Towards a Standardized Business Process Model for LLMOps
Maria Chernigovskaya
1 a
, Damanpreet Singh Walia
2 b
, Ksenia Neumann
2 c
, Andr
´
e Hardt
1
,
Abdulrahman Nahhas
1 d
and Klaus Turowski
1
1
MRCC VLBA, Otto-von-Guericke University, Universitaetsplatz 2, Magdeburg, Germany
2
BIRD Lab, Otto-von-Guericke University, Universitaetsplatz 2, Magdeburg, Germany
Keywords:
LLM, Large Language Model Operations, LLMOps, Standardization, Process Model, BPMN.
Abstract:
The generalization and standardization of the Large Language Model Operations (LLMOps) life cycle is
crucial for the effective adoption and management of Large Language Models (LLMs) in a business context.
Researchers and practitioners propose various LLMOps processes however they all tend to lack formalization
in their design. In this paper, we address the absence of a standard LLMOps model for enterprises and
propose a generalized approach to adopting LLMOps into existing enterprise system landscapes. We start
by identifying the state-of-the-art LLMOps processes through a systematic literature review of peer-reviewed
research literature and gray literature. Considering the scarcity of relevant publications and research in the
area discovered during the initial stage of the research, we propose a generic, use-case-agnostic, and tool-
agnostic LLMOps business process model. The proposed model is designed using the Business Process Model
and Notation (BPMN) and aims to contribute to the effective adoption of LLM-powered applications in the
industry. To the best of our knowledge, this paper is the first attempt to systematically address the identified
research gap. The presented methods and proposed model constitute the initial stage of the research on the
topic and should be regarded as a starting point toward the standardization of the LLMOps process.
1 INTRODUCTION
Large Language Models (LLMs) have emerged as a
technological breakthrough in the field of Artificial
Intelligence (AI). Advanced neural network models
trained on vast amounts of data have demonstrated
human-like performance in complex natural language
processing tasks like sentiment analysis (Chen, 2024),
text generation (Liang et al., 2024), and translation
(He et al., 2023). Due to its remarkable capabili-
ties, enterprises began looking for ways to integrate
this powerful tool into their IT landscape and existing
business processes. To achieve effective and seamless
integration of LLMs, Large Language Model Opera-
tions (LLMOps) must be applied. LLMOps is a set of
engineering practices that emerged as an extension of
the Machine Learning Operations (MLOps) paradigm
to regulate the lifespan of LLM-powered applications
(Diaz-De-Arcaya et al., 2024). Standardization of
LLMOps may provide substantial advantages to en-
a
https://orcid.org/0009-0004-2763-6970
b
https://orcid.org//0009-0002-4044-5613
c
https://orcid.org/0000-0002-3713-8893
d
https://orcid.org/0000-0002-1019-3569
terprises seeking to efficiently deploy and manage
LLMs by ensuring homogeneity in deployment and
maintenance across various teams and projects.
The standardization and formalization of the LL-
MOps framework could be achieved conceptually and
later practically through various standard modeling
languages (e.g., Unified Modelling Language (UML),
Systems Modelling Language (SysML), etc.). How-
ever, to capture not only the technology part but also
its integration with the ongoing business processes,
the use of the Business Process Model and Nota-
tion (BPMN) becomes sensible, as it has been the
prevailing business process modeling language stan-
dard since its debut in 2004 (Resp
´
ıcio and Domin-
gos, 2015). Additionally, adopting BPMN practices
could contribute significantly to the standardization
of LLMOps by supplying a systematic framework for
deploying, fine-tuning and monitoring every step in-
volved in the LLM-powered application life cycle.
In this paper, we propose an use-case-agnostic and
tool-agnostic BPMN model to standardize the process
of adopting LLMOps into existing IT landscape. To
the best of our knowledge, this work is the first at-
tempt to address this problem systematically, while
856
Chernigovskaya, M., Walia, D. S., Neumann, K., Hardt, A., Nahhas, A. and Turowski, K.
Towards a Standardized Business Process Model for LLMOps.
DOI: 10.5220/0013377700003929
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 856-866
ISBN: 978-989-758-749-8; ISSN: 2184-4992
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

highlighting the state-of-the-art research findings. In
our work, we aim to answer the following Research
Questions (RQs):
RQ 1. What are the state-of-the-art LLMOps pro-
cesses proposed by academics and industry
practitioners?
RQ 2. How can a standardized LLMOps process
model be developed using identified state-of-
the-art processes and BPMN?
The main contribution of this paper is an easy-
to-adopt BPMN LLMOps process model for business
use. The resulting process model is an artifact, based
on the definition of artifact by (Hevner et al., 2004).
We derive use-case- and tool-agnostic processes from
the existing LLMOps processes being employed by
researchers and practitioners to ensure a compatible
process model to establish the foundation for subse-
quent works on standardizing the LLMOps life cy-
cle. Overall, we assert two claims: First, we demon-
strate the lack of a standard LLMOps process model
through a critical viewpoint on distilled information
from a literature review. Second, we propose LL-
MOps standardized process model that is generic and
ensures compatibility for easy adoption.
The paper proceeds as follows: in the subsequent
section 2 we systematically explore the state-of-the-
art LLMOps processes in the peer-reviewed and gray
literature with a goal to answer RQ 1.. The section
also provides a critical viewpoint on the shortlisted
literature, where the main limitations of the analyzed
works are discussed. In section 3 we present a con-
solidation of the identified in the literature LLMOps
and propose a design of a standardized LLMOps pro-
cess model, addressing RQ 2.. In the section 4 we
once again discuss the goal of the proposed work and
address its main limitations. We summarize the main
findings and contribution of this work and discuss fu-
ture research perspective in the section 5.
2 STATE-OF-THE-ART LLMOps
To the best of our knowledge, no study has been con-
ducted aiming to achieve the same research goal as
ours i.e. to propose a standardized process model
of LLMOps. To conduct the research systematically,
the method of Systematic Literature Review (SLR)
has been chosen, as suggested by (Okoli, 2015) and
(Kitchenham et al., 2004), as it provides a structured
approach of the relevant literature analysis and helps
to summarize its findings. Thus, a comprehensive
SLR has been performed in order to derive a LLMOps
standardized process model based on existing pro-
cesses and models proposed in the related research.
2.1 Literature Search
This section describes in a detail process of conduct-
ing an SLR based on the guide by (Okoli, 2015) on
the highlighted topic as well as its findings. The vi-
sual representation of the conducted SLR is depicted
in Figure 1 and it consists of five screening stages:
(0) retrieving articles based on a constructed out of
keywords search string from various databases, (1)
applying exclusion criteria, (2) reading title and ab-
stract, (3) reading full-text, (4) synthesizing peer-
reviewed sources with gray literature such as white
papers. Each stage consists of one or more steps and
displays remaining records in each stage. These steps
follow the subsequent execution order: selecting liter-
ature sources (scientific databases), defining a search
string, retrieving articles, removing duplicates, sev-
eral screening stages of the articles according to the
defined exclusion and inclusion criteria, enriching the
intermediate results by extending our SLR through
forward and backward search and gray literature.
The first step of conducting an SLR is to define
a search string that covers the first research question.
The constructed query is based on the identified key-
words and formulated as following: ((”LLM*” OR
”Large language model*”) AND (”process model*”
OR ”Operation*”)) OR (”LLMOps” OR ”LLM-OPs”
OR ”Large language model* operation*”). A broad
set of databases is chosen in order to ensure a high
level of inclusion. Query search was performed on the
9th of October, 2024 over thirteen databases: ACM,
Springer Link, Scopus, IEEE Xplore, ScienceDirect,
Web of Science, AlSel, Scitepress, mdpi, Wiley, Tay-
lor&Francis, Emerald Insight and Sage, accumulating
to 43 initially retrieved articles.
Our exclusion criteria: complete duplicates, not
peer-reviewed, closed access book, language other
than English, title / abstract / full-text not related as
per PRISMA list (Tricco et al., 2018). The complete
overview of the search process including yielded re-
sults is depicted by Figure 1. Firstly, out of the 43
results eight duplicates were removed. Out of the re-
maining 35 four were removed as they were not peer-
reviewed and eight due to original source being in
closed access. Followed by the filtration stages of ti-
tle and abstract relevance, 17 records were removed,
resulting into six full-text-related results. As the final
number of records happened to be too small, we de-
cided to extend the peer-reviewed literature through
forward and backward search (two additional sources
found), white papers (accumulating five), three books,
Towards a Standardized Business Process Model for LLMOps
857

Identification
of studies via
databases
   Records identified from:
   Databases (n = 43)
 ACM,n = 6
 Springer Link,n = 17
 Scopus, n = 4
 IEEE Xplore,n = 5
 ScienceDirect,n = 1
 Web of Science, n = 1
 AlSel,n = 0
 Scitepress,n = 0
 mdpi,n = 4
 Wiley,n = 0
 Taylor &Francis,n = 0
 Emerald Insight,n = 8
 Sage,n = 0
Stage 0: Retrieved articles
Records assessed for eligibility
(n = 35)
Stage 1: Exclusion criteria
Stage 2:
Read title and abstract
Stage 3:
Read full-text
Stage 4:
Synthesis
Records screened (n = 23) Records screened (n = 6)
Studies included in synthesis
(n = 17)
Not peer-reviewed records
identified from:
     Forward Citation search
     (n = 2)
     White papers (n = 5)
     Books (n = 3)
     Arxiv (n = 1)
Records excluded:
Full text not related (n = 0)
Records excluded:
Title & abstract not related
(n = 17)
Records excluded:
Not peer-reviewed (n = 4)
Closed access book (n = 8)
Language not English (n = 0)
Records removed before
screening:
Duplicates removed (n = 8)
Identification of studies
via other sources
Figure 1: Synthesis of the conducted SLR process with the corresponding screening stages (Stage 0 - Stage 4).
and a single article from Arxiv. White papers were
chosen based on the market share of the publish-
ing company selecting following prominent compa-
nies such as AWS, Amazon Web Services, Fujitsu,
AMD epyc, Dell Technologies, Nvidia, appliedAI,
salesforce, RedHat and Intel. However, after the con-
ducted SLR, the final set of relevant white papers was
reduced to five (Fujitsu, 2024), (Basak, 2024), (Kar-
takis and Hotz, 2023), (Datta et al., 2024), (Venkata-
pathy, 2023). Some leading LLM vendors (e.g., Open
AI
1
, Google Gemini
2
, etc.) were also considered in
the initial stages of the SLR. However, these compa-
nies were not included in the final set as they tend to
focus more on the products’ performance rather than
on their integration into existing business processes.
