Decision matrix: guided bot, AI assistant or manual quick replies

The discussion about how to handle WhatsApp almost always starts with the tool: which bot, which AI, which platform. It starts wrong. The tool is the consequence of two numbers the business can measure this week: how many messages come in per day and how many different reasons are behind those messages. With those two data points, the decision stops being an opinion and becomes an intersection in a table.

This page does not recommend a brand. It organizes the three possible configurations —manual quick replies, guided-reply bot, and AI assistant— according to the intersection of volume and variability, and explains how to calculate each variable with the data you already have in your WhatsApp inbox. If by the end of this read you can place your case in the matrix, the page has done its job.

The two variables that decide everything

Volume is how many incoming messages you receive per business day, counting all the channels handled by the same person. It is not how many contacts you have or how many conversations: it is messages. One conversation can have eight messages and another just one, and workload is measured in messages because each one requires a read and, almost always, a reply.

Variability is how many different reasons appear in those messages. It is not how many different words customers use: it is how many different intents there are. "How much is it?", "Do you have stock?" and "Do you ship to Córdoba?" are three reasons. Twenty people asking the price with twenty different wordings is one reason with twenty ways of writing it. That distinction is what decides whether a flow bot is enough or whether you need something that interprets free language.

How to calculate daily volume

1. 1Choose a full business week (five days) with no holidays or mass campaigns.
2. 2Count the incoming messages for each day, not the conversations or new contacts.
3. 3Add the five days and divide by five. That is your average daily volume.
4. 4Mark the highest day of the week. That is your peak, and it is the number you will size with, not the average.
5. 5Divide the peak by the number of hours someone is actually watching the inbox, not by the hours the business is open. The difference between those two figures is the message that goes unanswered.

The result of that last step is the metric that matters: messages per hour of actual service. A business with 200 daily messages and a single person watching the inbox for six hours has 33 messages per hour. Another with the same 200 messages but three people covering ten hours has 6.6 per hour. The volume is identical; the load per person is five times different, and so is the recommended setup.

How to calculate reason variability

1. 1Take the last 100 incoming messages and read them straight through.
2. 2Assign each one a short label that describes the intent: price, stock, appointment, shipping, complaint, order status, billing.
3. 3Don't create a new label for every nuance. If two messages ask for the same thing in different words, they go under the same label.
4. 4Count how many different labels you used. That number is your variability over 100 messages.
5. 5Look at what percentage of the 100 messages fell under the three most frequent labels. That percentage is the concentration: the higher it is, the more the same handful of reasons repeats.

Concentration is more useful than the total number of labels. A business with 25 labels where the top three cover 80% of messages has, in practice, three dominant reasons and a long tail. A business with 12 labels where the most frequent one covers 15% has twelve reasons that come up all the time. The first one can handle guided responses; the second one can't.

The matrix: what setup goes in each intersection

Now the two variables are crossed. On the vertical axis, messages per hour of actual service. On the horizontal axis, variability measured as different labels over 100 messages. Each cell says what setup is best and why. The volume rows use the cutoffs of 5 and 20 messages per hour because below 5 one person can manage to answer everything by hand, and above 20 the queue of unattended messages grows faster than any one person can compensate for.

What each configuration means in practice

• Manual quick replies: the team answers everything by hand, with saved text shortcuts for frequently asked questions. There is no automation that decides anything. It is the right configuration when volume is low or when each conversation is worth too much to delegate to a flow.
• Guided reply bot: the customer chooses between numbered options or buttons and receives the corresponding answer. It works when the reasons are few and stable, because every menu option has to be written, tested, and maintained. If a new reason appears, someone has to add it.
• AI assistant: the system reads the message in free language, proposes a reply, and a person approves it before sending. There is no menu to maintain because there are no flows: it learns from the real replies the team has already given. It is the configuration that tolerates high variability without anyone having to program each new case.

The operational difference between the last two is who maintains the system. A guided bot is maintained by writing and updating flows: each new reason is configuration work. An AI assistant is maintained by correcting suggestions: each reply the team approves or edits is information the system incorporates. The first scales with hours of configuration; the second, with hours of normal inbox use.

When the guided bot beats AI

It is not a competition where AI always wins. The guided bot wins in three specific cases. First, when the reason has to be resolved with exact data and does not allow interpretation: a menu that asks for an order number and returns the status cannot make a mistake, and a language model can. Second, when the business needs the customer to follow a fixed path for compliance or record-keeping reasons: there, rigidity is the feature, not the flaw. Third, when volume is high but the reasons are few and very stable, because the menu is built once and is hardly ever touched.

There is also a case where no automation is advisable: when volume is low and each conversation has a high value. A real estate agency that receives fifteen inquiries per day for high-priced properties gains nothing by automating the first reply, because the customer expects to be served by a person and that expectation is part of the service. Putting a bot there worsens conversion, it does not improve it.

Common mistakes when choosing

• Deciding based on a campaign peak. Peaks are handled with temporary reinforcement, not by redesigning the operation for the strangest day of the year.
• Measuring variability by counting distinct words instead of intents. Twenty ways of asking the price are one reason, not twenty.
• Building a guided bot with twelve menu options. Nobody reads twelve options; after five, the customer picks the first one that sounds similar and ends up in the wrong flow.
• Deploying AI without quick support answers. AI suggests, but the team still needs shortcuts for what repeats every day.
• Not defining what happens when no configuration solves the issue. There always has to be a path to a person, and that path has to be written before automating anything.
• Measuring success by the number of automated messages instead of by resolved conversations. A bot that responds a lot and resolves little is a wall, not an improvement.

How to know if the chosen configuration is working

The choice isn't validated the day it's implemented, it's validated four weeks later. There are three signs that the configuration is the right one: the time to first response goes down or stays the same, the number of conversations that end with a person unnecessarily goes down, and the team stops complaining that it answers the same thing all day. If, on the other hand, conversations escalated to a human due to interpretation errors go up, or the team spends more time correcting suggestions than responding, the configuration was too small or too big.

For the case of an AI assistant, the most honest metric is what percentage of suggestions is approved without editing. If that percentage goes up week after week, the system is learning from the team's real answers. If it stagnates, the cause is usually variability: the business may have more reasons than it thought, or the team's answers may not be consistent with each other and the system has no clear pattern to learn from.

Where Wando fits into this matrix

Wando is an Argentine SaaS that uses Meta's official WhatsApp API and works with the AI assistant scheme: it suggests answers and the team approves them with one click, and it learns from the real answers the business has already given instead of requiring programmed flows. That's why it falls into the medium and high variability cells of the matrix, not the guided bot ones. If your measurement places you in low variability and low volume, the matrix says manual quick answers and no tool improves that. The free trial lasts 14 days and doesn't ask for a card; current plans are at wando.online/pricing.