Decision matrix: when to respond with a person, AI, or an automatic message
Decision matrix: person, AI, or automatic message?
If you handle customer service on WhatsApp, at some point you'll ask yourself: "Should a person answer this, should I let a bot handle it, or should I send an automatic message?". The answer is not "always the same". It depends on three variables: the type of inquiry, the risk of making a mistake, and the customer's context. This matrix organizes your options so you don't improvise.
Here we're not talking about a specific tool, but a method you can apply with any system. The key is that AI doesn't replace the person: AI suggests, the person approves. That's called human-in-the-loop, and it's what prevents a machine error from ruining a customer's day.
The three options and when to use each one
| Option | What it is | When it's suitable | When it's NOT suitable |
|---|---|---|---|
| Person | A human writes the response from start to finish. | Complex inquiries, complaints, claims, negotiations, legal or billing issues. | Repetitive inquiries that already have a standard response: time is wasted and responses are slow. |
| AI with supervision (human-in-the-loop) | AI suggests a response and the agent approves or edits it with a click. | Medium volume of inquiries that repeat but aren't identical: hours, prices, availability, shipping. | When there's no one to review the suggestions or when the risk of error is extremely high. |
| Automatic message | Fixed message sent without human intervention: welcome, hours, receipt confirmation. | First contact outside business hours, delay notice, response to a question whose answer is always the same. | When the customer expects a personalized response or when the message could confuse or frustrate. |
The three criteria for deciding
Before looking at the full matrix, define the criteria. There are three, and they're answered with yes or no:
- Is it a repetitive inquiry? If you've already answered it more than five times this week, it's a candidate for automation.
- Is there risk if it gets it wrong? If an error causes loss of money, legal trouble, or damage to the relationship, a person has to be involved, no question.
- Does the customer need to feel heard? Complaints, claims, or emotional issues require human empathy, not a template.
Complete decision matrix
| Type of inquiry | Repetitive? | Risk if it fails? | Recommended option | Example |
|---|---|---|---|---|
| Schedule or location inquiries | Yes | Low | Automatic message | What time do you open on Mondays? |
| Price of a product or service | Yes | Medium (if the price changes) | Supervised AI | How much does the Pro plan cost? |
| Order status | Yes | Medium | Supervised AI | Where is my order? |
| Claim for a damaged product | No | High | Person | It arrived broken, I want it replaced. |
| Price or terms negotiation | No | High | Person | If I pay in advance, do you give me a discount? |
| Legal or billing inquiry | No | High | Person | I need the invoice with a different tax ID. |
| Initial greeting outside business hours | Yes | Low | Automatic message | Hi, are you open? |
| Open question about a service | Sometimes | Medium | Supervised AI | How does the warranty work? |
| Complaint about response delay | No | High | Person | I've been waiting for two days and no one responds. |
How to apply the matrix in practice
- 1Classify your most frequent inquiries: write down the last 50 you received and mark which ones repeat.
- 2Apply the three criteria: repetitive, risk, need for empathy.
- 3Define standard responses for low-risk repetitive ones: that goes to an automatic message.
- 4Set up AI to suggest responses for the rest, but have a human approve them before sending.
- 5Review the matrix once a week: the types of inquiries change and what was repetitive may no longer be.
Frequent mistakes when deciding
- Automating everything: the customer notices the lack of empathy and the brand loses trust.
- Putting a person on everything: it's slow, expensive, and the team burns out on inquiries that could resolve themselves.
- Using AI without supervision: if the AI makes a mistake on a claim, the problem gets worse. Human supervision is not optional, it's the standard.
- Not updating automatic messages: an old schedule or outdated price generates more inquiries than it resolves.
- Ignoring context: the same question may require a person one day and an automatic message another, depending on the customer's tone.
Example application in a real business
Imagine a food place that receives 100 messages per day. 60% are questions about hours, address, and deliveries: they can be resolved with an automatic message. 30% are inquiries about the menu or order status: the AI suggests a response and the manager approves it with one click. 10% are complaints or special requests: a person handles them. With that split, the team dedicates its time to what matters and the customer doesn't wait for the simple stuff.
These percentages are an illustrative example, not verified statistics. Each business has its own distribution.
Conclusion
The matrix is not a magic recipe: it is a starting point for you to decide with good judgment. What matters is that AI does not replace the person, but rather helps them. If a query is repetitive and low-risk, automate it. If it involves risk or needs empathy, have a person respond. And in between, AI with supervision: it suggests, but the human decides. This way you protect the customer relationship and you don't go crazy dealing with the same thing every day.
