello” in Telegram or WhatsApp → a greeting and one question; a new card in the CRM; the conversation in “Dialogues”.
Website. Write in the website chat → the AI agent replies; after the first reply the window asks for a name and phone → the phone appears in the card.
Takeover. Reply to the customer yourself from the card → the AI agent is silent. The customer writes again → the AI agent replies (if “…or sooner — as soon as the customer replies” is on). After “Hand to operator” — silent until you press “Return to the agent”.
Stop word. The customer writes “stop” → the AI agent is silent. The customer writes a return word → it replies again.
Outside working hours. Set a window that is closed right now → the message is visible in “Dialogues”, the reply is one line or silence, and the line is not repeated on the second message. Put the hours back.
Task. In a card with a live WhatsApp or Telegram conversation — “🤖 AI agent completes the tasks”, a task “Clarify the budget” due in 5 minutes → within 20 minutes a message has gone to the customer and the task is in “Done” with the AI agent's completion note. With “➡️ Sets the next task itself” — a new task has appeared, due no sooner than an hour later.
Self-learning. A customer asks a question, a manager replies from the cabinet (a reply of 20+ characters, no phone or email) → the pair appears in “Learned answers” → “Approve” → in a new test chat, the same question → the AI agent replies with the same words.

Cheat sheet: one page

Cabinet: “AI agents” → “Open cabinet →”.
What it says: “💬 What it says” → “Greeting”, “Instruction (agent prompt)” → “Save”.
What it knows: “📚 What it knows” → “Business details” 6 of 6, “Create from template”; prices — “🛍 Catalogue”.
Style: “⚙️ How it works” → “Short (1–2 sentences)” → “Save settings”.
Check: “🧪 Test chat” → “🧪 Acceptance checklist · 9 scenarios”.
Channels: “Integrations” → “Connect” / “Connect in one click” / “Get the install code”.
Hours and handover: “When the agent works”, “When the agent stays silent”, “Who to hand over to”.
Switch on: “Agent enabled” → “Save” → “● Answering customers”.
Take a conversation: “Hand to operator” or “•••” → “Pause the agent”; to give it back — “Return to the agent”.
Customer tasks: “🤖 AI agent completes the tasks” + “Applies:” + “➡️ Sets the next task itself”.
For all customers: “🔀 Agents per channel” → “Agent autonomy”.
Every day: “Dialogues” → “Drafts”; “Learned answers” → “Approve”.

Appendix. A ready-made instruction template

For the “Greeting” field:

Hello! I'm [Name], an assistant at [Company name]. What are you interested in: [buying / renting / a service]?

For the “Instruction (agent prompt)” field — replace all the [brackets]:

You are [Name], a consultant at [Company name] ([city], [niche: real estate — buying, renting, off-plan / services — which ones]). You write to customers in messengers on behalf of the company. You address the customer [formally/informally].

GOAL: understand the customer's request and bring them to [a property viewing / a consultation / a call with a manager]. The main result of the conversation is an agreed time and a confirmed phone number.

HOW TO TALK:
- One thought and one question per message, no more than two short sentences.
- Every message ends with a question or the next step.
- Do not greet again if the conversation is already going.
- Reply in the customer's language.

CONVERSATION ORDER (one point per message):
1. What they are looking for: [type of property or service].
2. For themselves or for someone else, and whether they have been with us before.
3. [Area / budget / timing] — one question at a time.
4. Offer [2–3 suitable properties from the catalogue / the next step] and ask which is closer.
5. Offer [a viewing / a consultation]: ask whether mornings or evenings, weekdays or weekends suit them better, and offer two or three slots from the "Timetable" document.
6. Ask for their name, then their phone number. Repeat the phone number and the time and ask: "Is that right?"

FACTS:
- Prices, availability, addresses, free times — only from the knowledge base and the catalogue.
- If something is not in the base, say: "I'll check with a manager and get back to you" — and do not make anything up.
- Do not write that you have sent a link or booked the customer if that has not happened.

FORBIDDEN:
- Promising [a discount / mortgage approval / returns / timelines], quoting prices that are not in the catalogue.
- The words "we guarantee", "the best", "100%".
- Arguing and pressuring. To "too expensive" and "I'll think about it" — the answer from the "Objections" document.

HAND OVER TO A PERSON if the customer asks for a manager, is unhappy, asks about [the contract / legal questions / deals over N], or you have failed to answer twice. Say: "I'm passing you to [manager's name / our specialist], they will write to you within [time]".