ESTIMATING PRODUCTIVITY GAINS IN DIGITAL AUTOMATION Mauricio Jacobo-Romero1 Danilo S. Carvalho1 André Freitas12 1Department of Computer Science The University of Manchester Manchester UK

2025-04-29 0 0 1.3MB 11 页 10玖币

侵权投诉

ESTIMATING PRODUCTIVITY GAINS IN DIGITAL AUTOMATION

Mauricio Jacobo-Romero1, Danilo S. Carvalho1, André Freitas1,2

1Department of Computer Science, The University of Manchester, Manchester, UK

{mauricio.jacoboromero, danilo.carvalho, andre.freitas}@manchester.ac.uk

2Reasoning & Explainable AI group, Idiap Research Institute, Martigny, Switzerland

andre.freitas@idiap.ch

ABSTRACT

This paper proposes a novel productivity estimation model to evaluate the effects of adopting Artiﬁcial

Intelligence (AI) components in a production chain. Our model provides evidence to address the

"AI’s" Solow’s Paradox. We provide (i) theoretical and empirical evidence to explain Solow’s

dichotomy; (ii) a data-driven model to estimate and asses productivity variations; (iii) a methodology

underpinned on process mining datasets to determine the business process, BP, and productivity; (iv)

a set of computer simulation parameters; (v) and empirical analysis on labour-distribution. These

provide data on why we consider AI Solow’s paradox a consequence of metric mismeasurement.

Keywords AI productivity ·Solow’s paradox ·Process mining

1 Introduction

Every time we integrate new technology into a company’s operation, we assume it will increase the current productivity;

because the same volume of production would be reached out using smaller inputs of other factors of production,

or, with the same resources, we might increase the production [

]. Solow-Swan’s model showed that increments in

capital and labour only conduct to a time increase until the organisation reaches a steady-state growth. Meanwhile,

technological changes deliver permanent growth [

]. However, in the 1970s and 1980s, the computing revolution, the

massive introduction of computers in all industry sectors, did not create signiﬁcant productivity growth [

]. The Nobel-

Prize-winning economist, Robert Solow, noticed this abnormality and stated: "we see the computer age everywhere

except in the productivity statistics". This argument is called the "Solow’s paradox" [4].

Since Solow’s paradox formulation, it caught the researcher’s attention. By the mid-1990s, they proposed four groups

of explanatory arguments: (1) false hopes, (2) productivity mismeasurement, (3) concentrated distribution and rent

dissipation, and (4) implementation and restructuring lags[5].

In 1996, researchers ﬁnally deciphered the paradox: “We conclude that the productivity paradox disappeared by 1991, at

least in our sample of ﬁrms” [

]. According to this study, productivity models were based on loosely limited statistical

analysis; these prototypes underestimated the economic beneﬁts of computers. Additionally, new information systems

implementation was problematic due to a shortage of specialised personnel.

Our central motivation is developing a methodology to interpret how digital automation (accelerated by recent advances

in the AI space) impacts production chains and causes labour displacements. Relevant literature contains few models

to forecast the effect of AI integration into business processes. These techniques require considerable amounts of

interventional data to produce results. Unfortunately, well-controlled interventional studies in production chains are not

widely available[7].

With these requirements and constraints in mind, we sought to: (i) provide theoretical and empirical evidence to

address the Solow’s paradox (”lack of evidence of productivity improvement for AI-based interventions”); (ii) develop

a micro-economic level, data-driven model to estimate and assess productivity variations after automation interventions

within business process (BP) workﬂows; (iii) building a methodology that uses process mining to support productivity

estimation and analysis within BP workﬂows; (iv) determine a systematic method to parametrise productivity estimations

arXiv:2210.01252v2 [cs.AI] 8 Oct 2022

Estimating productivity gains in digital automation

for BP workﬂows; (v) and provide an empirical analysis of labour-redistribution of the model on a well-document

public case study.

This work is organised as follows: Section II provides related work, and Section III introduces key economic modelling

concepts. Section IV describes our methodology. The proposed model is explained in section V. Section VI and VII

introduce results and discussion, respectively. Lastly, we present our conclusions.

