Use casesServices

Outsourced complaints desk

One team answers written complaints for several clients, each with its own rules, tone, and deadlines.

Kinds of AI
SLMsAgentic AI
Who decides
Team lead
FIG. 1 · Routed complaintIllustrative
Client B (retail) · complaintLate delivery

Agent · facts from the client's systems

  • DeliveryPromised date passed by 12 daysAPI

SLM · client B's policy and reply

  • ClauseDelay over 10 days → voucher per the matrixclient policy §3
  • ProposedVoucher within the matrix, from the calculator
  • ReplyDraft in client B's tone
  • SLM
  • Agent

Why a person decidesNot a template case, so the drafted reply goes to the team lead.

DecidesTeam lead

1. The desk

Today

  1. ReadEvery complaint is read and coded by hand.
  2. RouteDeadlines and regulator escalations are tracked in spreadsheets.
  3. WriteComplaint handlers switch between clients, mix up policies and compensation rules, and answer each complaint by hand.
  4. DecideThe team lead checks replies against each client's rules by hand.

With the system

  1. ReadThe routing model, trained for each client, reads the complaint and returns its category, priority, deadline, flags, and confidence.SLM
  2. RouteThe agent sends it down one of three paths by the client's rules (template reply, drafted reply, or team lead) and gathers the facts from the client's order and CRM systems.Agentic AI
  3. WriteThe client's own SLM drafts the reply in that client's tone with the policy clause cited; a calculator proposes compensation from the client's matrix.SLM
  4. DecideReplies from approved templates go out; the team lead reviews every other reply and decides, as before; anything outside the matrix needs the desk supervisor.Person decides

What lands on the desk

The team lead gets the complaint with its facts already pulled, the client's own clause that applies, and a reply drafted in that client's voice. The compensation comes from the client's matrix, and nothing outside a template goes out without the team lead.

FIG. 2 · Routed complaintIllustrative
Client B (retail) · complaintLate delivery

Agent · facts from the client's systems

  • DeliveryPromised date passed by 12 daysAPI

SLM · client B's policy and reply

  • ClauseDelay over 10 days → voucher per the matrixclient policy §3
  • ProposedVoucher within the matrix, from the calculator
  • ReplyDraft in client B's tone

Agent · deadline

Answer due in the policy window; the clock started at receipt.

  • SLM
  • Agent

Why a person decidesNot a template case, so the drafted reply goes to the team lead.

DecidesTeam lead

2. The system

For your technical team — skip it if you like.

How it works

In

  • The complaint: email, portal, letter

  1. SLM · understands

    Complaint router

    The request-routing model, trained per client

    Reads

    • Everything under In

    Out

    • Category, priority, deadline
    • Flags and confidence

  2. Agent · prepares

    Routing agent

    Router

    Reads

    • The routed complaint

    Out

    • A path by the client's rules
    • Template, drafted reply, or team lead
    • Facts from the client's order and CRM APIs

  3. SLM · understands

    Client reply writer

    A separate adapter, trained for this client only

    Reads

    • The facts
    • The client's policy
    • The compensation matrix

    Out

    • A reply draft in the client's tone
    • The policy clause cited
    • Compensation by a calculator from the matrix

  4. Decision

    Templates and team lead

    Reads

    • The reply

    Out

    • Replies from approved templates go out
    • Everything else waits for a team lead
How the agent's work is arrangedRouter

Router

The first component already classifies, so the agent only sends each complaint down one of three fixed paths by the client's rules.

How the models are trained and tested

Your expert writes the test first. A step the test does not need is skipped, and we say why.

Show the training steps

Client SLM (one per client)

  1. Separation

    One model (an adapter or a full small model) per client; one client's data never trains another client's model; separate storage, keys, and logs.

  2. Test set

    Per client, 100–200 complaints graded by that client's senior complaint handlers (policy, tone, compensation).

  3. Dataset

    That client's policies, past complaints, final replies, and outcomes (reopened or escalated).

  4. Continued pre-training (CPT)Skipped

    Skipped: the task needs no field language the base model lacks; fine-tuning teaches each client's format.

  5. Supervised fine-tuning (SFT)

    On accepted replies.

  6. Reinforcement learning (RL)Skipped

    RL or preference optimization (DPO, which learns from a draft paired with the supervisor's corrected version) only once volume allows; skipped at launch.

  7. Regression test

    Per client, before every model update.

Where the agentic system stops

Replies from the client's approved templates go out; everything else waits for a team lead.

Show its tools, limits, and record
Tools and access
Read the client's order and CRM APIs; send only approved templates; draft everything else.
May not
Send a non-template reply, grant compensation outside the matrix, or mix data between clients.
Who decides and the written limit
Team leadReplies from the client's approved templates go out; everything else waits for a team lead. The client's policy owner sets the compensation matrix; anything outside it needs the desk supervisor.
Stops when
Money above the matrix, legal wording, a regulator or arbitration escalation, or low routing confidence: the complaint goes to a team lead.
Record
The client, route, facts, clause cited, and who sent the reply.

How agentic systems stay inside limits

3. Value and conditions

Where the time goes, and what your institution will measure.

Where the value comes from

  1. Handling time, and deadlines met.
  2. Fewer policy errors across clients.
  3. Faster onboarding of a new client.

What your institution measures

  • Average handling time
  • First-contact resolution
  • Reopen rate
  • Regulator or arbitration escalations
  • Policy errors in audits
  • Onboarding time for a new client

Measured in the Decision Blueprint

Baseline handling time and audit error rate per client set the starting point; your institution sets the target.

Conditions

The rules, the data, and the known error, with its guard.

  • The desk is a data processor (veri işleyen) for each client under KVKK; contractual separation between clients is audited.
  • Brand tone is kept per client.
  • Money and legal wording are always reviewed by a person.

Start here

We take a sample of complaints and the policies of two or three clients, measure handling time and audit errors per client, and check what each client's templates already cover before anything is built.

Request a Decision BlueprintWhat the Decision Blueprint covers
  • SLMsAgentic AI

    Contact-center speech: quality and compliance on every call

    TodayEvery call is recorded, quality hears two in a hundred, and off-the-shelf speech-to-text mangles your product names.

    With the systemEvery call checked for quality and compliance; the coaching point in context.

    Team lead decides

  • SLMsAgentic AI

    Agent assist

    TodayNew call-center agents take months to be productive and put customers on hold to search the knowledge base.

    With the systemAnswers drafted while the customer is on the line.

    Call-center agent decides

Equalize the gap.

Bring one decision. Leave knowing the first step.

A short description of the decision is enough; no confidential records.

Start with a Decision BlueprintTalk to the engineer

https://eqlx.ai/use-cases/complaints-desk · Printed copy; the website is the current version.