Decision matrix: automatic, AI-suggested or manual replies by inquiry type and risk

The problem isn't "AI yes or no," it's what kind of inquiry you're facing

Almost every discussion about automated customer service starts off wrong: people argue about whether the business "uses AI" or "handles things manually," as if it were a single decision for the whole channel. In practice, three modes coexist in the same inbox and the real work is assigning each conversation to one of them. A hair salon appointment and a complaint about a product that arrived broken come in through the same number and can't get the same treatment.

This page doesn't recommend a tool: it builds the criteria. Two variables decide, and with those two you fill in the table. The first is the type of inquiry: whether the answer is always the same or changes depending on the case. The second is the risk: what happens if the answer comes out wrong, incomplete, or out of context. When both are low, automating is free. When either goes up, you need a person at some point in the chain, even if they're not writing the first line.

The third option, the middle one, is the most ignored: AI suggests and a person approves with one click. It's not a chatbot that answers on its own or an agent typing from scratch. It's a draft with context that someone on the team reviews. For many small businesses that's the only AI mode that makes sense at first, because it doesn't require trusting that the system gets it right without supervision.

The three modalities, defined without ambiguity

An automatic reply isn't just the welcome greeting. It includes auto-replies by schedule (outside business hours, holidays), receipt confirmations, and any text triggered by a known condition. Its limitation is that it only works for the foreseeable: if the inquiry resembles the condition but isn't identical, the automatic reply answers anything with total confidence.

AI suggestions are different because the text is generated for that specific conversation, using what was already said before. In Wando, for example, the AI learns from the real answers the team gave before and proposes the draft; the team approves it with one click. That's the human-in-the-loop mode: the system doesn't answer on its own by default. The advantage isn't the speed of sending but the speed of writing: the agent stops building the answer from a blank page.

The matrix: query type crossed with risk

Read the table like this: find the query in the row that best describes it and then move down the risk column. If you're unsure between two rows, choose the lower one (the more specific one), and if you're unsure between two risk levels, choose the higher one. Doubt almost always means there's missing context to automate.

How the risk of a query is measured

Risk is not the seriousness of the topic in the abstract: it's how expensive it is for the answer to be wrong. It's scored with four questions, and just one positive answer is enough to move up a level. You don't need a complicated system; just noting the level in the conversation tag is enough to later measure how many went to manual.

• Does the answer involve money? A wrong price, a discount that doesn't exist or a poorly put-together quote is paid for with the business's margin or with an angry customer.
• Does the answer involve sensitive data? Documents, health data, minors' data or payment information shouldn't go through generated text without review.
• Does the answer promise something the business doesn't control? A delivery time, an appointment date or an external approval. If the business doesn't guarantee it, it's not automated.
• Is the answer hard to reverse? A "yes, I'll change it for you" sent by mistake is binding; an "I'll look into it and confirm" is not. The less reversible, the higher the risk.
• Has the customer already shown frustration? If it's the second or third message about the same topic, automation reads as disinterest and escalates the conflict.

With those five questions, the level works out like this: low risk is when none were positive, medium when one was positive and it can be corrected in the same thread, high when more than one was positive or when the query involves money, sensitive data or a commitment from the business.

The escalation circuit: how a conversation moves from one mode to another

The matrix only works if the mode change happens at a specific point. If there is no escalation rule, the conversation stays on automatic until the customer gets angry, and that's when it moves to manual with all the context lost. These are the four signals that should take a conversation out of automatic or suggested mode and put it in a person's hands.

An operational detail that gets overlooked: when the conversation escalates, the person who takes it needs to see the full history, including the automatic reply that was already sent. If the agent starts without that, they will ask things the customer already answered and the escalation feels like starting over.

The edge cases that break the matrix

There are inquiries that fall right on the edge and aren't resolved by the table. These are the four cases that always come up and how they are handled.

• Outside business hours with high risk: the temptation is to leave it on automatic because no one is there. The right move is not to answer the inquiry but to acknowledge receipt with a fixed text that promises nothing and clarifies when there will be a response. The inquiry itself stays for manual.
• Known customer asking something simple: even if the inquiry is low risk, if the customer is recurring and the history is long, the AI suggestion is usually better than the fixed text, because it can take into account what was already discussed. Automating here feels cold.
• Ambiguous inquiry: if it's not clear what they're asking for, none of the three modes works well. It goes to manual and the first reply is a question, not a solution.
• High volume with a small team: when there are no people to approve suggestions, the discussion isn't AI or manual but which inquiries can be taken out of the flow with an honest automatic reply. If there are none, the problem is staffing, not the tool.

What to measure to know if assignment is working well

The matrix can be audited. You don't need a complicated dashboard: it's enough to tag each conversation with the mode that was used and review three numbers per week. If the manual percentage rises without volume rising, the matrix is poorly calibrated or the team is avoiding the suggestion. If the automatic percentage rises and complaints appear about out-of-place responses, the risk is being underestimated.

One caution with the data: if you're going to analyze what time messages arrive to decide your auto-responses by time slot, use the message's timestamp (the one that comes from the channel), not the date it was saved in your system. If you ever imported history, the save date piles everything up on the day of the import and any conclusion about time slots comes out wrong.

Common mistakes when building this matrix

• Defining the mode by channel and not by inquiry: "automatic on WhatsApp, manual on Instagram" doesn't survive the first complaint that comes in through WhatsApp.
• Automating the greeting and nothing else: that's not automating customer service, it's greeting faster. The work is still entirely on the other side.
• Using the AI suggestion as an automatic response: if no one approves it, it stopped being a suggestion. The mode is defined by who sends it, not by who wrote it.
• Not keeping a record of which mode was used: without that there's no way to audit the matrix or to know if the criterion works.
• Changing the matrix every time a strange case appears: strange cases go to manual and that's it; the matrix is reviewed by period, not by incident.
• Promising resolution times to the customer from the automatic mode: the time depends on the case and on who takes it, so it's best to acknowledge receipt without committing to an hour.

The proof that the matrix is well built is simple: anyone on the team can take a new conversation, place it in a row and know whether the system answers it, whether they approve it, or whether they write it from scratch. If deciding requires asking the owner, the matrix isn't written yet.