Decision matrix: manual messages, templates, or AI automation
When to reply manually, when to use a template, and when to let AI suggest
The most common mistake when setting up WhatsApp Business is not choosing the wrong tool, but applying the same logic to all conversations. A customer asking about business hours does not deserve the same treatment as one filing a complaint about an order, and a promo sent once a month is not managed the same as a daily appointment reminder.
This matrix organizes decisions along two axes: the type of interaction (inquiry, notification, complaint, sale, follow-up) and its frequency (sporadic, recurring daily, mass). This answers the underlying question: should this be handled by a person, sent via a template, or managed by an AI system that suggests and a human approves?
The three modes of responding
Before the matrix, it is worth defining the three modes, because the decision is not binary between "manual" and "automatic." There is a third path that combines both.
- Manual message: a person writes the response on the spot, without system help. It works for the unpredictable: a complaint with context, a negotiation, an unusual inquiry.
- Template (HSM): a message pre-approved by Meta that is sent outside the 24-hour window. It works for one-way notifications: appointment confirmation, payment reminder, promo. Template approval can take up to 48 hours.
- AI-assisted with human approval: the system suggests a response based on the business's real history and the person approves it with a click. It works for repetitive but varied cases: inquiries about hours, price, shipping, availability.
The key difference between template and AI-assisted: the template is a fixed message sent the same to everyone; AI-assisted adapts the response to each conversation. The template does not expect a reply; AI-assisted does, because the cycle is ask-respond-approve.
The matrix: which mode to use based on interaction and frequency
The table is read like this: find the row that describes the type of interaction and the column that describes its frequency. The intersection indicates the recommended mode. It is not an absolute rule: it is the starting point to avoid thinking about each case from scratch.
| Type of interaction | Occasional (1-5 per week) | Repetitive daily (5-50 per day) | Massive (campaign or notice to many) |
|---|---|---|---|
| Simple query: hours, address, product price | Manual. There are few and they don't justify setting anything up. | AI-assisted. The answer is always similar but with variations (which product, which price). | Not applicable. You don't send a campaign to answer queries. |
| Complex query: shipping to another province, warranty, invoicing | Manual. It requires context and judgment. | Manual with AI support. The AI suggests a draft but the case needs human review. | Not applicable. |
| One-way notice: confirmed appointment, payment received, order status | Template. There are few, but it's still better for them to go out automatically and leave a record. | Template. This is the ideal case: it goes out automatically every day at the same time. | Template. A campaign is scheduled and sent to the segment. |
| Reminder: appointment tomorrow, invoice due, pick up order | Template. It goes out automatically and doesn't require a response. | Template. Can be automated by schedule. | Template. A campaign is scheduled. |
| Claim or complaint | Manual. Always. A complaint poorly answered by a bot is the most expensive mistake. | Manual with AI preparing context. The AI can summarize the history and suggest a tone, but the decision is human. | Not applicable. Complaints are not sent in bulk. |
| Consultative sale: quote, price negotiation, conditions | Manual. Each sale is different. | Manual with AI preparing drafts. The AI speeds things up but a person closes the deal. | Not applicable. |
| Simple sale: product with fixed price and known stock | Manual. If there are few, it's not worth setting up. | AI-assisted. It answers price, stock, and payment methods, and escalates to a human if the question gets complicated. | Template for a one-off promo. The rest, AI-assisted. |
| Post-sale follow-up: delivered order, satisfaction survey | Manual. There are few and they require tact. | AI-assisted with backup template. The AI suggests, the person approves. | Template. Sent to everyone who purchased in the period. |
How to apply the matrix in practice
The matrix solves 80% of cases. The remaining 20% are situations where the recommended mode is not enough. These are the most common and how to get out of them.
- The business receives 3 queries per day but all are the same type: "Do you ship?". Even if there are few, if the answer is always the same, it's better to use AI-assisted or a saved quick reply. Frequency is not the only criterion: repetitiveness also counts.
- An approved template becomes invalid because the hours or price changed. Templates need to be reviewed, not set once and forgotten.
- A customer replies to a reminder template with a question. At that point the template no longer serves: the conversation moves to manual or AI-assisted, depending on volume.
- The AI-assisted mode suggests an inappropriate response. That's why the mode is "assisted with human approval": the system proposes, the person decides. If the suggestions are bad, you need to review the history the system is learning from.
Frequent mistakes when choosing a mode
These are the four mistakes most commonly seen in businesses that are just setting up WhatsApp Business. Recognizing them saves time and avoids angry customers.
- Automating everything with templates: templates are one-way. If the customer replies, someone has to handle it. A template that generates 50 replies that nobody reads is a ticking time bomb.
- Not using templates for what they are: a reminder for an appointment that is sent manually every day is work that could be saved. The template doesn't replace the person: it takes the repetitive work off their plate.
