Support Manual

In the sidebar the module is called Support. Inside you will find four tabs:

The Setup guide card marks each step automatically. If a step that was Ready goes back to pending, the setup broke there.

1. Create a channel. Go to Channels → New Channel. Type an easy-to-recognise name and role, such as "Main Support" and "Support assistant".

2. Connect WhatsApp. Scan the QR from the company phone: WhatsApp → Linked devices → Link a device. Do not use a teammate personal number.

3. Assign an AI model. The model decides how well it understands the customer and how many credits it consumes. Also confirm whether it supports images or audio.

4. Write the instructions. Define identity, scope, hours, rules, allowed data and when it must hand off to a person.

The Dashboard is an activity view, not an inbox. Use Refresh to query the data again.

The Messages sent by date, Calls by date and Credits spent charts help compare activity and cost over time.

Open Channels → New Channel. A channel must have, at minimum, a name, role, model and instructions before serving customers.

4.1Basic fields

4.2Connection type and mode

For WhatsApp (QR) you can choose:

• Scan new number: creates a session and requests a QR.
• Use existing channel: shares an already-connected line. You must pick the Existing number and complete the Whitelist; that agent will answer only those numbers.

Type numbers with country code and no +. Example: 593991234567.

1. Click the channel row to open it. In the general tab, look for Scan to connect WhatsApp.

2. On the company phone open WhatsApp → Linked devices → Link a device.

3. Scan the code. If it shows Waiting for QR code… for too long, use Regenerate QR.

4. Wait for the Connected state and verify the right phone shows up.

Channel changes save when you leave a field or change a selection. Confirm ✓ Saved. If Could not save appears, retry before closing.

6.1Choosing the model

A cheap model spends fewer credits, but may follow complex instructions worse. A more capable model usually understands long conversations, images and ambiguous requests better. The list shows cr/M, that is, credits per million tokens.

• If the model shows no image capability, do not expect it to interpret photos or receipts.
• If it does not understand audio natively, configure Audio transcription model.
• If you leave transcription empty, a voice note the model cannot hear will not be understood correctly.

6.2Writing the system prompt

The Agent role is just a label. The System prompt is the real control: it defines how it talks, which data it may use and what it must never invent.

You are the support assistant of [ISP NAME].

GOAL
- Help with service, balance, payment and plan questions.

ANSWER STYLE
- Answer in clear, friendly, brief English.
- Ask one question at a time when information is missing.
- Never invent prices, coverage, dates or technical fixes.

RULES
- Human support hours: [HOURS].
- If the customer asks for a person, use the human handoff.
- If you cannot solve it in two tries, hand off to a human.
- If the case was solved but someone should review it, use the human follow-up and continue the conversation.
- Never promise a visit or refund without confirmation.

ISP DATA
- Plans and terms: [SUMMARY OR POINTER TO THE FUNCTION].
- Payment methods: [DATA].
- Emergency contact: [NUMBER].

In the functions subtab you decide which actions the AI may run. Enabling a function does not replace the instructions: the prompt must say when to use it.

Indicators may show Enabled, Disabled or Always On.

7.1Reports

Every handoff and every follow-up is logged on the platform, in Log › Reports. The tab shows a counter of unread reports and, clicking a row, the case detail:

7.2Notifications

• Human Notification Number: receives over WhatsApp a copy of every handoff and follow-up.
• Payment Registration Notification Number: receives the notice when a payment is logged from the conversation.

Use full country code and number, no +.

8.1Number Filter

Numbers are entered without +. A wrong country code makes the filter miss.

8.2Delays

• Message Delay (ms): briefly waits and groups consecutive messages before asking the AI. The usual starting value is 2000 ms.
• Send Delay (sec, min–max): adds a random pause between outgoing messages. It also paces broadcasts and helps avoid overly fast sending.

8.3Auto-Takeover

When enabled, if a person answers from the connected WhatsApp account, PlenoAgent automatically pauses the AI for that contact during the Auto-Pause Duration (min).

In WhatsApp (QR) channels, the voice subtab uses a second linked device, separate from the chat device.

1. In voice, use Pair voice device and scan the QR as an additional linked device.

2. Wait for Voice device paired.

3. Select a Voice model. Without a model, calls are not answered.

4. Pick a Voice and use Test this voice with a short sentence. The test uses the real model and spends at least one credit.

5. Write the Voice call prompt and enable Answer calls with AI.

The voice prompt must be brief and conversational. Do not use tables, markdown or long lists: the content will be read aloud. If left empty, the general instructions are reused.

