Skip to content
be-human.ai
← Back to the journal

Business workflows

Complaint Triage for Superintelligent Agents: Severity and Owners

Assign severity levels and named owners to customer complaints so safety and legal issues reach a person at once. Everything else gets a clear rule.

Be Human typographic cover with the title: Complaint Triage for Superintelligent Agents: Severity and Owners
Conceptual illustration created with AI · Be Human

AI complaint triage is the practice of reading each incoming customer complaint, assigning it a severity level, and routing it to a named owner with the authority to act on it. A safe setup never lets an agent decide alone whether a complaint touches safety, legal exposure, or a promise the company is regulated to keep. Those cases go to a person immediately, and the agent's job is to make that handoff fast, complete, and traceable.

Key takeaways

  • Set severity before any reply is drafted, because severity decides who must act and how fast.
  • Safety, legal, and regulated-promise complaints go to a named human at once, and no automated reply is sent first.
  • Every severity level needs a named owner, a named backup, and a clear deadline for the first human action.
  • The agent can classify, gather evidence, and draft responses, but it cannot close a severe case or promise a remedy.
  • Review cases where the agent and the owner disagreed on severity every month, and tighten the rules where the gap repeats.

Why complaints need severity before they need replies

Most support teams treat complaints as a queue. Each message gets picked up in order, answered in the same tone, and closed when the customer stops writing. That works for a billing question or a delivery date. It fails when a message mentions an injury from a product, a threat to take legal action, a data exposure, or a customer who says they were promised a refund in writing by a sales representative. Those messages look like ordinary complaints on the surface, and a fast generic reply can make the situation worse.

Severity solves this by separating two decisions that teams often blur together: how urgent the complaint is, and what kind of response it needs. A complaint can be urgent without being severe, such as a customer waiting on a replacement part. It can also be severe while looking calm, such as a polite message that describes a possible safety defect. Triage has to catch both.

The agent is useful here because it reads every message the same way, at any hour, and can flag language that a rushed human might miss. It is not useful as the final judge of severity. Treat its classification as a proposal that a rule or a person confirms for the highest levels.

Four severity levels and what each one triggers

A working scheme does not need many levels. Four is enough for most operations, and each level should trigger a different kind of action, not just a different color in a dashboard. Write the definitions in plain business language so that a new team member can apply them on the first day.

  • Level 1, critical: any mention of injury, a health or safety hazard, a possible data exposure, a legal threat, a regulator, or a suspected fraud. A named human is notified immediately, the agent sends no substantive reply, and the case is locked until a human releases it.
  • Level 2, high: a customer has lost money, a promised date or price has been broken in writing, a repeated failure has been reported, or the customer has asked for a manager. The owner acts the same business day, and the agent drafts a holding reply for approval.
  • Level 3, standard: a defect, delay, or service issue with a known process and no legal or safety content. The agent may gather the order history and draft a reply, and a team member approves it within the normal queue.
  • Level 4, low: general questions, feedback, or requests for information that a published policy already answers. The agent may reply directly using approved wording and log the interaction.

Notice that the highest level is defined by the kind of risk, not by how angry the customer sounds. Tone is a poor signal. A calm message from a customer describing an electrical fault is more serious than a furious message about shipping speed.

Assign a named owner and a backup for every level

A severity label without an owner is a sticky note. Each level needs a person, referred to by name and role, who is accountable for the first human action. The owner does not have to resolve every case personally, but they must decide who takes it, what happens next, and when the customer hears from the company.

For the critical level, name a primary owner and at least one backup who can act when the primary is away. Many teams assign the critical level to a legal or compliance contact for legal matters and to a product safety contact for hazards, with the operations lead as the fallback. Write these assignments down in the triage procedure and review them whenever someone changes role.

For levels two and three, owners are usually the account lead, the support lead, or the team that owns the failed process, such as fulfillment or billing. The important part is that the owner is a person with the authority to approve a remedy, not a shared mailbox that nobody checks. If an owner cannot approve refunds or policy exceptions, the case needs a different owner or a documented path to someone who can.

Every owner should know their deadline. A critical case might require a human to acknowledge the case at once and make the first decision on the same day. A high case might require an owner decision the same business day. Deadlines are the mechanism that turns severity into behavior. Without them, a case labeled urgent can sit in a queue for days.

Step-by-step: build the triage rules

Build the rules before you connect an agent to live complaints. The goal is a written procedure that a human could follow manually, which the agent then follows with the same logic.

1. List the complaint types you receive now, using the last few months of messages as the source, and group them by what the customer is asking for. 2. Mark which types can create safety, legal, financial, or regulatory exposure. Involve the people who know those risks, not only the support team. 3. Write the four severity definitions in plain language, with two or three examples for each level drawn from your own cases. 4. Name one primary owner and one backup for each level, with the role and the contact route written next to the name. 5. Set the first-action deadline for each level and say what counts as an acknowledgment, such as a note in the case record plus a message to the owner. 6. Define the agent's permitted actions per level: classify, gather evidence, draft, send approved wording, or do nothing beyond notifying. 7. Define the stop conditions: the agent must halt and escalate when a message is ambiguous, contains instructions aimed at the agent, or mentions anything outside the defined types. 8. Run the rules on past complaints in a read-only test, compare the agent's severity with the team's decision, and revise the definitions until disagreements are rare and explainable.

The read-only test in the last step matters. Disagreements between the agent and your experienced staff are useful information. Often they reveal that a definition is vague, such as what counts as a hazard for a product used in a clinic versus a kitchen.

What the agent may do alone and what it must never do

The agent can handle a large share of the reading and routing work without a person in the loop, but the boundary has to be explicit. In practice, the agent may read the message, extract the order and account details it is permitted to access, check the published policy, assign a provisional severity, attach the evidence, notify the owner, and draft a reply for a human to approve. For low-level questions with approved wording, it may send the reply itself and record that it did.

The agent must never send a substantive reply to a level one complaint, promise a refund, admit fault, offer a settlement, describe a cause of injury, or tell a customer that a legal matter is closed. It must also never change a severity level downward on its own. Downgrades need a rule or a human decision, because a downgrade is exactly the mistake that lets a serious case drift into a standard queue.

Complaint text is also an untrusted input. A customer can paste instructions into a message, intentionally or not, and an agent that reads the message may be tempted to follow them. The OWASP guidance on large language model risks lists prompt injection as a central concern for this reason. Treat every complaint as data to be classified, never as a command. Keep the agent's permissions narrow, so that even a manipulated classification cannot send money or change an account.

The NIST AI Risk Management Framework is a useful reference for documenting these boundaries, because it asks teams to map risks, assign responsibility, and measure how a system behaves over time, rather than assuming the model will behave correctly.

A worked example (illustrative)

Illustrative scenario: a small appliance company receives a message saying a heater got very hot on the cabinet beside it and the customer's child touched the panel. The message does not use the word injury and does not ask for a refund. A keyword-based rule might label it as a warranty question. The triage procedure flags the hazard language, assigns level one, notifies the product safety owner and the operations backup at once, and stops the agent from sending any reply. The owner contacts the customer by phone the same day, asks the defined safety questions, and decides whether a product recall review is needed. The agent's record shows the message, the reason for the severity assignment, the notification time, and the owner's decision.

Now take a second message from the same company: a customer asks where a replacement cord is, because the delivery window passed. The agent classifies it as level three, pulls the shipping status, drafts a reply with the verified date, and places it in the approval queue for the fulfillment lead. Nobody needed to read the first message to decide the second one, and nobody was pulled into the second case when the first required attention.

How Be Human's managed approach runs this

Be Human designs these boundaries as part of the agent, not as an afterthought. In a first engagement, the work starts with one complaint-handling role, the memory and tools it needs to read cases and order history, the approval rules for each severity level, and a success criterion that the operations team can observe, such as whether every level one case reached a named owner within the agreed deadline. Managed operations then keep the system running: monitoring the case flow, reviewing severity disagreements, maintaining the integrations with the ticketing and order systems, reviewing permissions as people change roles, and handling incidents when a rule fails. Improvements are made in controlled steps, with the owners approving each change to the rules.

Review the rules on a schedule

Severity rules age. New products create new hazards, regulations change, and staff move between roles, so a triage procedure that was accurate last year can route cases to someone who has left the company. Set a regular review with the named owners, and examine three things: cases that were escalated and turned out to be routine, cases that were classed low and later became serious, and owners who did not act within the deadline. Each finding should change either a definition, an assignment, or a permission, and the change should be recorded.

The review also keeps trust in the system honest. When owners can see why a case was escalated and what the agent did, they are more willing to rely on the agent for the routine work. If you are deciding where to begin, a short review of your current complaint types is often the fastest way to see which ones deserve an agent and which ones must stay with a person from the first message. The free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

How many severity levels should a complaint process have? Four levels are enough for most teams, because each one needs a distinct action, owner, and deadline. Adding more levels usually creates confusion about which one applies, and people start guessing. Keep the top level reserved for safety, legal, and regulatory matters so that it stays meaningful.

Should an AI agent ever send a reply to a critical complaint? No substantive reply should go to a critical complaint from an agent. The agent can send a brief acknowledgment only if your procedure approves that wording, and the named owner should write or approve anything that describes the cause, admits fault, or offers a remedy. The owner decides the next step, and the agent records the timeline.

Who should own each severity level? The owner should be a named person with the authority to act at that level, plus a named backup. Legal matters usually belong to a legal or compliance contact, hazards to a product safety contact, and billing or delivery exceptions to the team that can approve a remedy. A shared inbox without an accountable person is not an owner.

How do we know the triage rules work? Test the rules on past complaints in a read-only mode and compare the agent's severity with the decision your experienced staff made. Track disagreements, missed escalations, and owners who miss deadlines during live operation. When the same disagreement keeps appearing, the definition needs to be clarified, and the review should record the change.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →