Decision matrix: when an agent should resolve alone, when to consult and when to escalate

What this matrix is for

This page does not discuss whether automating is a good idea. It assumes there are already people (or a person) handling conversations and that the real problem is something else: two different operators facing the same message make different decisions. One answers something that commits the business; another checks on something they already knew how to handle. Both are costly: the first because of the error it creates, the second because of the time it takes from the person who should be resolving.

The matrix classifies each inquiry into one of three levels of intervention: N1, resolve on your own; N2, check before answering; N3, escalate to another person or area. The classification does not depend on who the operator is or how much experience they have: it depends on the type of inquiry, the risk of the answer, and whether the decision can be undone. That is what makes it usable as a shared reference.

The most common mistake when building a rule like this is defining it by topic. "The salesperson handles prices, everyone else handles the rest" sounds orderly but breaks down as soon as a case comes up that crosses two topics. Here the level is assigned based on three axes that are combined in a table, and the topic is only one of the inputs.

The three axes that define the level

Before looking at the table, it is worth understanding what is being measured. Each inquiry is evaluated against three axes, and the final level is the highest one produced by any of the three. They are not averaged or offset: if one axis gives N3, the inquiry is N3 even if the other two give N1.

• Risk: what happens if the answer is wrong. A misstated schedule detail is corrected in the next message; a misstated price commitment, delivery deadline, or health condition creates a problem that cannot be fixed with a message.
• Reversibility: whether the decision can be undone at no cost. Offering an appointment and rescheduling it is reversible. Confirming coverage, applying a discount, or canceling a service is not.
• Scope: whether the answer commits only that conversation or sets a precedent. An exception granted once will be requested again, and the operator who granted it without checking left the problem for the next person.

The risk axis is the most underestimated one. The useful question is not "is it serious?" but "what would have to happen for this to become a problem?". If the answer to that question includes a person making a claim, a refund, or an explanation to a third party, the risk is high even if the inquiry seems trivial.

Matrix by type of inquiry and risk

How to read the table in the moment

In practice, no one is going to read fourteen rows while they have a client waiting. The table is studied once and then applied with a short rule that sums it up: if the answer compromises money, deadlines, health or something that cannot be undone, it is not N1. Everything else resolves itself.

That rule covers most cases. For the rest, the operator needs to be able to locate the inquiry in the table quickly. That is why it is best to have it at hand in the same place where you provide support, not in a separate document that no one opens. If the team uses a shared inbox, the table can live as a pinned note; if it uses another tool, as a pinned message in the internal channel.

• If you are unsure between two levels, take the higher one. The doubt is information: it means the case is not covered by the rule and it is best to resolve it with someone else.
• If the inquiry is N2 and the person you need to consult is not available, don't answer anyway. Let the client know you are verifying and come back when you have the answer.
• If the inquiry is N3, the operator is no longer the point of contact. Escalating is not passing the message: it is passing the context and letting the client know that another person is continuing.

The handoff: what has to travel with the conversation

A poorly done escalation costs almost the same as not having escalated at all. If the person receiving the case has to reread the whole conversation to understand what is happening, the time the operator saved was paid double by the person who received it. The handoff has to carry the minimum context so that the other person can respond without asking the customer the same thing again.

Edge cases that get classified wrong

There are inquiries that systematically land in the wrong level, not out of carelessness but because their form does not match their risk. These are the ones that come up again and again when decisions are audited.

• The frequent customer asking for an exception. It is classified as N1 because "it is always granted," but granting it without checking sets a precedent. It is N2 even if the expected outcome is obvious.
• The inquiry that looks informational but asks for confirmation. "Can you confirm it arrives on Thursday?" is not a schedule question: it is a deadline commitment. N2.
• The complaint expressed as a question. "Why was I charged this?" is not looking for an explanation, it is looking for a correction. It is answered as N2 even if the tone is neutral.
• The operator's own mistake that they can fix alone. If the operator made a mistake and can correct it at no cost, it is N1 with a later notice, not N3. Escalating everything you did wrong saturates the receivers.
• The inquiry that was already escalated and comes back. If the customer insists on a case that is already in someone else's hands, it is not re-escalated: the person holding it is notified and the customer is told it is still in progress.
• The request from another area of the business. When the inquiry is internal, this matrix does not apply: it is resolved through internal channels and is not mixed with customer service.

How to know if the matrix is working

A decision matrix that is never reviewed becomes outdated within weeks: new products appear, conditions change, new people join. The way to know if it still works is not to ask the team whether they think it's fine, but to look at three specific signals.

The review doesn't need a meeting. It's enough for someone to look at the week's escalated conversations and mark which ones could have stayed at a lower level and which ones should never have left N1. With two or three adjustments per month, the matrix stays up to date.

Common mistakes when building this matrix

These are the ones that almost always show up in the first version of an escalation rule. Recognizing them before writing it saves you the correction afterward.

• Defining the level by person and not by query. "What Juan resolves, Ana doesn't" is not a matrix, it's a permanent exception that breaks when Juan isn't there.
• Defining the level by channel. The same request is equally risky via WhatsApp, Instagram or phone. The channel changes the format of the response, not the level.
• Not writing down the level for frequent queries. If the ten most common queries don't have an assigned level, each operator decides it again every time.
• Escalating without saying what is expected. An escalation without a specific request is a message the other person has to interpret, and that takes longer than having resolved it.
• Never lowering a level. If everything that was ever escalated is still N3, the matrix only grows and ends up being ignored.
• Confusing level with urgency. An N1 query can be extremely urgent and an N3 can wait until the next day. They are different axes and are logged separately.

Template for writing your business's matrix

The table on this page is a starting point, not a ready-made configuration. Every business has a handful of query types that are its own and that do not appear here. The way to complete it is to list the real queries from the last two weeks and assign each one the three axes.

1. 1Gather the last 50 or 100 closed conversations. If there are many more, take a sample by day of the week so atypical days are not left out.
2. 2Group by type of request, not by customer. "Price," "appointment," "complaint about delay," "billing" are types. Ten or fifteen types cover almost everything.
3. 3For each type, answer the three questions: what happens if the answer is wrong, whether it can be undone, and whether it sets a precedent.
4. 4Assign the highest level produced by any of the three axes. Do not average.
5. 5Write the level next to the query type and share it with everyone who handles customers. If someone disagrees with a level, it is discussed once and stays in writing.
6. 6After two weeks, review the escalations and adjust. The first version always has two or three levels set wrong.

The result is a short table, one page long, that can be kept on hand. It does not replace the operator's judgment: it frames it. An operator with judgment and no matrix makes good decisions most of the time and decisions different from those of their colleagues almost always. The matrix does not seek to make judgment uniform; it seeks to ensure that risky decisions do not depend on who was on shift.