2.2 Literature Synthesis
In order to combine our findings, a thematic synthesis
as proposed by (Cruzes and Dyba, 2011) was applied.
Their multi-step approach helps to identify recurring
themes from multiple studies, analyze and interpret
them, so the conclusions can be drawn in the system-
atic reviews. In other words, it synthesizes the find-
ings identified within primary studies. After the SLR
was completed and full-text analysis of the final re-
sults was conducted, four main groups aka ”themes”
were identified within the final set of studies. The
proposed categorization consists of shortlisted litera-
ture from Stage 4 of SLR and is based on RQ 1. to
identify state-of-the-art LLMOps processes with ad-
ditional categories for LLMOps tool(s), LLMOps uti-
lization and LLMOps related actors. The overview of
all four groups can be found in Table 1.
1
https://openai.com/[03.02.2025]
2
https://gemini.google.com/[Accessed on 03.02.2025]
The first group consists of thirteen studies and
describes either LLMOps processes or their stages.
The second group is assigned to category ”LLMOps
tool(s)” and contains studies, in which authors pre-
sented or discussed at least one LLMOps tool. In the
third identified category, all studies that describe LL-
MOps utilization, were grouped. Finally, the fourth
category, where the authors described LLMOps actors
and their responsibilities, concludes Table 1 with just
two studies. It it worth mentioning, that some of the
publications fall under multiple thematic categories as
they cover multiple topics at the same time.
2.3 Critical Viewpoint
To evaluate the quality of the conducted review, we
present our perspectives on the literature within the
“LLMOps processes” cluster. From the perspective
of AWS, one of the leading cloud service providers,
(Kartakis and Hotz, 2023) highlights the distinc-
tion between MLOps, Foundational Model Opera-
tions, and LLMOps. The authors also detail the pro-
cesses, actors, and tools involved in LLMOps, pro-
viding practitioners with a comprehensive framework
that encompasses all critical dimensions relevant to
businesses. Expanding on this work, (Basak, 2024)
presents a guide for the operation and management
of LLMs using Apache Airflow. While both works
effectively delineate the various processes, roles, and
responsibilities associated with LLMOps, they do not
provide detailed work and adoption analysis is also
missing. Furthermore, the literature demonstrates a
noticeable preference for tools within AWS’s service
catalog, such as Amazon Managed Workflows for
Apache Airflow.
The white paper by TruEra and Intel (Datta et al.,
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
858

Table 1: Categorization of the identified studies based on the thematic category.
Category Literature
LLMOps processes (Basak, 2024), (Datta et al., 2024), (Diaz-De-Arcaya et al., 2024),
(Huang et al., 2024), (Kartakis and Hotz, 2023), (Kamath et al.,
2024), (Park et al., 2024), (Parnin et al., 2023), (Reddy et al.,
2024), (Roychowdhury, 2024), (Shan and Shan, 2024), (Venkata-
pathy, 2023)
LLMOps tool(s) (Arawjo et al., 2024), (Basak, 2024), (Datta et al., 2024), (Fujitsu,
2024), (Huang et al., 2024), (Kamath et al., 2024), (Kartakis and
Hotz, 2023), (Park et al., 2024), (Venkatapathy, 2023), (Wang and
Zhao, 2024)
LLMOps utilization (Chen, 2024), (Kamath et al., 2024), (Parnin et al., 2023), (Shan and
Shan, 2024), (Shi et al., 2024), (Venkatapathy, 2023)
LLMOps actors and responsibilities (Basak, 2024), (Kartakis and Hotz, 2023)
2024) provides a detailed exploration of the LLMOps
workflow, consolidated based on the emerging LL-
MOps technology stack available in the market. The
authors acknowledge the limited listed technology
stack while emphasizing that the workflow’s develop-
ment, grounded in these emerging tools, enhances its
validity and achieves tool-agnosticism by encompass-
ing a variety of technologies. Furthermore, the white
paper’s broad focus on LLM use cases adds to its gen-
eralization and potential for wider adoption. How-
ever, the workflow presented is predominantly tech-
nical in nature, highlighting a notable absence of gen-
eralized approach and relevant strategic processes.
The authors of (Diaz-De-Arcaya et al., 2024) con-
solidate the definition of LLMOps and its various
stages through a synthesis of existing literature. Their
unified definition emphasizes the need for LLMOps,
presenting it as a customized MLOps approach tai-
lored to address business challenges such as cost man-
agement, technical hurdles, and the selection of tools
and infrastructure. The authors also identify and com-
prehensively define the stages of LLMOps. However,
the absence of a visual model and detailed informa-
tion on the sequence of stages is a notable limitation.
While the stage numbering implies a sequential flow,
this is not explicitly clarified. Furthermore, the cor-
relation between the defined stages and the strategic
challenges faced by businesses mentioned in the LL-
MOps definition appears incomplete, as these chal-
lenges are not adequately reflected in the stages. Al-
though the model presented is sufficiently general for
LLM use cases, the lack of detailed information on
tools makes it challenging to assess its tool-agnostic
nature.