Max call duration (sec) caps cost and avoids calls left open forever. The allowed range is 30 to 3600 seconds; empty uses the server value.

Use Unlink only if you are going to pair the voice device again.

The Log › Chats subtab allows searching by phone and shows Contact, Channel, Last activity and State.

Use the row control to pause or resume the AI. When pausing manually, make sure a person will continue the conversation over WhatsApp.

Click any row to open the detail. On the left you will see the conversation. AI answers and function runs can be opened to review:

• Raw request and Raw response of the model.
• Arguments and Result of a function.

The Log › Calls subtab groups by contact: channel, last call, duration, total calls, credits spent and state.

• Pause AI on calls stops the AI from taking that contact calls.
• Resume AI on calls enables it again.
• Refresh queries the log again.

Call pausing is independent from chat pausing. If a contact needs fully human service, check both places.

The tab gathers three blocks: the two Billing automations (Payment Thanks and Payment Reminder) and the manual Broadcasts. The first two are configured once here and switched on or off per customer from each subscription toggles in Billing.

12.1Payment thanks and reminder

Each card asks for the same: the sending channel, the sending moment (when the payment is approved or a few hours later; on the due date or a few days before), the allowed sending window with its timezone, and the message template with variables and preview.

If you enable attaching the PDF (receipt or statement), it goes out with the sender data loaded in Billing → Settings → Operation. If that data is not there yet, the message still goes out, just without the PDF.

12.2Broadcasts

In the Broadcasts block you can send the same message to several phones from a connected channel.

1. Pick a channel in Connected state. If none shows up, go back to Channels.

2. In Phone numbers, type one number per line, with country code and no +. Numbers separated by comma or semicolon are also accepted.

3. If Billing has active customers, use Import all customers and review the list before continuing.

4. Type the Message and check it contains no personal information belonging to a single customer.

5. Click Send Broadcast. The platform confirms how many messages were queued.

13.1Before starting the day

• Confirm the needed channels are Connected.
• Check there are credits and the module is working.
• Open Log › Chats and look for paused conversations needing follow-up.
• Check Log › Reports: the counter shows the cases the AI escalated and nobody has read.
• If you use voice, confirm the device is still paired.

13.2After changing instructions

• Wait for ✓ Saved.
• Test from another phone with a new chat, or delete only a trial conversation to avoid old context.
• Verify a normal answer, a refusal, a handoff and each enabled function.
• Check credits spent and the technical inspector if the answer was not as expected.

13.3When a person steps in

1. Confirm the chat shows Paused (auto) or pause it manually.

2. Answer from the same connected WhatsApp account.

3. When done, resume manually or let the auto-pause expire, depending on the case.

The channel is connected, but does not answer.

Confirm it is active, has an AI Model, there are credits and the module is working. Also check Number Filter and the chat state.

It answers some numbers and not others.

Check Whitelist, Blacklist and country codes. If the line is shared with another agent, confirm which list holds each contact.

The AI talks over a person.

Enable Auto-Takeover and raise the Auto-Pause Duration (min). Verify the person answers from the same linked account.

The AI stays silent after an intervention.

Find the contact in Chats. If it is Paused (manual), use resume; if automatic, wait the remaining time or resume.

It does not understand voice notes.

The main model must support audio, or the channel must have an Audio transcription model. Also check credits.

It does not interpret a photo or receipt.

Select a model with image capability. For payments, enable Payment Registration and confirm Billing is configured.

It does not hand off to a person.

Enable Human Handoff and add to the prompt when to use it. Test an explicit request: "I want to talk to a person". If only Human Follow-up is on, the AI reports the case but never pauses the chat: that is the expected behaviour.

The AI reports a case but nobody finds out.

Reports live in Log › Reports even with no number configured. If you expected the WhatsApp notice, check Human Notification Number and the delivery state inside the report.

It does not answer calls.

Confirm Voice device paired, Answer calls with AI, a Voice model, credits and global voice enablement. Check whether the contact is paused in Calls.

No channels show up for a broadcast.

Broadcasts need a connected WhatsApp QR channel. Connect or reconnect one from Channels.

The broadcast takes long.

It is intentional: it goes out in the background and respects the configured delay. Active conversations take priority. Do not lower the delay aggressively.

An answer is wrong or invented.

Open the row in Chats and check Raw request, Raw response and functions. Fix the prompt, add the missing information or pick a more capable model.