What to do when the matrix is not enough: edge cases
The matrix covers most cases, but there are situations that escape simple classification. For example, a query that is repetitive but the customer raises with an aggressive or confused tone. There, the recommended option would be AI with supervision, but the risk that the suggestion doesn't capture the customer's anger is high. In such cases, the practical rule is: if the message has more than three exclamation marks, contains insults, or words in all caps, it goes straight to a person even if it is repetitive. AI can suggest, but the human has to be attentive to those signals.
Another edge case is the query that seems simple but hides a second intention. A classic example: "How much does it cost?" can be a price query (repetitive, low risk) or the start of a negotiation (non-repetitive, high risk). The difference lies in the context: if the customer has asked before, if they mention the competition, or if they ask for a discount, the matrix leans toward a person. The recommendation is never to automate a query that has a negotiation component, even if it starts as a price question.
| Situation | What it seems | What it really is | What to do |
|---|---|---|---|
| Customer asks price and adds 'can you give me a discount?' | Repetitive price query | Negotiation | Person, always |
| Message with aggressive tone or all caps | Repetitive query | Emotional complaint or claim | Person, even if repetitive |
| Customer asks something you already answered before | Repetitive | Frustration over lack of response | Person, with an apology included |
| Question that seems simple but is from a new customer | Repetitive | First impression of the brand | AI with supervision, checking that the response is clear and friendly |
Difficult questions that nobody answers
There are uncomfortable questions that the matrix does not resolve. The first: what happens if the AI suggests an incorrect answer and the agent does not notice? The honest answer is that it can happen, and that is why supervision is not just approving with a click: it is reading the suggestion, verifying the data, and only then sending. If the agent approves without reading, the system fails. The recommendation is that teams have a mental checklist: is the data correct? Is the tone appropriate? Does it answer what was asked? That reduces errors.
Another difficult question: when to migrate an AI query to a person? The rule is that the customer must always be able to ask to speak to a human. If the supervised AI does not resolve in two attempts, or if the customer explicitly asks, the conversation goes to a person. That is not a system failure: it is a sign that the query is more complex than it seemed. The matrix must include that escape rule, or the customer will leave frustrated.
The third difficult question is about the cost of being wrong. The matrix says "high risk" or "low risk," but it does not quantify. The recommendation is that each business defines its own threshold: if an error costs more than X pesos or more than X customers, it is high risk. That threshold changes by industry. A real estate agency has high risk in almost everything; a food place has high risk only in claims. Defining that threshold in writing helps the team not improvise.
How to adjust the matrix when the business changes
The matrix is not static. If the business launches a new product, if hours change, or if a strong competitor enters, queries change and so does the classification. The recommendation is to review the matrix once a month, not once a year. In that review, look at the latest queries you received and ask yourself: which ones are no longer repetitive? Which ones have become riskier? Did any new type of query appear that was not in the matrix? That monthly adjustment prevents the system from becoming outdated.
The matrix also needs to be adjusted when the team changes. If a new agent joins, supervised AI can be more useful because the agent does not know the answers from memory. If the team shrinks, the automatic message can cover more cases than it did before. The matrix must reflect the team's real capacity, not the ideal one. If there is no one to supervise, supervised AI is useless; better a clear automatic message and the rest handled by a person when possible.
Frequently asked questions
What does human-in-the-loop mean?+
It's a model where AI suggests a response and a person approves or edits it before sending. This combines machine speed with human judgment, avoiding mistakes that could damage the customer relationship.
Is an automatic message suitable for a complaint?+
No. A complaint requires the customer to feel heard. An automatic message can increase frustration. The right thing is to escalate to a person who can resolve the issue with empathy.
How do I know if an inquiry is repetitive?+
If you've already answered it more than five times in the same week, it's a candidate for automation. You can keep a simple log in a spreadsheet or use the statistics from your support tool.
What happens if the AI makes a mistake?+
If the AI is supervised, the error is caught before sending. That's why human approval must be part of the process, not an optional extra. Without supervision, the risk of an error costing a customer is high.
What do I do if a customer gets angry because AI responded instead of a person?+
If the customer explicitly asks or seems frustrated, move the conversation to a person immediately. AI with supervision is not meant to replace humans, but to speed them up. Having a clear escape rule prevents the customer from leaving.
How do I know if an inquiry is high or low risk?+
Define your own threshold: if a mistake costs you money, a customer, or a legal issue, it's high risk. That threshold varies by industry. Write it down and share it with the team so everyone decides the same way.
How often should I review the decision matrix?+
Once a month is fine. Look at recent inquiries, check if new types have appeared, and adjust the classification. The matrix becomes outdated quickly if the business changes.