A shortlisted book, (Huang et al., 2024), provides
an explanation of LLMOps, emphasizing its needs
and benefits. It effectively highlights the differences
between MLOps and LLMOps. Regarding LLMOps
processes, the book focuses on and elaborates upon
processes related to LLM security. Additionally, it
lacks a visual representation of a process model. An-
other shortlisted book, (Kamath et al., 2024), dis-
cusses certain aspects of LLMOps and presents an ex-
ample of an LLMOps architecture for an LLM-based
chatbot. However, the proposed architecture is com-
plex and overly detailed, and its lack of compliance
with standard modeling notations makes it challeng-
ing to interpret at a strategic organizational level.
(Park et al., 2024) introduces an LLMOps
pipeline, termed “LlamaDuo”, designed for migra-
tion from service-oriented LLMs to smaller, locally
manageable models. While the proposed pipeline
shares some similarities with other shortlisted LL-
MOps processes, its objectives differ from the in-
tended focus of this work. In (Parnin et al., 2023), the
authors gathered challenges and workflows for devel-
oping copilot-like products through interviews with
professional software engineers. The iterative process
model outlined in the study describes all underlying
processes by incorporating the perspectives of these
practitioners. Although the process model does not
employ any standard modeling notation, it effectively
provides a comprehensive representation of LLMOps
processes tailored to this specific use case.
The causes and impacts of hallucinations in LLMs
are examined in (Reddy et al., 2024), along with
strategies to mitigate this issue. While the article also
presents an LLMOps process model, it lacks a clear
explanation of its origin and underlying rationale.
This drawback limits its applicability and leads to the
exclusion of this process model from further consider-
ation in this study. The short article (Roychowdhury,
2024) proposes a three-stage LLMOps process model
for finance-focused LLM products, starting with the
definition of the business case. However, as the pro-
cesses are presented in an abstract manner and lack
Towards a Standardized Business Process Model for LLMOps
859

detailed explanations, it is challenging to achieve a
comprehensive understanding of the model.
The research article (Shan and Shan, 2024) intro-
duces the 4D LLMOps process model, outlining best
practices and application scenarios for implementing
LLMOps in enterprises. The authors also provide a
list of potential tool stacks that can be utilized within
the proposed framework. However, the rationale be-
hind the naming of the stages is unclear, as the terms
do not maintain a consistent level of abstraction. For
instance, “Deploy” is a fundamental operational task,
while “Deliver” encompasses multiple operational ac-
tivities. Furthermore, the model does not adhere to
standard modeling notations, which may limit its clar-
ity and adoption. The white paper by Dell Technolo-
gies (Venkatapathy, 2023) provides a comprehensive
outline of the validated life cycle design for generative
AI in retail use cases. The author thoroughly explains
all elements involved in the design and presents a tech
stack based on NVIDIA technologies. However, the
life cycle lacks adherence to standard modeling no-
tations, which hinders its simplification and broader
adoption.
In summary, the definition of LLMOps pro-
vided by (Diaz-De-Arcaya et al., 2024) establishes
the criteria for assessing the completeness of LL-
MOps lifecycle models. Among the twelve reviewed
sources on LLMOps processes, only seven—(Basak,
2024), (Datta et al., 2024), (Diaz-De-Arcaya et al.,
2024), (Kartakis and Hotz, 2023), (Parnin et al.,
2023), (Roychowdhury, 2024), and (Shan and Shan,
2024)—present well-defined LLMOps models that
align with the focus of this research. The absence of
a comprehensive LLMOps process model with stan-
dardized notation and consistent terminology remains
evident. A detailed assessment of the identified mod-
els for standardization is provided in the following
section.
3 LLMOPS STANDARDIZED
PROCESS MODEL
To assess the level of standardization in LLMOps
business process models, we performed a subjective
evaluation of their completeness and generality due to
the absence of a predefined set of criteria. For com-
pleteness, we evaluated whether the LLMOps models
incorporate both strategic and technical aspects of the
LLMOps life cycle, as defined in (Diaz-De-Arcaya
et al., 2024), and whether they effectively address
key business challenges. For generality, we examined
three key factors:
• Use of Standard Modeling Notation. To ensure
ease of interpretation and clarity.
• Use-Case-Agnostic. To facilitate adoption across
a diverse range of LLM use cases.
• Tool-Agnostic. To support unbiased adoption by
accommodating a wide variety of available LL-
MOps tools.
This approach allows us to systematically evaluate
the strengths and limitations of the proposed models.
The presented comparison in Table 2 aims to as-
sess generality of the identified state-of-the-art LL-
MOps models from the literature. Here, columns rep-
resent generality criterion: Use of Standard Notation,
Use-Case-Agnostic and Tool-Agnostic. Literature ci-
tation is used as identifier for identified LLMOps
models. ”+” and ”−” used as marking scheme show-
ing compliance and non-compliance respectively. As
it can be observed, none of the identified publica-
tions presents a fully standardized approach for gen-
eralizing LLMOps. Additionally, none of the re-
viewed literature utilized BPMN in their proposed
artifacts. Regarding the completeness of LLMOps
models, (Parnin et al., 2023) and (Shan and Shan,
2024) offer the most comprehensive process models
compared to other identified approaches.
3.1 Redesigning LLMOps Life Cycle
The seven shortlisted LLMOps models were assessed
against the aforementioned criteria for standardizing
LLMOps process models. However, none of the iden-
tified models fully met the criteria. Consequently, a
method was devised to redesign the LLMOps life cy-
cle from the identified models, resulting in a standard-
ized LLMOps process model.
Firstly, process models from the shortlisted liter-
ature were consolidated, as illustrated in Figure 2, to
redefine processes into consistent process group cate-
gories (as shown in legend) to meet the requirement
of RQ 2., generic LLMOps process model artifact.
Here, the proposed process group’s color is used to
mark processes of models based on the responsible
activity areas. This classification aids businesses in
identifying relevant roles for these activities.
Secondly, six process groups were identified
within the shortlisted models, providing a high-level
overview of the LLMOps life cycle:
1. Strategic. Processes focused on defining and for-
mulating business needs.
2. Data. Processes related to data collection, pro-
cessing, transformation, and management.
3. LLM. Processes centered on the selection, fine-
tuning, and evaluation of LLMs.
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
860

Table 2: Comparison table to assess the generality of the identified state-of-the-art LLMOps models from the literature.
Literature
Generic LLMOps Model
Use of Standard Notation Use-Case-Agnostic Tool-Agnostic
(Basak, 2024) − + −
(Datta et al., 2024) − + +
(Diaz-De-Arcaya et al., 2024) − + −
(Kartakis and Hotz, 2023) − + −
(Parnin et al., 2023) − − −
(Roychowdhury, 2024) − − −
(Shan and Shan, 2024) − + +
Data LLM Dev
OPs
Compliance
Strategic
Legend for Process Groups:
(Datta et al., 2024)
Prompt
Engineering
Test & Test Prompt
Lineage
Chain Prompts &
Applications
Input / Output
Filtering &
Guardrails
Rating
Mechanisms
Develop & Deploy
Web application
Select and Fine-
tune or Use Fine-
tuned Model
Test Functionality
New Test Set
Input / Output &
Rating Interaction
Create User Profile
& Share Data
Backend Frontend
(Diaz-De-Arcaya et al.,
2024)
(Kartakis and Hotz, 2023)
(Basak, 2024)
(Parnin et al., 2023)
(Roychowdhury, 2024)
(Shan and Shan, 2024)
Data Preparation
Vector Database
Index
Construction
Model Fine Tuning App Creation
Prompt
Engineering,
Tracking,
Collaboration
Evaluation and
Debugging
Model Training
Model Deployment
and Inference
App Hosting Monitoring
Model Selection Evaluation Deployment Monitoring
Data Privacy &
Security
Data Management Ethics & Fairness
Distill
Start
Adaptation
DeliverDeployDiscover
Exploration Implementation Evaluation Productization
The
Prototype
Scaling Evolution
Figure 2: Consolidated view of the identified LLMOps models with colored areas of responsibilities.
4. Development (Dev). Processes involving the in-
tegration of LLMs into system infrastructures.
5. Operations (Ops). Processes responsible for de-
livering and maintaining the product to end users.
6. Compliance. Processes ensuring periodic audits
and adherence to ethics, fairness, data privacy, and
security.
Thirdly, these process groups and their underly-
ing processes are abstractly represented in Figure 3,
which depicts an infinite knot diagram of LLMOps
process model as per aforementioned redefined pro-
cess groups on an abstract level. Color consistency
is maintained with Figure 2. This diagram presents
the sequence of processes in the LLMOps life cycle,
drawing inspiration from the MLOps model outlined
in the following ML4Devs article
3
. Lastly in the sub-
sequent subsection, this LLMOps process model is
further formalized and standardized using BPMN.
3
https://www.ml4devs.com/articles/mlops-machine-
learning-life-cycle/ [Accessed on 14.12.2024]
Data
LLM
Dev
OPs
Compliance
Strategic
Figure 3: Infinite loop representation of LLMOps process
model derived from Figure 2.
3.2 LLMOps BPMN Model
The BPMN model depicted in Figure 4 is the cen-
tral artifact of this paper. It aims to generalize
and standardize the steps required to effectively in-
tegrate LLMOps into various business scenarios. The
model summarizes the findings of analyzed scientific
and gray literature and reflects the consensus among
Towards a Standardized Business Process Model for LLMOps
861

academia and industry with respect to the LLMOps
workflow definition. Finally, the presented artifact
aims to highlight the use-case and tool-agnostic na-
ture of the proposed LLMOps representation.
The presented model consists of one main pool
and a black box pool that depicts an IT infrastructure.
The main pool of the model describes the complete
LLMOps workflow within the organization divided
into two lanes, each illustrating a distinct functional-
ity level: strategic and operational. Each defined level
is represented by a team that performs the respective
roles. The strategic team consists of key decision-
makers in charge of defining goals and objectives, de-
veloping strategies, and supervising the overall execu-
tion. The operational team is responsible for carrying
out day-to-day tasks that align with the business strat-
egy. In this scenario, we consider a development team
to be a part of the operational team as it is responsi-
ble for performing tasks that are operational in nature.
Depending on the size and type of the enterprise, the
exact roles and activities of both teams may differ.
The decision to illustrate the IT infrastructure as
a black box entity is motivated by two main reasons.
Firstly, viewing IT infrastructure as a black box en-
ables our model to stay general enough and adaptable
to various infrastructure configurations (on-premises,
cloud, hybrid, etc.). Secondly, the goal of the model is
to outline key touch points between two pools and the
associated impact of the infrastructure without delv-
ing into its internal activities and technicalities as both
will strongly depend on the specific technology stack
and their vendors.
The process starts with the launch of the LLMOps
project and an outline of its general scope and ob-
jectives. Here, the desired outcome of the project,
that depends on the application of an LLM, must be
clearly defined. It could, for example, be targeting the
enhancement of the customer experience, automation
of tasks, improvement of decision-making, etc. De-
fined goals are subsequently translated into business
requirements, which typically comprise objectives,
benefits, constraints, value propositions, stakeholder
expectations, and use case descriptions. The defined
business requirements are then passed onto the oper-
ational team to derive functional requirements. Func-
tional requirements are technical specifications that
engineers and developers must follow to implement
the expected functionalities, while aligning with busi-
ness and non-functional requirements. Requirement
engineering is an important activity, especially with
regard to complex systems, as it guarantees that all
parties involved have a shared understanding of the
project objectives and expectations.
Once functional and non-functional requirements
are defined, the operational team proceeds with the
design of artifact architecture. The architecture de-
sign is a foundational step, that lays out a blueprint
of how exactly the application will be integrated into
the existing system. It outlines the application’s high-
level structure and defines its main components, their
functions, and connections with one another. The pro-
posed architecture is passed onto the strategical team
to evaluate its alignment with the defined objectives.
If the proposed architecture is approved, the blueprint
is finalized and passed on to development activities.
If the provided architecture does not pass the approval
step, it is subjected to further refinements until it sat-
isfies the set requirements.
In alignment with the strategic workflow, the oper-
ational team proceeds with the activities focusing on
the data management and preparation. At this step,
the relevant data is collected, cleaned, normalized,
and transformed to the form suitable for usage within
an LLM model. Additionally, the data quality evalu-
ation process is performed. Data preparation step is
critical as it ensures quality, validity, and usefulness
of the used data throughout the life cycle. Following
this step, the operational team selects a suitable pre-
trained LLM model that best meets the functional and
non-functional requirements and is compatible with
the business scenario. At this step, factors like model
size, model type, costs (e.g., personnel, licensing, in-
frastructure), and technical requirements must be con-
sidered. Within this work we assume that a typical en-
terprise, due to the lack of expertise, high cost and sig-
nificant computational complexity, would rather not
invest in training their own LLM model and rather use
a pre-trained one.
After selecting the base model, the decision to
fine-tune LLM is made. A fine-tuning phase is re-
quired to tailor the model for a particular type of task
and data. As shown in the proposed BPMN model,
the fine-tuning stage can be skipped if the base model
already performs sufficiently to serve the intended
use. At this point, the strategic team activities (e.g.,
finalization of the project blueprints) integrate with
the operational team’s workflow, with a touch point
right before the artifact development activity. Artifact
development is a transformational phase that merges
the approved architecture design and requirements to
produce a functional LLM artifact integrated with all
its components. Such components might include vari-
ous APIs, databases, user interfaces, embedding tools,
etc.
The subsequent evaluation stage examines
whether the produced artifact fulfills the business,
functional, and non-functional requirements before
proceeding to the deployment phase. The assessment
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
862

InfrastructureLLMOps
yes
Strategical Team
Definition of
scope and
objectives
Business
requirements
engineering
Initiation of the project
Finalized
blueprints
Business Data
no
Bussiness
requirements
are met?
Adjust or retire
the artifact?
retire
Project terminated
Business value
assessment
Artifact architecture
approved?
yes
no
adjust
Data
transformation &
synchronization
Approval of
artifact
architecture
Initial
blueprints
Operational Team
Functional
requirements
engineering
Definition of the
artifact
architecture
Data preparation
& Management
Base Model
selection
Functional &
non-functional
requirements
yes
Fine-tuning
required?
Model
fine-tuning
Artifact
development
Evaluation
Deployment
no
Artifact works
as intended ?
Artifact
integration
Monitoring &
Maintanance
Lifecycle assessment
Artifact
Retirement
Artifact retired
Vector Database
incident
occured
no
yes
Figure 4: Standardized and generalized design of LLMOps process model using BPMN 2.0.
Towards a Standardized Business Process Model for LLMOps
863

process might involve artifact performance testing
(e.g., response time and throughput evaluation),
output accuracy, etc. The evaluation process consists
of several cycles of the artifact assessments and cul-
minates in two possible outcomes: a verified artifact
suitable for deployment or additional refining if inad-
equacies or inconsistencies with the requirements are
discovered. If the artifact fails to perform as intended,
the root cause has to be identified in the preceding
activities, with the worst-case scenario necessitating
the functional requirements to be refined.