2 Related Work

In the mid-2010s, early Artiﬁcial Intelligence (AI) ﬁrm adopters stated that the acquisition of Natural language

Processing (NLP) systems generated savings and productivity increments [

]. However, empirical evidence showed

that productivity beneﬁts were not easy to track [

]. AI integration exhibited the conditions of a Solow’s Paradox redux.

This phenomenon motivated several works. They aimed to identify the reasons for the absence or emergence of AI

productivity effects [

]. These studies followed two main approaches: pessimistic reading of the empirical past and

optimism about the future [

]. The ﬁrst group suggests that AI provided beneﬁts, but statistics did not accurately capture

them. The second one states that AI advantages are not yet part of the business operation [

]. Both perspectives employ

macroeconomic information and lack the suitable tools to seize the data up to the task level [

]. As a consequence,

productivity assessment transfers AI beneﬁts to other evaluation axes. The outcome, then, is an averaged result[

]. On

the other hand, the optimistic course implies that there is a period in which technology is not mature enough to produce

a discernible inﬂuence on productivity growth[13, 1].

It is clear, then, that a model capable of measuring productivity variations at a task level can provide the methodological

support to disambiguate the current controversies. Some other works have found that a mismatch between technology

and workers’ skills negatively affects productivity [

]. AI systems assemble new "hybrid" duties that become part of

the automated process[

]. The analyses conclude that labour composition is another factor that might harm productivity

estimation [

]. In other words, AI systems might require more labour to execute the automated BP [

]. Hence,

process inputs would change, and so the production volume [

]. Therefore, labour composition is another relevant

factor for AI productivity assessment.

Apart from these two main study segments, we identiﬁed a third one: AI Capital. AI capital is the portion of the

income generated by those activities automated with AI systems [

]. AI cash inﬂows are not challenging to track [

Unfortunately, it is complex to determine the root cause of these income streams [21]. Some authors state that market

conditions are responsible for the generated wealth [

]. In other words, AI Capital models require considerable data to

produce results.

The three approaches employ either Cobb-Douglas equations or variations to represent companies’ production [

]. This function family is exponential [

]. Hence, parameter determination requires considerable effort [

]. To ease

computations, researchers employ logarithmic techniques. In this fashion, linear regression methods are suitable to

determine production parameters[

]. Unfortunately, business processes store averaged information across different

databases[

]. Hence, AI-systems contributions dilute across the business [

]. Determining the productivity gains

behind digital technologies provides a fundamental translation of value between the technology investment and the

ﬁnancial results [1, 10].

To tackle this issue, we explored emerging process mining frameworks which induce and business processes based

on event logs[

], where an event is the basic unit of a BP. Each event provides basic information such as completion

timestamp, executor, and activity costs[

]. There are several process mining log formats. However, process mining

researchers and practitioners adopted the IEEE XES format as a standard [

] which has a formal and canonical schema,

recording compulsory ﬁelds and offering the possibility of including customised data in XML structure[

]. We

found that process mining would provide adequate granularity and a feasibility argument to facilitate the measurement

of productivity changes, dialoguing with the lower barriers provided by logs in contrast to competing more formal

frameworks (such as Business Process management - BPM workﬂows).

Our examination of the relevant literature revealed a three-point gap: 1) labour productivity analysis should cover the

task level characterisation[

], 2) the interest in Solow’s Paradox grew more towards measuring the impact of shifts

in labour input than towards the identiﬁcation of the sources of AI capital streams[

], 3) process mining might provide

the required data to produce better productivity evaluations[27].

3 Estimating Productivity: Key Deﬁnitions

In this section, we outline key critical deﬁnitions used throughout this paper.

Estimating productivity gains in digital automation

Business Processes (BPs).

A Business Process (BP) is a series of coordinated activities that deliver a service or

product[

]. These production chains are usually represented with Business Process Model and Notation(BPMN)[

Each task is an abstraction of a production step that is parametrised in terms of production time and skill-set distribution

[32]. These sketches provide information on task ownership, dependencies, information inputs and outputs [33].

Productivity.