- Confusing AI-assisted with autonomous bot: AI that responds on its own without supervision is not what is recommended by default. The safe mode is for AI to suggest and a human to approve. That avoids 90% of serious errors.
- Choosing the mode based on the tool rather than the interaction: if the business already has a CRM, there's a tendency to send everything via template. If it already has AI, there's a tendency to automate everything. The matrix works the other way: first the type of interaction, then the tool.
How to size the team according to the chosen mode
The chosen mode defines how many people are needed to handle the workload. The formula is simple: number of messages per day, divided by the number of messages one person can respond to per hour, multiplied by the hours of service.
| Mode | Messages per hour per person | When it is justified |
|---|---|---|
| Pure manual | 20-30 | Less than 50 messages per day, with complex conversations that require context. |
| Manual with assisted AI | 40-60 | Between 50 and 200 messages per day. AI prepares the draft and the person only approves or corrects. |
| Templates for notifications | No person required | One-way notifications do not consume human time. They free up capacity for what does require it. |
| Assisted AI with multi-agent | 60-100 | More than 200 messages per day. Several people approve suggestions in parallel. |
Questions to decide before configuring
Before touching any configuration, answer these five questions. If you don't have the answer, don't configure anything yet.
- How many messages does the business receive per day, on average, in the busiest week?
- How many of those messages are repetitive inquiries (same topic, same answer)?
- How many are notifications that the business initiates (reminders, confirmations, promos)?
- How many people handle the service today and how many hours per day?
- What happens when a customer responds to a template? Who handles it and in how much time?
With those answers, the matrix applies itself: repetitive and one-way goes to templates; repetitive but with variations goes to assisted AI; unpredictable and sensitive stays manual. The result is a system where the person does what only a person can do, and the machine takes care of the rest.
Wando (wando.online) uses the official WhatsApp API from Meta and its AI suggests responses that the team approves with one click, learning from the business's real responses. Current plans can be consulted at wando.online/pricing.
The hidden cost of choosing the wrong mode
Choosing the mode is not just about efficiency: it has a direct and indirect cost. If you use a template for a complaint, the customer feels ignored and the complaint escalates. If you manually respond to something that was repetitive, you pay for hours of someone who could be closing sales. If you let AI act unsupervised, the risk of a public error is real. The cost of a bad response is not the time it took to write it, but the customer who doesn't come back. That's why the matrix is not an optimization suggestion: it's a way to avoid losses that don't show up in the monthly report.
| Mode | Direct cost | Indirect cost |
|---|---|---|
| Manual | Staff hours | Slow response during peak hours |
| Template | Meta message per conversation | Customer responds and no one attends |
| AI-assisted with approval | Tool plan | Suggestion that needs correction, but with supervision |
The edge case: when the recommended mode is not enough
There are situations that don't fit into the matrix because they mix types of interaction. The most common case: a customer who starts with a simple query (AI-assisted mode), moves to price negotiation (manual mode), and ends with a complaint about delay (manual with context). The mode is not chosen once per conversation, but per segment. The practical rule: if the conversation becomes complex, the system should be able to escalate to a human without the customer noticing the change. In Wando, that's solved with human approval: AI suggests, the person decides, and if the case gets complicated, the agent takes control.
| Situation | Initial mode | Signal to change | Mode to change to |
|---|---|---|---|
| Customer asks price and then asks for discount | AI-assisted | Question about special conditions or volume | Manual |
| Customer responds to a reminder template with a complaint | Template | The message is not a yes or no, it's a complaint | Manual with AI summarizing history |
| Customer makes a simple query but in an aggressive tone | AI-assisted | Offensive language or threat of cancellation | Manual, with priority |
| Customer asks for info not in the history | AI-assisted | AI suggests something generic or doubtful | Manual, and check if info needs to be loaded |
Tough questions nobody answers
Guides explain how to use each mode, but not what happens when things get complicated. These are the questions that often go unanswered and how to approach them.
- What happens if the AI suggests a response that is wrong but no one reviews it? In Wando, the default mode is with human approval: the suggestion is not sent on its own. But if the team is overloaded and approves without reading, the error is human, not systemic. The solution is to lower the volume or add staff, not blame the AI.
- How many templates should you have approved? Templates are approved once and reused. But if the price or schedule changes, they must be updated. The recommendation is to have a monthly review calendar, not wait for a customer to flag the error.
- What happens with messages outside business hours? Wando has auto-responses by schedule. But note: the auto-response is not a template or an AI response. It is a notice saying "we'll get back to you during such hours." It does not replace customer service, only delays it.
- Can Instagram and WhatsApp be mixed in the same inbox? Yes, Wando integrates Instagram and Messenger. But the matrix applies the same: a query via Instagram is not different from one via WhatsApp. The mode is chosen by content, not by channel.
What to do when a template is not approved
Meta reviews templates and may reject them. The approval period is up to 48 hours, but it is not guaranteed. If a template is rejected, the most common mistake is trying the same thing with different words. What you need to check is the reason: Meta usually indicates whether it is due to excessive promotional content, lack of opt-in, or ambiguity. Reminder or confirmation templates are usually approved more easily than promotional ones. If you need to send a promo, it is best to have a previously approved template and not improvise on the date.
| Rejection reason | What it means | How to resolve it |
|---|---|---|
| Excessive promotional content | The template looks like spam or direct sales | Reduce to useful information, without exaggerating benefits |
| Lack of opt-in | The recipient did not consent to receive messages | Verify that the contact has accepted communication |
| Ambiguity | The purpose of the message is not clear | Rewrite with a clear objective: confirm, remind, inform |
| Invalid format | Syntax errors or poorly used variables | Review Meta's documentation on template format |
The role of webhooks in the matrix
The matrix decides the response mode, but not how it connects with the rest of the tools. That's where webhooks come in. Wando has integrations with Make, Zapier, n8n, and Slack. A typical case: a customer pays via Mercado Pago and the webhook triggers a confirmation template. Another: a purchase in the e-commerce creates a contact in Google Sheets and is tagged for a campaign. Webhooks do not change the matrix, but they allow the chosen mode to execute without friction. If the matrix says "template for payment received," the webhook ensures the template goes out on its own when appropriate.
| Trigger | Automatic action | Response mode |
|---|---|---|
| Payment confirmed in Mercado Pago | Send payment confirmation template | Template |
| New purchase in e-commerce | Create contact in Google Sheets and tag | Not applicable (it is registration, not a response) |
| Customer message outside business hours | Send schedule auto-response | Schedule auto-response |
| Event in Slack (e.g., stock alert) | Notify the team | Not applicable (it is internal) |
Frequently asked questions
What's the difference between a WhatsApp template and an automatic reply?+
A template (HSM) is a message pre-approved by Meta that is sent outside the 24-hour window, ideal for one-way notifications like reminders or confirmations. An automatic reply is a message sent instantly when someone writes, within the 24-hour window. The template doesn't expect a response; the automatic reply does, because it starts a conversation.
When should I use assisted AI instead of replying manually?+
When there are more than 50 messages per day and most are repetitive queries with variations: hours, prices, shipping, availability. The AI suggests a response based on the business's real history and a person approves it with one click. If the volume is lower, replying manually is simpler and doesn't justify setting anything up.
What happens if a customer replies to a template?+
The template is one-way: it doesn't expect a response. If the customer replies, the conversation enters the 24-hour window and someone has to handle it. That's why it's a good idea to define who responds to those replies and how quickly, before sending a template campaign.
How long does it take for a WhatsApp template to be approved?+
Meta's approval of a template can take up to 48 hours. That's why it's best to have templates approved in advance and review them periodically: if an hour or price changes, you need to update them and wait for approval again.
Does assisted AI replace a human agent?+
No. The recommended mode is for AI to suggest responses and a human to approve them with one click. That combines the speed of automation with human judgment for cases that require it. AI learns from the business's real responses and improves over time, but the final decision is always made by a person.
What happens if I choose the wrong mode?+
The cost isn't just time: using a template for a complaint angers the customer, and replying manually to repetitive messages wastes hours. The matrix helps avoid those indirect costs, which don't show up in the monthly report.
Does assisted AI replace a person?+
No. The AI suggests and the person approves with one click. That human-in-the-loop cycle is what prevents serious errors. AI speeds things up, but the final decision is always made by a person.
What do I do if the AI suggests an incorrect response?+
In Wando, the AI suggests and a person approves with one click. If the suggestion is bad, it isn't sent: the person corrects or discards it. The system learns from the business's real responses, so suggestions improve over time. If you see repeated errors, review the history the system is using to learn.
How long does it take for a template to be approved in Meta?+
The allowed time is up to 48 hours, but it can be less. It's not guaranteed. If the template is rejected, check the reason Meta indicates (promotional content, lack of opt-in, ambiguity) and fix it. Don't try the same thing with different words.
Can I use templates to answer queries?+
No. Templates are one-way messages: they're sent outside the 24-hour window and don't expect a response. If a customer replies to a template, the conversation moves to manual or assisted AI depending on volume. For repetitive queries, assisted AI with human approval is recommended.
Do webhooks replace the need to choose a mode?+
No. Webhooks automate the trigger of actions (for example, sending a template when a payment is confirmed), but the decision of which mode to use for each type of interaction is still made by the business according to the matrix. The webhook executes, it doesn't decide.
What happens if a customer replies to a reminder template with a question?+
The template no longer applies to that conversation. The customer's message enters the 24-hour window and must be answered manually or with assisted AI, depending on complexity. If it's a simple question (hours, address), assisted AI can suggest a response; if it's a complaint, it goes to manual.