![](bg2.png)
cated than others. Further, a customer service e-mail
might contain two paragraphs of text, one detailing
a technical issue, and the other one an order errand.
As such, the e-mail topic would be sorted as Invoice,
TechicalIssue, and Order. This would further affect
response time.
1.1 Aims and Objectives
In this study we investigate the possibility to predict
the time-to-respond for received e-mails based on its
content. If successful, it would be possible to adjust
the schedules for customer support personnel in or-
der to improve efficiency. The two main questions
investigated in this work are as follows. First, to what
extent it is possible to predict the time required by
customer support agents to respond to e-mails. Sec-
ond, to what extent it is possible to predict the time it
takes customers to respond to e-mails from customer
support personnel.
1.2 Scope and Limitations
The scope of this study is within a Swedish setting,
involving e-mail messages written in Swedish sent to
the customer service branch of the studied telecom
company. However, the problem studied is general
enough to be of interest for other organizations as
well. In this study, e-mails where no reply exists have
been excluded, as it has been suggested to be a sep-
arate classification task (Huang and Ku, 2018). Fur-
ther, time-to-respond (TTR) is investigated indepen-
dent of the workload of agents, and the content of the
e-mails.
2 RELATED WORK
Time-to-Respond, or responsiveness, can affect the
perceived relationship between people both posi-
tively and negatively (Church and de Oliveira, 2013),
(Avrahami and Hudson, 2006), (Avrahami et al.,
2008).
Investigations into mobile instant messaging (e.g.
SMS) indicates that it is possible to predict whether a
user will read a message within a few minutes of re-
ceiving it (70.6% accuracy) (Pielot et al., 2014). This
can be predicted based on only seven features, e.g.
screen activity, or ringer mode.
Responsiveness to IM has been investigated, and
been predicted successfully ( 90% accuracy) (Avra-
hami and Hudson, 2006). The paper where limited
to messages initiating new sessions, but the model
where capable of predicting whether an initiated ses-
sion would get a response within 30s, 1, 2, 5, or 10
minutes. Predicting the response time when inter-
acting with chatbots using IM have also been inves-
tigated, within four time intervals < 10s, 10 − 30s,
30 − 300s, and > 300s (Accuracy of 0.89), but also
whether a message will receive a response (Huang
and Ku, 2018).
Similarly to IM, response time in chat-rooms have
also been investigated, with one study finding that
the cognitive and emotional load affect response time
within and between customer support agents (Rafaeli
et al., 2019). In a customer support setting, the cogni-
tive load denotes e.g. the number of words or amount
of information that must be processed. TTR predic-
tions have also been investigated in chat rooms (AUC
0.971), intending to detect short or long response
times (Ikoro et al., 2017).
However, it seems that there is little research that
have investigated predicting the TTR of e-mails in a
customer support setting. This presents a research gap
as it has been argued that e-mails are a distinct type
of text compared to types of text (Baron, 1998). Re-
search indicates that it is possible to estimate the time
for an e-mail response to arrive, within the time inter-
vals of < 25 min, 25 − 245 min, or > 245 min (Yang
et al., 2017). Similarly, research has been conducted
on personal e-mail (i.e. non-corporate) (Kooti et al.,
2015). However, this investigates quite small TTRs
which, although suitable for employee e-mails, might
not conform to the customer support setting according
to domain experts. Further, the workload estimation
of customer support agents work resolution benefits
from an increased resolution, i.e. more bins.
3 DATA
The data set consists of 51, 682 e-mails from the cus-
tomer service department from a Swedish branch of
a major telecom corporation. Each e-mail consists of
the:
• subject line,
• send-to address,
• sent time, and
• e-mail body text content.
Each e-mail is also labeled with at least one label. In
total there exists 36 distinct topic labels, each inde-
pendent from the others, where several of these might
be present in any given e-mail. The topics have been
set by a rule-based system that was manually devel-
oped, configured and fine-tuned over several years by
domain expertise within the corporation.
ICEIS 2020 - 22nd International Conference on Enterprise Information Systems
306