After being evaluated, the artifact is deployed into
the company’s infrastructure. This step ensures that
the artifact is technically functional and effectively
integrates with running other applications, databases,
and daily workflows. It usually involves setting up the
software environment, allocating the required com-
puting resources, and carrying out limited live tests
to observe its behavior in real-world settings. The
follow-up artifact integration activity is responsible
for seamlessly integrating the artifact with the enter-
prise’s current IT systems and processes. It includes
integration with standard enterprise IT systems (e.g.,
Enterprise resource planning (ERP), Customer Rela-
tionship Management (CRM)), data synchronization
between databases used for day-to-day business ac-
tivities and the databases containing representations
suitable for an LLM (e.g., vector storage), general
workflow alignment, user interface integration, and
data privacy compliance validation.
The following phase is the monitoring and main-
tenance of the deployed and integrated into the pro-
duction environment artifact. This stage consists of
two distinct activities with different sets of respon-
sibilities. Monitoring involves tracking the artifact’s
activity in real-time to ensure the alignment of the ar-
tifact’s behavior with the requirements defined in the
earlier stages of the project. It comprises tasks such as
performance monitoring, output quality assessment,
anomaly and error detection, log analysis, and ensures
continuous compliance with the data security and pri-
vacy regulations. Additionally, continuous mainte-
nance of the software (e.g., security updates) and in-
frastructure components ensure that the artifact stays
functional, manages incidents, and adheres to contin-
uously evolving data and production environment re-
alities (e.g., hardware generation changes and licens-
ing changes). This phase includes activities such as
regular artifact updates and correcting any potential
errors detected during monitoring.
Once the artifact is integrated, its life cycle must
be evaluated on a regular basis to ensure that it con-
tinues to provide the intended value while being rel-
evant to evolving business requirements, significant
technology developments (e.g., new model types, ef-
ficiency improvements), and user feedback. The life
cycle assessment is additionally supported by a busi-
ness value assessment, which serves as a systematic
method applied to establish the project’s continuous
viability. This approach assists the strategic team in
decision making on the project’s extension, adjust-
ments, or termination depending on the outcomes of
the assessment.
If the artifact proves its practicality, it continues
to operate until the next assessment cycle, or it might
be determined that it requires specific adjustments to
meet business requirements. If, based on the eval-
uation performed, the artifact is considered outdated
or no longer valuable, it is retired accordingly, with
all its dependencies being transitioned, archived, and
finally decommissioned from the production environ-
ments. The exact decommissioning and archival pro-
cedures strongly depend on the nature of the project
as well as the compliance regulations it falls under, if
any. The project termination process is initiated con-
currently with the artifact retirement.
4 DISCUSSION
Based on the key findings of the research, we can con-
clude that the relevant literature for the defined prob-
lem is notably sparse. Furthermore, none of the ana-
lyzed sources proposed a formalized or standardized
depiction of LLMOps processes with the underlying
activities. The majority of proposed representations
in academic literature were predominantly broad and
theoretical. Conversely, in industrial literature (white
papers), the presented models were rather specific and
tailored for the particular technological stacks. As a
result, there was a lack of consistency and formaliza-
tion in the LLMOps presentation in the collected lit-
erature.
To address the lack of consistency and generaliza-
tion, we sought to consolidate the identified variations
of LLMOps processes from academic and industrial
fields into a single formalized representation. By ana-
lyzing the presented LLMOps processes, we selected
stages of the process on which numerous authors ap-
peared to have a consensus and consolidated them in
Figure 2 and Figure 3. Following this approach, we
built a research foundation for our business process
model depicted in Figure 4.
The resulting business process model is the re-
search artifact of this work that presents the main
stages and activities required for LLMOps integra-
tion. We utilized the standard modeling notation like
BPMN and formalized LLMOps workflow based on
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
864

the literature findings to achieve a high degree of gen-
eralization of the proposed model. By this approach
we try to ensure its compatibility for easier adoption
in the business contexts.
4.1 Limitations
Although the proposed artifact is based on an exten-
sive and consistent literature analysis and designed
using a standard modeling language, we acknowledge
that we can not yet claim an overall standardization
and generalization. To reach this point, a number of
real-world use cases must be applied and tested to
prove the proposed model’s practical validity. We rec-
ognize this limitation and aim to address it in our fu-
ture work.
5 CONCLUSION AND OUTLOOK
Successful integration of LLMs into an enterprise
presents numerous challenges and necessitates the
adoption of LLMOps. Standardizing LLMOps prac-
tices may provide significant advantages to businesses
seeking to effectively manage the LLM life cycle
while ensuring workflow consistency across various
teams and projects. In this paper, we conducted an ex-
tensive SLR on the existing standard LLMOps mod-
els in both academic and industrial literature. Fur-
thermore, we consolidated an overview of identified
LLMOps processes and designed a cyclic LLMOps
representation in Figure 2 and Figure 3 as a visual
summary of the analyzed literature. Both figures are
utilized as a trustworthy research foundation for our
artifact.
Based on the SLR findings, we concluded that
none of the discovered paper proposed a formal-
ization or standardization model for LLMOps adop-
tion. Moreover, none of the obtained findings relied
on some standard modeling languages (e.g., BPMN,
Unified Modeling Language). To address the iden-
tified research gap, we proposed an use-case and
tool-agnostic LLMOps business model designed us-
ing BPMN. The model is designed to formalize and
standardize the main steps required to effectively in-
tegrate LLMOps into various business scenarios. The
conducted SLR and obtained findings answer RQ 1..
