Skip to content
Aplora
eCommerce and retail
AI Workflow Automation

Client under NDA

An online store's support front line stopped growing with its order volume

An EU online store: 300–800 enquiries a day across chat and e-mail, with support scaled only by hiring. As the catalogue grew, the share of repeat questions grew faster than revenue did.

Published: 2026-08-24 · updated: 2026-08-24

An agent in a headset working at a laptop in an open-plan office
Illustrative photo of the working context, not a screenshot of the client's system

  • Enquiries that reach an operator

    −25–40%

  • Ticket handling time

    −30–50%

  • Support satisfaction (CSAT)

    +10–18%

In short

  • Problem: support scaled only by hiring, while the share of repeat questions grew with the catalogue.
  • Solution: topic and urgency classification, answers from the knowledge base, a ticket summary for the operator, and a content backlog built from real questions.
  • Result: 25–40% fewer enquiries reach an operator, handling time is 30–50% lower, CSAT is 10–18% higher — per the client's own figures.
  • Systems: helpdesk, store platform, internal knowledge base.

Context

  • An online store selling retail in the EU; region and product category under NDA.
  • Process volume: 300–800 enquiries a day across chat and e-mail.
  • Team: the support front line plus operations management.
  • Systems: helpdesk, ticket history, internal knowledge base; some integrations under NDA.
  • The constraint showed itself at seasonal peaks: during a sale, enquiry volume grew faster than staff could be rostered.

Baseline

Baseline metrics with their sources. Without them, any later result has nothing to be compared against.
  • Operators are absorbed by repeat questions
  • Complex enquiries get lost in the queue
  • The knowledge base goes stale faster than it is fixed
  • More volume runs straight into hiring more operators
  • Before the work started we fixed: the share of enquiries reaching an operator, average ticket handling time, and CSAT from the post-resolution survey.
  • Data source: the client's helpdesk exports.
  • The comparison ran on the same support channels and was adjusted for sales seasonality: a sale on its own changes both the volume and the mix of enquiries.
  • There are no measurement dates, calculation formula or approving owner behind these numbers — which is why they are published as a reported result rather than an audited one.

Diagnosis

Which hypotheses were considered, why this one was chosen, what was assumed, and the condition under which we would have stopped.
  • Three hypotheses were on the table: too few people on the front line, poor routing, or too many enquiries that need no human at all.
  • The third was chosen: reviewing ticket history showed that a sizeable share of questions repeat verbatim and the answer already exists in the knowledge base — nobody was getting to it.
  • The assumption: the knowledge base is substantively correct and the problem is access, not content. It was tested against a sample of closed tickets.
  • Stop criterion: if the assistant cannot answer at least half of the repeating topics correctly on historical enquiries, the hypothesis counts as unconfirmed and the work stops.

What we implemented

  • Data sources: ticket history, the knowledge base, order statuses from the store platform.
  • AI components: topic and urgency classification, answers from the knowledge base, a ticket summary for the operator, extraction of unanswered topics.
  • Business rules: payment, complaint and personal-data topics go straight to a person, with no attempt to answer.
  • Integrations: helpdesk, store platform, knowledge base.
  • Human checkpoints: the operator sees and can correct any answer before it is sent; a contested ticket returns to the manual queue.
  • Monitoring: every answer is logged, and the share handed to an operator and the share of corrected answers are tracked separately.

How the process changed

Before

6 steps
  1. The enquiry lands in a shared queue
  2. The operator reads it in full
  3. The operator looks for an answer in the base or asks a colleague
  4. The operator writes the answer in their own words
  5. The answer goes to the customer
  6. The ticket closes; the repeating topic is recorded nowhere

After

5 steps
  1. The enquiry is classified by topic and urgency
  2. An answer from the knowledge base, or a clarifying question
  3. A complex ticket goes to an operator with a ready summary
  4. The ticket closes
  5. Unanswered topics enter the knowledge base backlog

What was stuck

Operators were absorbed by repeat questions, complex enquiries got lost in the queue, and the knowledge base went stale faster than anyone could fix it.

  1. 1The enquiry is classified by topic and urgency
  2. 2An answer from the knowledge base, or a clarifying question
  3. 3A complex ticket goes to an operator with a ready summary
  4. 4The ticket closes
  5. 5Unanswered topics enter the knowledge base backlog
  6. Measured result

Results

Enquiries that reach an operator

−25–40%

Reported by: Client figures: before and after on the same support channels, adjusted for seasonality

Ticket handling time

−30–50%

Reported by: Client figures: before and after on the same support channels, adjusted for seasonality

Support satisfaction (CSAT)

+10–18%

Reported by: Client figures: their own post-resolution survey

Economic impact

  • The effect is freed front-line time: an enquiry closed without an operator costs only infrastructure.
  • Cost to serve an order goes down, but that alone adds no revenue — the saving materialises when the freed hours go into complex enquiries rather than into idle time.
  • Money figures are not disclosed: the front-line hourly rate and cost per contact are the client's commercial information.

Adoption

  • The support team was trained before launch rather than after.
  • An explicit rule was introduced for when an operator steps in: a stop-list topic, a second failed answer, or the customer asking directly.
  • The assistant became part of the standard handling process, not a separate tool used when someone remembers it.
  • The solution owner is support operations management, which also watches the share of corrected answers.
«We used to just hire people. Now half the routine questions close themselves and the operators work the hard cases.»

Client under NDA — Operations director, eCommerce

What's next

  • Scaling: adding the remaining support languages.
  • Next initiative: returns and exchanges — the same flow, but with external carrier systems involved.
  • What we decided against: letting the assistant decide about money — refunds, policy exceptions, discounts. That stays with a person.

Have a similar workflow? Let's check whether the hypothesis transfers

Another company's result is not a promise. It does show where to look.

All cases