Productivity is the relation between the output generated by a BP and the required resources to create

this result [34]. Then, productivity can be described as follows:

P=Y

T.(1)

where production/output,

, is the estimated volume of products/services generated by a ﬁrm, and

is the labour input,

the number of worked hours, to produce these outputs [29].

Solow’s paradox.

"‘You can see computers everywhere but in the productivity statistics’, wrote Robert Solow in 1987.

His dictum spawned several decades of economic research aimed at solving the mystery that has become known as the

‘Solow Paradox’: massive investment in computers but no net gain in productivity" [

]. "Solow’s paradox" attracted the

attention of researchers and, after several years, some examinations concluded the paradox disappeared, at least for

the analysed sample: "our sample consisted entirely of relatively large ‘Fortune 500’ ﬁrms" [

]. From the proposed

conjectures, mismeasurement, was one of the most analysed. "The closer one examines the data behind the studies of

IT performance, the more it looks like mismeasurement is at the core of the ’productivity paradox’" [35].

Based on this, we decided to orchestrate our efforts toward expanding an answer for the "Solow’s Paradox". Thus, we

propose a model to analyse productivity variations due to the introduction of automation changes, in the digital context,

motivated by the growth in the adoption of AI.

Stochastic queues.

A stochastic queue is a system that consists of two parts: a server and a queue. An event item

arrives in the queue. The time for the item to be processed is dependent on the waiting (queueing) time and the service

(processing) time[

]. Our strategy is to substitute each BP activity for a queue system. Each system belongs to a

speciﬁc queue category. In this manner, it is possible to simulate the introduction of automation in a given BP. We

deﬁned three types of queue systems, dialoguing with the Cobb-Douglas equation: low-skilled workers, high-skilled

workers and automated systems. Queue system type selection depends upon the activity description.

4 Methodology

We propose a novel method to estimate productivity and labour distribution changes of BPs after the introduction of

automation. For that purpose, we designed the experiment described in ﬁgure 1.

We analysed two datasets, 2012 and 2017, from the BPI Challenge. They describe a process that belongs to the same

ﬁnancial institution[

]. The process changed over the time. Automation was integrated into the process. Thus, we

found information about the BP state before and after the intervention.

As part of the BPI Challenge, participants produced a report on every workﬂow being represented via the logs. We

employed this information to identify the labour composition of the BP.

We applied the proposed parametrised queue-based productivity model and were able to produce a report on productivity

ﬂuctuations and queue system parameters. Our model presented a novel method to analyse productivity variations a

priori and posteriori with minimum information requirements (building upon existing process mining frameworks).

Moreover, the same model can be used to report labour displacements.

5 Proposed model

Economists usually express production functions as Cobb-Douglas equations. This family of functions are exponential

and depends on two variables: labour and cost of capital. The exponent is known as the elasticity factor and reﬂects the

impact of these two inputs in the production[39].

Unfortunately, the Cobb-Douglas functions do not provide information on labour distribution. Thus, we employed an

equivalent production function:

Y= (L+AX)αH1−α

where

is the production of some good, that employs high-skilled labour hours,

, low-skilled labour hours,

, an

input good that substitutes low-skilled labour,

, and applies

α∈(0,1)

to trace the relative shares of the high and

low-skilled labour at a rate A[16].

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ESTIMATINGPRODUCTIVITYGAINSINDIGITALAUTOMATIONMauricioJacobo-Romero1,DaniloS.Carvalho1,AndréFreitas1;21DepartmentofComputerScience,TheUniversityofManchester,Manchester,UK{mauricio.jacoboromero,danilo.carvalho,andre.freitas}@manchester.ac.uk2Reasoning&ExplainableAIgroup,IdiapResearchInstitute,Martign...

展开>> 收起<<

ESTIMATING PRODUCTIVITY GAINS IN DIGITAL AUTOMATION Mauricio Jacobo-Romero1 Danilo S. Carvalho1 André Freitas12 1Department of Computer Science The University of Manchester Manchester UK.pdf

共11页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

ESTIMATING PRODUCTIVITY GAINS IN DIGITAL AUTOMATION Mauricio Jacobo-Romero1 Danilo S. Carvalho1 André Freitas12 1Department of Computer Science The University of Manchester Manchester UK

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: