Harnessing AI in Industrial Policy:

Turn Data into Action

Minho Kim (KDI)

For Global Development Learning Network
 On the background

Formulation of Industrial Policy

STRATEGIC
INDUSTRY PRODUCTIVITY Resilience -
OBJECTIVE DECARBONISATION
AUTONOMY
STRATEGY
DESIGN

Technology
Mission Oriented Place-Based Sectoral
TYPE Focused

Instrument
The strategy
provides a Strategy Instrument

rationale for Instrument

instrument
choice
CHANNEL Between Within
Demand Governance
(Supply)
INSTRUMENT

(Supply)
CHOICE

SCOPE Horizontal Targeted

2
Source: Criscuolo et al. (OECD, 2022)
 On the background

Taxonomy of policy instruments

3
Source: Criscuolo et al. (OECD, 2022)
 On the background

Industrial Policy in the Fourth Industrial Revolution

▪ After the global financial crisis, countries are paying renewed attention to the
importance of manufacturing and implementing policies to improve
manufacturing competitiveness

• Germany ‘Industrie 4.0’, USA ‘Manufacturing USA’,

Japan ’Japan Revitalization Strategy’, China ‘Made in China 2025’ etc.

▪ These policies aim to make manufacturing production more efficient, more

flexible, and higher value-added based on the concept of the smart factory.

▪ Introduction to Korea’s Policy in the Digitalisation (AI Integration) of

Manufacturing

• South Korea has been actively promoting the digital transformation of its
manufacturing sector
• Based on a research: Smart Factory: Economic Impacts and Policy
Implications (Minho Kim, Sunghoon Chung, and Changkeun Lee)

4
 On the background

Industrial Policy in the Fourth Industrial Revolution

Source: https://www.bossard.com/en/smart-factory-logistics/blog/2018/02/what-are-
industry-40-and-the-internet-of-things-all-about/
5
Korea’s Manufacturing Innovation Strategy

▪ Korean economy relies heavily on manufacturing compared to other developed

countries

▪ Korea continuously comes up with measures to strengthen its competitiveness

in manufacturing, especially supporting SMEs.

▪ The main strategy is to have ‘30,000 smart factories by 2022’.

• It supports the introduction of smart factory-related technologies to

manufacturing SMEs.

• Smart factory is selected as eight leading projects of 'Innovative Growth

Leading Project‘

• The budget for smart factories has risen sharply, resulting in a budget of
about KRW 1 trillion (USD 850 million) in 2019.

6
So far, so good?

▪ Concerns about the effectiveness of Korea's smart manufacturing

innovation strategy

• ‘30,000 smart factories by 2022’ is very similar to ’30,000 SMEs IT

support project’ in terms of its target and delivery process

• Only half of enterprises actually used IT systems built at the time*

* SME IT Service Information Survey Report, Korea Chamber of Commerce and Industry(2005)

7
 Questions

Something can be improved with the policy..

What is smart factory?

How do we know how smart a factory is?

What are the consequences of the factory smartization?

Which factories are becoming smart?

(What makes a factory become smarter?)

Did the government support help to increase the level of smartness?

8
 What we did..

✓ We participated in the field trip program to learn about smart factory

• Hannover Messe 2018 in Germany

• Trends and techs in manufacturing
• Camera and AI sensing defects
• Sensors to detect rpm and temperature of engine
• Siemens AG factory (making motors)
• BCG Innovation Center for Operations (Demo factory making candies)

https://www.youtube.com/watch?v=KUKl-c5uWbM

✓ Technical definition

A factory that digitizes and networks all processes, products, and resources
(Zuehlke, 2018; Kagermann et al., 2013; Bitkom e.V. et al., 2016).

9
 What we did..

✓ We visited and interviewed factories in Korea that implemented smart factory

technologies

• Factory A (oil filters for autos):

• Error occurred after installing MES(Manufacturing Execution System)
=> Return to the old practice

• Factory B (plastic products):

• Not sure about the effects of installing MES, but applied to get subsidy
from the government

• Factory C (industrial pumps):

• Real time sharing of information between production department and
sales department

10
 How we define smart factory

✓ Key Features of Smart Factory

(1) ICT & OT based System Integration (Kagermann et al., 2013)

1. Vertical integration of production system

2. Horizontal integration of the product value chain

(2) Data Share and Use (McAfee and Brynjolfsson, 2012)

1. Sharing (digitized) data as the lifeblood

2. Using the data for decision-making

• Other features are interesting but somewhat optional.

- Flexible & adaptive productions, human-machine interactions, ...

11
• Horizontal and Vertical Integration of Factory System

12
 Economic Interpretation
• Smartness of a factory as a particular form of organizational capital.

• System integration needs complementary management practices

• Implementing one IT can fail to work for various reasons
“If you’re not careful, the dream of information integration can
turn into a nightmare.” &
“Those companies that stressed the enterprise, not the system,
gained the greatest benefits.” (Davenport, 1998).

• Data share and use (DSU) are organizational practices.

• Collection of and reliance on data for managerial activities (Data-
driven decision making, DDD) has grown rapid in the U.S. and is
strongly associated with productivity increase
(Brynjolfsson and McElheran, 2016 & 2019)
• DSU requires managerial time and worker efforts to find, to
measure, and to utilize data.

These two jointly form a factory-specific capital called smartness.

• In this sense, factory accumulates—not installs—the smartness.

“smartization” vis-a-vis automation and robotization
13
 Goal

▪ Provide an empirical analysis on the impact of factory’s smartization on its

performance and employment

▪ Make suggestions on policies to effectively support SMEs

1. Measure the smartness level of each factory by surveying 1,000 domestic

manufacturing plants.

2. Study the effects of the smartization

3. Find out factors affecting the smartization

14
 Survey Data
Survey Sampling
• Target 1,000 sample factories in six 3-digit and two 5-digit industries.

15
Measure of the Factory Smartization
▪ How do we know how smart a factory is?

• System Integration (SI)

• Questions on degrees of horizontal and vertical integration (C5 & C6)
• The level of SI is calculated as the average of the two answers.

• Data Share and Use (DSU)

• Questions on degrees of data share (C11) and use (C10)
• The level of DSU is the average of all answers of 10 sub-questions.

• Factory Smartization is measured by the average of SI and DSU.

16
Level of smartization
▪ The level of smartization improved in 2017 compared to 2015

• SI level is at 1~2 out of 5

• DSU level shows bell-shaped distribution

17
Change in the level of smartization

18
 Empirical Model

The effects of factory smartization

• Δ = 2 years between 2015 and 2017

19
The effects of factory smartization Results

• Technology alone does not guarantee productivity gains. 20

 Empirical Model

The effects of factory smartization

▪ The performances are likely to vary with the type of production process
1. Small batch or job shop: individual or small group production
• More variety, complex (e.g., prototype, mould, furniture)
2. Assembly line: Step-by-step production along the production line
• Complex, many parts (e.g., electric and electronic products, car)
3. Continuous process: automated facilities running non-stop
• Mass production of homogeneous goods (e.g., steel, beer, paper)

21
 Results

▪ What are the consequences of the factory smartization?

• Induces improvements of productivity and cost efficiency.

• Increased daily production for all types of process
• Reduced lead time for continuous process
• Reduced defect rate for assembly line

• Conducive to develop a new product variety and attract a new customer.

• for all types of process but more for small batch

• Effects are heterogeneous depending on the type of production process.

22
Determinants of Factory Smartization Empirical Model
▪ Model

• 𝑇𝑒𝑐ℎ : 12 key technologies in smart manufacturing

• 5 operation technologies(PLC, RFID, SCADA, big data analytics, CPS), 7
information technologies(PLM, MES, ERP, SCM, FEMS, cloud, AI)

• Structured managment practices for incentives as in Bloom et al (2012)

− pratices on connecting bonus, promotion, and reassignment or dismisal
to the performance and ability of both managers and non-managers

• CEO’s risk preference, interest in process innovation

• dummy for the presence of a department or personnel dedicated to

technology-based process optimization,

• quintiles by the employment share of process technicians,

• dummy for the presence of labor union

23
Determinants of Factory Smartization Results

24
 Results

▪ Which factories are becoming smart? (What makes a factory become smarter?)

• Technology adoption leads to smartization

• There is a complementarity between human resource management

practices and technology adoption.
• The level of smartness increases disproportionately with the level of
structured human resources management.

• The presence of an ICT division and CEO's willingness to upgrade the

manufacturing system are also significantly important factors

25
 Results

▪ Which factories are becoming smart? (What makes a factory become smarter?)

• It is inefficient to support technology adoption when the factory is not

ready!

Digitization Management Smart

Leadership Skills
(Technology) Practice Factory

26
 Policy recommendations

Support for SMEs

▪ Clearly show long-term commitment to support manufacturing

firms
• Due date and target numbers in ‘30,000 smart factories by
2022’ are visible strategy but this gives an incentive for firms to
hurry to get government supports.

1. Take “pre-consulting & post-support” approach

• before providing support, there needs to be an accurate diagnosis of
the target firm's organizational practices, needs, and expectations for
the effects of the smartization.
• This approach will be more effective than merely helping firms to
introduce specific technologies.

27
 Policy recommendations

Support for SMEs

2. Build a platform that provides one-stop services for SMEs.

Case: Singapore, Enterprise Singapore

Source: https://www.enterprisesg.gov.sg 28
 Policy recommendations

Support for SMEs

3. Make the government's support system smart (government smartization)

to increase the effectiveness of the support.

Case: Germany, Map of Industrie 4.0 use cases

Source: https://www.plattform-i40.de/
29
 Policy recommendations

▪ Cooperation between SMEs and large corporations is crucial in pursuing smart

manufacturing in the Korean business environment.

▪ New program: When the host organization (large company, etc.) builds a
smart factory in cooperation with small and medium-sized companies, the
government supports part of the cost.

▪ Making test centers accessible to SMEs

▪ Establishment of smart factory test beds by region (6 locations)

▪ Running smart manufacturing innovation center at Technopark

(19 locations)

▪ Enabling regional clusters

30
 Another research: Applying AI on government program

▪ How to build a government system for data-driven decision making

• Show an application of machine learning on targeting firms using

government support programs recipients data.

• Explore the possibilities and cautions of applying AI on government

program.

31
 Questions

Something can be improved with the policy..

▪ The problem of finding the policy target is common:

in business support policy (R&D support, export support, start-up)
in welfare policy (cash and consumption coupon, energy voucher, or
supports for people or business under Covid-19 crisis)

When the government finds support targets with a limited budget,

does the government make decision based on evidence or data?

Once it hands out money to business or people, does the government use
the results to improve the efficacy of support?

32
 Questions

▪ Despite high expectations for AI’s predictive power to drive innovation in

the public sector, Korea has yet to fully apply it to core policy tasks—
such as identifying target groups or managing follow-up actions.

▪ Current AI usage in Korea’s public services remains limited, often restricted

to automating simple, repetitive tasks.

▪ A few leading countries are beginning to use AI-driven analysis to select

promising policy targets and monitor project performance in real time.

▪ This marks a shift from traditional ex-post evaluation to data-driven, ex-

ante policy forecasting, enabling more proactive and effective decision-
making.

33
 Policy

Government Credit Programs for SMEs

▪ Many governments implement policy instruments to help SMEs access

financing, enabling their growth and job creation.

▪ Public guarantee schemes

• offer SMEs guarantees for bank and trade credits, transferring the risk
from financing institutions to public institutions
• currently available in nearly 100 countries (Green, 2003)

▪ Government subsidized loans

• provide credits at a cheaper interest rate than the average market
interest rate or that are not available from other sources.

▪ Relative to all outstanding business loans to SMEs in 2018,

• Korea: guaranteed 9.7%, provide direct loans 0.6%
• Japan: guaranteed 8.3%, provide direct loans 7.9%
• US : guaranteed 4.6% (Financing SMEs, 2020).

34
 Policy

SME subsidized loans and public guarantees in Korea

▪ Public guarantee schemes guarantee from 85% to 100% of the credit

amount for private companies, with a cap set at KRW 3 billion.

▪ Subsidized loans follow the policy fund base rate which has consistently
been 1% to 2.2% lower than the average market interest rate from
2012/Q4 to 2017/Q4, with a limit set at KRW 4.5 billion in 2015.

▪ Both public guarantees and subsidized loans are equally important

sources of accessing credit for SMEs.

▪ The decision to provide a subsidized loan or public guarantee is based on

their application for funding, as to the feasibility of the applicant's
technological and business plans, management expertise, and growth
prospects.

35
 Policy

Selecting Suitable Targets for Government Credit Programs

▪ We propose using predicted sales growth as a key criterion to identify

suitable firms for these programs.

• it aligns with the objective of promoting growth of firms facing credit

constraints and
• can serve as key performance indicator (KPI) for evaluating the
impact of government support.

▪ The Korean Ministry of SMEs and Startups has explicitly designated sales
growth rates as the primary performance indicators in their annual
performance and accountability report.

▪ Banerjee and Duflo (2014) show that "if the firm is credit constrained, an
expansion of the availability of bank credit will lead to an increase in its
total outlay, output and profits, without any change in market borrowing."

36
 Related literature

Policy shift towards data-driven decision-making

▪ Berryhill et al. (2019)., Ubaldi et al. (2019), OECD (2019)

▪ ML applications in diverse areas: Fraud detection, crime prediction, traffic
management, public service optimization.

Impacts of subsidized loans or public credit guarantee schemes on sales

and employment

▪ Brown and Earle (2017), Bertoni, Martiand Reverte (2019), Hottenrott and
Richstein (2020)

Applying machine learning in resource allocation

▪ Andini et al. (2018): Italy's tax rebate program by targeting consumption

constrained households
▪ Kleinberg et al. (2018): Crime risk to enhance judges' bail decisions
▪ Sansone and Zhu (2021): Social security data to identify individuals at risk
of long-term income support
▪ Andini et al. (2022): Public credit guarantee programs by targeting
creditworthy and credit-constrained firms 37
 Machine Learning

▪ Machine learning, a representative methodology in artificial

intelligence (AI), is defined as an algorithm that helps prediction or
decision-making by using experiences or observations as training
data.

▪ Machine learning => supervised learning & unsupervised learning

▪ Supervised learning is based on a model that predicts the output (𝑌)

using one or more inputs (𝑋).

1. It estimates a model by using a part of the given data X and Y as a

training sample.

2. Predicts the result using the rest of the data as a test sample.

38
 Application

✓ Applied machine learning to find target of business support policies by

predicting firms’ growth rate

• I classified companies with high predicted growth potential and those with
low growth potential

• Then, using the classification for targeting beneficiaries, compared the

performance between the targeted beneficiary candidates and the
companies that were actually supported.

• Data:
1. KED DB: financial and basic information of firms (2010-2016)
2. SIMS DB: SME support information (2010-2015)

The model was predicted for each group by limiting the population to
companies with less than 6 years of experience in the analysis of support for
start-up companies, and companies with more than 7 years of experience in
the analysis of support for established companies.

39
40
 Results

Machine Learning Predictions and Actual Outcomes

▪ The figure shows the average growth rate of sales for both subsidized
firms and not subsidized firms across deciles of predicted growth.

41
 Results

Firm performance comparison by ML prediction target

▪ Panels display the average growth rates of sales and total assets,
respectively, for firms in the top 30% and bottom 70% based on the
predicted sales growth.

42
 Conclusion

AI for Smarter SME Policy Targeting

▪ By helping to screen out unfit firms and select recipients aligned with policy
goals, AI can greatly enhance the effectiveness of public support programs.

▪ We need a shift from the current support system (ex post evaluation of project
effectiveness) to a data-driven system where targets are selected based on
their characteristics.

▪ In a KDI survey in 2020, respondents pointed out that what hinders

transitioning into a system of AI-applied, data-driven policymaking in the
public sector are:

1. incomplete standardization and linkage of policy information between

governmental ministries

2. lack of expertise in technology utilization in the public sector.

43
 Policy recommendations

Toward Data-Driven Policymaking: Strategy

▪ To enhance policy effectiveness across sectors—education, healthcare, public

safety, national defense, and business support—a national-level strategy is
needed to promote data-driven policymaking.

▪ AI adoption in the public sector should begin by supporting capable public

institutions with tailored assistance in system building, data platforms,
security, consulting, and training.

▪ Key strategic actions include:

• Integrating dispersed data systems into a unified platform
• Reorganizing administrative processes for efficient policy coordination
• Establishing trust-based public-private partnerships

44
 Policy recommendations

Toward Data-Driven Policymaking: Action Plan

▪ The government must analyze the entire policy cycle—planning,

implementation, and evaluation—to identify where AI can add value.

• It is necessary to clearly recognize the areas where AI technology can

effectively help policy decision-making.

▪ A step-by-step approach is needed:

• Define clear objectives
• Select and share relevant data
• Build secure infrastructure
• Ensure transparent operations

▪ The final step is to apply AI-analyzed data to improve project planning and
operations.
• Monitoring real-world performance data from programs is essential.
• These insights should be fed back into the policy cycle to continuously
refine and enhance the effectiveness of ongoing and future programs.

45
 Discussion

▪ How do we actually select the policy targets?

• Just rely on AI?

• Type 1 error (false positive): Among the selected companies, an
unintended company (e.g., a company with not high growth rate) may
have been selected.
• Type 2 error (false negative): a desirable target is not selected.

• Model predicted only one-dimension (sales growth). Is this enough

information to support the company?
• the purpose of government support is to support a company that has
business or technical potential but lacks funding.

46