Alphaweb

AI agent · Customer service · Sort & route

Complaint classification and routing

Complaints arrive through every channel and wait to be read before anyone can act. The agent classifies each one against your definition, routes it to the team that owns it, and keeps the unclear and sensitive ones for a person.

all channelsclassifyrouted to the owning teamfollow-up, linked to the open casevulnerability or legal threat, to a lead
1,800 complaints / month · 82% handled by the agent · 18% to a person
Typical volumes for this process, not a client figure.
Today

Complaints land in a shared inbox; someone reads each one, picks a category and forwards it.

With the agent

Classified, routed with the clock started; sensitive and unclear ones go to a person.

Where the time goes today

Complaints reach you through every door at once: the complaints mailbox, the general contact form, chat transcripts, call notes typed by the contact centre, scanned letters, and the occasional message forwarded by a director. A triage team reads each one, decides whether it is a complaint or an ordinary query, finds the customer and the order, picks a category and forwards it to the team that owns the problem. Where your rules set a response deadline, it runs from arrival, not from opening.

Most of the time goes on context, not on the decision. The triager searches the customer record by email address, then by name, then by the order number buried in the third paragraph. Long threads must be read to the end to learn whether this is a new complaint or the fourth chaser. A message that raises two issues, say a late delivery and a rude driver, is forwarded to one team and the second issue disappears.

The costly errors are quiet ones. A complaint logged as a query never gets a formal response and never reaches the root-cause figures. A misrouted case bounces between billing and logistics while its deadline runs. A customer who mentions a bereavement or financial hardship is handled like everyone else, because the triager was reading for the category, not for the person.

How the agent works

  1. Collect every channelThe agent watches the mailbox, web forms, chat exports, call notes and the scanning queue. Each message is logged once, with its channel and its time of receipt, which is the time any deadline runs from.
  2. Find the customerIt matches the sender to an account and an order using email address, phone number, references quoted in the text and names. If that customer has an open case on the same issue, the message is linked to it instead of opening a new one.
  3. Apply your definitionIt tests the message against your written definition of a complaint and records which part of the definition was met, quoting the sentence that meets it. Messages clearly outside the definition go to the service queue as queries.
  4. Classify and flagIt assigns category, product and issue from your taxonomy, splits a message raising several issues into linked parts, and looks for the markers your policy names: vulnerability, legal action, contact with a regulator or the press, safety, financial loss.
  5. Route and timeIt assigns the owning team from your routing table, sets the response deadline from your rules, and sends an acknowledgement from an approved template where you allow it.
  6. Hand over the restMessages it cannot classify with confidence, and every message carrying a vulnerability or legal marker, go to the triage lead with the reason and the passages it relied on.

What stays with a person

The triage lead keeps the boundary cases. Whether an angry message about a price rise is a complaint or feedback depends on your definition and sometimes on context the text does not give. Vulnerability markers always reach a person, whatever the confidence score, because treating a vulnerable customer as routine costs far more than the minute it takes to look. The same applies to threats of legal action and anything that mentions a regulator or the press.

The agent does not decide redress, write the substantive response or close a complaint; those stay with the owning team. Nor does it log a message as not a complaint without leaving a trace: those decisions are sampled every week, because under-recording is the error nobody notices until an audit finds it.

What it reads, what it produces

It readsIt produces
Complaints mailbox, web forms, chat transcripts, call notes and scanned lettersA complaint case with category, product, issue and the quoted text supporting each
Customer accounts and order history in the CRMFollow-ups linked to the existing case instead of new duplicates
Open cases in the complaints systemAcknowledgements from approved templates, stating the case reference and deadline
Your written complaint definition and category taxonomyA queue for the triage lead, each item with the reason it was held back
The routing table: which team owns which category and productA daily list of cases whose deadline falls in the next two working days
Your response-deadline rules and acknowledgement templates

Controls that come with it

How you know it works

Is your process ready?

The five candidacy checks are explained, with an exam, in the free Module 01.

What goes wrong

Questions we get

Does the agent reply to the customer?

Only with an acknowledgement, and only from templates you approve. It confirms receipt, gives the case reference and states the response deadline your rules set. It does not write the substantive answer, offer redress or apologise for something specific, because those depend on facts the owning team has not yet established. You can switch acknowledgements off entirely and use the agent for classification and routing alone.

What happens when one message contains several complaints?

The agent splits it into linked parts, one per issue, each with its own category and owning team, and keeps the original message attached to every part. The customer receives one acknowledgement listing the issues. Where your rules require a single response per customer, the parts are grouped under a lead case owned by the team with the most serious issue.

How does it handle vulnerable customers?

Your policy names the markers: bereavement, illness, disability, financial difficulty, a third party writing on someone's behalf. The agent looks for them in the text and in any flag already on the customer record. When it finds one, the case goes to a person with the passage highlighted, whatever the classification confidence. The agent does not judge how vulnerable someone is; it makes sure a person sees it.

What do we need before starting?

A few hundred past complaints with the category and team they should have had, confirmed by your triage lead rather than copied from the system, since the history contains the errors you want to remove. That set measures the agent before go-live and after every change. You also need the written definition, the taxonomy and a current routing table.

Want this agent on your process?

Tell us about your version of this process — volumes, systems, what goes wrong. A person answers with an approach and a price, usually within two working days, or tells you it is the wrong project.