How AI reply suggestions work

When a message arrives, Wando generates a suggested reply. The suggestion is not sent on its own: you review it, edit it if needed, and approve it. That way you stay in control of what is said to the customer.

What context the AI uses

• Your business's system prompt (tone, rules, industry).
• The templates you have uploaded.
• Previous corrections you made by hand (few-shot learning).
• Knowledge files and business documents.

How to improve it over time

Every time you correct a suggestion, that example can be used as a reference for future replies. The more you use Wando and make corrections, the more the suggestions adapt to your style.

What happens when the AI makes a mistake

Nothing is sent without your approval. That is the point of human-in-the-loop: if the suggestion is wrong, you discard it, edit it, or write the reply from scratch. The mistake never reaches the customer.

What matters is what you do afterward. If you correct the suggestion before sending it, that correction becomes an example and feeds into the learning. If instead you delete everything and write outside of it, the AI has nothing to learn from and the same mistake can happen again.

For the system to improve, it's better to edit on top of the suggestion rather than ignore it. It's the difference between correcting and throwing away.

• If the suggestion is wrong, edit on top of it: that way the correction is recorded.
• If the problem repeats with the same type of query, context is usually missing from the system prompt or the knowledge files.
• If the tone doesn't match, review the tone rules first before touching each response by hand.

What Wando's AI does NOT do

It doesn't respond on its own without supervision by default. Wando suggests and the team approves. If you're looking for a bot that answers without anyone watching, this isn't that.

It's not a flow-based chatbot. There are no decision trees to program or maintain when the business changes. The AI works from the real context of the business and the corrections you made.

Nor does it replace Meta in any way: template approval, API limits and conversation prices are defined by Meta. Wando uses the official WhatsApp API (WhatsApp Cloud API) and templates are approved within up to 48 hours. Any other timeframe or cost depends on Meta and can be checked in its official documentation.

Example: a query from a pizzeria, step by step

An audio comes in at 21:40: "Do you deliver to Villa Urquiza and until what time?".

Wando transcribes the audio and generates a suggestion. The business's system prompt says that delivery to Villa Urquiza costs $X and that the kitchen closes at 23:30, so the suggestion starts with those two pieces of information.

The owner reads it: the schedule is right, but he wants to clarify that the last order is taken 20 minutes before closing. He edits that part and approves. The reply goes out through WhatsApp Business.

That edit stays as an example. The next time someone asks about delivery hours, the suggestion will probably already include the cutoff of 20 minutes before. Nothing had to be programmed: it was corrected once and the system picked it up.

Common mistakes when using AI suggestions

The most frequent one is leaving the system prompt empty or generic. If you don't tell it the industry, tone and rules, the AI improvises and the suggestions come out flat. Fill it in once and improve it every now and then.

The second one is approving without reading. The suggestion is a draft, not an automatic send. If you approve everything blindly, the customer receives replies that nobody reviewed and the learning gets contaminated with examples you didn't want.

The third one is correcting from scratch. If you delete the suggestion and write everything again, you lose the signal. Edit on top of what's already there.

• Load the system prompt with industry, tone and rules before demanding precision from the AI.
• Review before approving: it's a draft, not a send.
• Edit on top of the suggestion instead of rewriting from scratch.
• Add knowledge files with the information that comes up most often (shipping, hours, payment methods).
• If the same mistake comes back, don't patch it response by response: fix the context.

How long it takes for the improvement to show

There's no guaranteed timeframe. It depends on how many corrections you make and how repetitive the business's inquiries are. A shop with ten frequently asked questions will notice the adjustment sooner than one with very varied inquiries.

What you can measure is your own edit ratio: if you touch the suggestion less and less before approving, the context is working. If you're still correcting the same thing every week, the problem is in the system prompt or the knowledge files, not the AI.

Message and campaign statistics give you the volume; the rest you see in the day-to-day of the inbox.

How it combines with the rest of the inbox

The inbox with AI suggestions coexists with the other tools: templates (HSM) and campaigns for messages that go out to many contacts, contacts with tags for segmenting, multi-agent so several people can assist at once, and auto-replies by schedule for what comes in outside business hours.

Suggestions are for one-to-one conversation. Templates and campaigns, for mass messaging. Don't mix them: a template approved by Meta has a different format and different rules than a one-off reply.

If you need to connect the inbox with other tools, Wando has webhooks to Make, Zapier, n8n and Slack, Google Sheets for contacts and Mercado Pago for payments. The channels are WhatsApp (Meta's official API), Instagram and Messenger.

How the AI is trained without you programming anything

This is the question that comes up most often when someone hears "AI that learns": who trains it, how long does it take, do you need to know about prompts? The short answer: you train it without realizing it, every time you correct a suggestion before approving it.

The mechanism is few-shot learning. Instead of retraining a model (which is expensive, slow and not under your control), Wando passes the AI real examples of how your business responded. Those examples come from your corrections and from the responses you approved as is. The AI reads them as reference before suggesting the next response.

That's why editing the suggestion instead of discarding it matters so much: an edit is a labeled example ("this was wrong, this is right"), while a discard leaves no signal. It's not magic or a model that retrains itself: it's context that accumulates.

• You don't have to write technical prompts or program anything: the context is built from what you already do in the inbox.
• Every correction adds an example; every discard does not.
• The learning lives in your account, not in a global model: what your business learns is not mixed with another's.

What data you should load before expecting good suggestions

The AI doesn't guess. If the system prompt is empty, it will suggest generic responses that don't work for your industry. Before demanding precision, give it material.

The bare minimum you need: industry, tone (formal, informal, with or without emojis), what you do and what you don't do, and the hard rules of the business. For example: "we don't take reservations via WhatsApp", "shipping is only within CABA", "we don't give prices over chat, we redirect to the website". These rules prevent the AI from promising things you can't deliver.

Then, the knowledge files: hours, shipping zones, payment methods, return policies, frequently asked questions. Everything you answer a thousand times a week goes there. The more specific the material, the fewer edits you'll have to make later.

• Industry and tone: without this, suggestions come out flat and out of character.
• Hard rules: what the AI must never promise or state.
• Knowledge files: hours, shipping, payments, policies, FAQ.
• Loaded templates: they serve as a reference for format and approved content.

How to measure whether suggestions are improving

There's no magic number or dashboard that tells you "you improved 40%". What you can look at is your own edit rate: out of every ten suggestions you approve, how many did you touch before sending? If that proportion drops week after week, the context is working.

The second indicator is repetition. If you correct the same type of error every week (same wrong hours, same tone, same missing info), it's not the AI that's failing: there's missing context in the system prompt or in the knowledge files. That error can't be fixed response by response.

Message and campaign statistics give you conversation volume, but they don't measure suggestion quality. For that, the reference is your eye on the inbox and the count of edits you make.

• Lower proportion of edits → the context is working.
• The same error repeats → missing context, not a model problem.
• Many suggestions discarded without editing → review the system prompt: it's probably too generic.

Full example: an e-commerce store with a stock inquiry

A text message comes in at 15:20: "Hi, do you still have the black t-shirt in size M? How much is it with shipping to Rosario?".

Wando generates a suggestion. The business's system prompt states that prices are checked on the website and that shipping to Rosario has a fixed cost. The knowledge files contain the list of available sizes. The suggestion starts with the stock and redirects the price to the website, as the rule requires.

The owner reads it: the stock is correct, but he wants to add that if the purchase is made today, it goes out in the same day's dispatch. He edits that part and approves it. The reply is sent via WhatsApp Business.

That edit becomes an example. The next stock inquiry with shipping will probably include the mention of same-day dispatch. Nothing was programmed: it was corrected once and the system took it as a reference.

• The system prompt prevented the AI from making up a price.
• The knowledge file resolved the stock without the owner having to type it.
• The edit added a valuable piece of information (dispatch today) that the AI can now reuse.

How the AI is prevented from making up information

The real risk of any generative AI is that it fills in what it doesn't know. At Wando, the defense is twofold: on one hand, the human-in-the-loop, which prevents a made-up reply from reaching the customer without you seeing it. On the other, the loaded context, which gives it real material so it doesn't have to improvise.

The rule of thumb: if information isn't in the system prompt, in the templates, or in the knowledge files, the AI doesn't have it. It can suggest something plausible, but not verified. That's why it's best to explicitly load what you do want it to say and, if needed, an instruction like "if you don't know the price, ask them to check the website."

When the data can vary (Meta's conversation prices, approval timelines, API limits), the correct reference is always Meta's official documentation. Wando doesn't define those rules and can't assert them for you.

• If the data isn't loaded, the AI doesn't have it: it doesn't make it up well.
• Instructions like "if you don't know, hand off" prevent made-up answers.
• Meta's data (prices, timelines, limits) is looked up in its official documentation, not assumed.

When it's best for the AI not to suggest anything

Not every conversation benefits from a suggestion. There are cases where an automatic reply takes away instead of adding, and it's best for the team to write from scratch.

Delicate complaints, legal matters, emotionally charged complaints, or one-off negotiations: there, a "correct" suggestion can sound cold or out of place. Better for a person with context about the relationship to write it.

It's also best to turn off suggestions when the business is changing something (new prices, a current promotion, a change of hours) and the context isn't updated yet. On those days, the AI may be suggesting with old information. Update the system prompt or the files before trusting the suggestions.

• Complaints and sensitive matters: human reply, no suggestion.
• One-off negotiations: the AI doesn't know the history of the relationship.
• Recent changes (prices, hours, promos): update the context first or turn off the suggestion until it's up to date.