KT Sparks

Vehicle rental · Europe · multi-billion-euro rental group with hundreds of thousands of cars

Car Rental Contact Centre: Email Triage and One-Click Agent Assist

Extending a rental, cancelling it or swapping the car meant an agent paging through Salesforce screen by screen while the customer held on. Emails had to be read by someone just to find out what was being asked. We added UiPath robots behind the agents. Emails covering five request types are now classified and processed automatically, and on a call the same requests take a single click.

Car Rental Contact Centre: Email Triage and One-Click Agent Assist
Industry
Retail Customer Service
Function
Customer Service & Contact Centre

Results

5
request types now automated
Extension, cancellation, roadside assistance, vehicle change, drop-off change. This is scope, not a saving.
2
channels sharing one automation layer
Email runs unattended, phone runs attended. This is scope, not a saving.
1 click
for each agent action during a call
Replaces several Salesforce screens and status changes. Call time was not recorded.

01

The challenge

The client ranks among the largest car rental groups in Europe. It is a multi-billion-euro business, and its contact centre runs on Salesforce across several offices.

Calls and emails came in with the same small set of requests: extend the rental, cancel it, send roadside assistance, swap the vehicle, or change where it is returned. For each one an agent had to step through several Salesforce screens and status changes by hand. Emails added a step before that: a person read every message only to decide which of the five requests it contained, and only then could the work begin.

Where the cost showed up

  • Longer calls. Each additional screen adds time to the current call and to every call queued behind it.
  • Longer queues. While an agent is navigating Salesforce, the next customer waits.
  • Manual email sorting. Each message had to be read and sorted by a person before any action was taken.
  • Records left half-done. Status changes spread over many screens, made while a customer is on the line, are exactly where records get only partly updated.
  • Revenue lost. When an interaction is slow or dropped, a rental goes unextended or a booking stays unchanged, and that money is gone.

02

What we did

We delivered UiPath automation that connects to Salesforce in two ways: the API and UI automation. It covers two channels. On email, unattended robots read customer messages and act on them. On the phone, an attended agent-assist tool reduces a multi-screen Salesforce task to one click while the call is in progress.

Email: no human in the loop

  1. Reading the intent. Communications mining processes every incoming customer email, tags it with one of the five request types and attaches a confidence score. Nobody has to read emails to sort them anymore.
  2. A confidence threshold. Emails with high confidence are placed in a queue as transaction items for an unattended robot. Anything below the threshold is not processed automatically, so the robot never acts on a guess.
  3. Doing the work in Salesforce. The robot completes the request in Salesforce, whether that is an extension, a cancellation, roadside assistance, a vehicle change or a new drop-off location.

Phone: agent assist in one click

  1. A pop-up of actions. As a call arrives, the agent gets a pop-up listing the actions that apply.
  2. A single click. Picking an action starts the automation in the background, so the agent no longer steps through screens and status updates.
  3. The agent keeps talking to the customer while Salesforce is updated by the robot.

API first, UI where it has to be

Each channel writes to Salesforce both through the API and through UI automation. That meant we could automate every action, including those Salesforce does not offer through its API. The email robots and the phone agents share one integration layer, so a request is processed identically whether it came in by email or by phone.

Stack

LayerTools
AutomationUiPath Studio with attended and unattended robots
Email classificationCommunications mining, intent labels with confidence scores
CRMSalesforce via API and UI automation
Agent toolsAgent-assist pop-up for the contact centre
Distributing workTransaction processing through queues

03

The outcome

Instead of paging through screen after screen, agents click once, and the most frequent email requests get sorted and processed without a person reading them first.

BeforeAfter
Sorting incoming customer emailsA person read each oneLabelled automatically with a confidence score
Handling a high-confidence emailManual work in SalesforceDone by an unattended robot
Changing a rental during a callSeveral screens and status updatesA single click
Emails with low confidenceA person read each oneLeft alone by the robot, no guessing
One request arriving by email or phoneTwo separate manual routesA shared Salesforce automation layer
  • Calls got shorter. What used to take several screens now takes one click, and call durations fell.
  • Each agent handles more transactions, and productivity rose in every one of the client's offices.
  • High-confidence emails for the five request types were sorted and processed with no manual reading.
  • Revenue went up. The client reported higher revenue. This comes from the delivery team's recollection and is not a documented figure.
  • A direction, not a measurement. The engagement did not record volumes or timings, so the numbers on this page describe scope rather than savings.

What to watch

Salesforce puts out three major releases every year, and any of them can shift the screens a UI robot relies on. A one-click button that breaks in the middle of a call does more harm than having no button. Wherever Salesforce exposes an action through the API, we used it, which keeps most of the work safe from screen changes. The steps automated through the UI are the ones to retest around every release.

For a simpler version of this build, see refund processing for in-flight wifi at a UK airline group. There the mailbox had a single purpose, so no intent classification was needed.