The proposed BPMN model in Figure 4 addresses and
answers RQ 2..
As discussed previously, in the subsection 4.1, the
presented work is subject to certain limitations that we
intend to tackle in our future work. Firstly, we aim to
test the proposed artifact on the real-world use cases
from various domains to prove its generalization. The
discovered findings might lead to further adjustments
and refinements of the presented model. Secondly, we
intend to elaborate more on specific type of activities
and the related tasks and roles. Therefore, we believe
that the presented results and artifact should not be re-
garded as final but rather considered as starting point.
REFERENCES
Arawjo, I., Swoopes, C., Vaithilingam, P., Wattenberg, M.,
and Glassman, E. L. (2024). ChainForge: A Visual
Toolkit for Prompt Engineering and LLM Hypothesis
Testing. In Mueller, F. F., Kyburz, P., Williamson,
J. R., Sas, C., Wilson, M. L., Dugas, P. T., and
Shklovski, I., editors, Proceedings of the CHI Confer-
ence on Human Factors in Computing Systems, ACM
Digital Library, pages 1–18. Association for Comput-
ing Machinery.
Basak, P. (2024). Unlocking fmops/llmops with airflow:
A guide to operationalizing and managing large lan-
guage models. Technical report, Airflow Summit
2024.
Chen, T. (2024). Challenges and opportunities in integrat-
ing llms into continuous integration/continuous de-
ployment (ci/cd) pipelines. In 2024 5th International
Seminar on Artificial Intelligence, Networking and In-
formation Technology (AINIT), pages 364–367.
Cruzes, D. S. and Dyba, T. (2011). Recommended steps for
thematic synthesis in software engineering. In 2011
international symposium on empirical software engi-
neering and measurement, pages 275–284. IEEE.
Datta, A., Sen, S., and Bandyopadhyay, A. (2024). Llmops
explained. Technical report, TruEra and Intel Corpo-
ration.
Diaz-De-Arcaya, J., L
´
opez-De-Armentia, J., Mi
˜
n
´
on, R.,
Ojanguren, I. L., and Torre-Bastida, A. I. (2024).
Large language model operations (llmops): Defini-
tion, challenges, and lifecycle management. In 2024
9th International Conference on Smart and Sustain-
able Technologies (SpliTech), pages 1–4. IEEE.
Fujitsu (2024). Fujitsu composite ai. Technical report, Fu-
jitsu Limited.
He, Z., Liang, T., Jiao, W., Zhang, Z., Yang, Y., Wang,
R., Tu, Z., Shi, S., and Wang, X. (2023). Exploring
human-like translation strategy with large language
models. arXiv preprint arXiv:2305.04118.
Hevner, A. R., March, S. T., Park, J., and Ram, S. (2004).
Design science in information systems research. MIS
quarterly, pages 75–105.
Huang, K., Wang, Y., Goertzel, B., Li, Y., Wright, S., and
Ponnapalli, J. (2024). Generative AI Security: Theo-
ries and Practices. Springer.
Kamath, U., Keenan, K., Somers, G., and Sorenson, S.
(2024). Large Language Models: A Deep Dive.
Springer.
Kartakis, S. and Hotz, H. (2023). Fmops/llmops: Opera-
tionalise generative ai using mlops principles. Techni-
cal report, Amazon Web Services Inc.
Towards a Standardized Business Process Model for LLMOps
865

Kitchenham, B. A., Dyba, T., and Jorgensen, M. (2004).
Evidence-based software engineering. In Proceed-
ings. 26th International Conference on Software En-
gineering, pages 273–281. IEEE.
Liang, X., Wang, H., Wang, Y., Song, S., Yang, J., Niu, S.,
Hu, J., Liu, D., Yao, S., Xiong, F., and Li, Z. (2024).
Controllable text generation for large language mod-
els: A survey. arXiv preprint arXiv:2408.12599.
Okoli, C. (2015). A Guide to Conducting a Standalone Sys-
tematic Literature Review. Communications of the As-
sociation for Information Systems, 37.
Park, C., Jiang, J., Wang, F., Paul, S., Tang, J., and Kim,
S. (2024). Llamaduo: Llmops pipeline for seamless
migration from service llms to small-scale local llms.
arXiv preprint arXiv:2408.13467.
Parnin, C., Soares, G., Pandita, R., Gulwani, S., Rich, J.,
and Henley, A. Z. (2023). Building your own product
copilot: Challenges, opportunities, and needs. arXiv
preprint arXiv:2312.14231.
Reddy, G. P., Kumar, Y. P., and Prakash, K. P. (2024). Hal-
lucinations in large language models (llms). In 2024
IEEE Open Conference of Electrical, Electronic and
Information Sciences (eStream), pages 1–6. IEEE.
Resp
´
ıcio, A. and Domingos, D. (2015). Reliability of
bpmn business processes. Procedia Computer Sci-
ence, 64:643–650.
Roychowdhury, S. (2024). Journey of hallucination-
minimized generative ai solutions for financial deci-
sion makers. In Proceedings of the 17th ACM Inter-
national Conference on Web Search and Data Mining,
pages 1180–1181.
Shan, R. and Shan, T. (2024). Enterprise llmops: Advancing
large language models operations practice. In 2024
IEEE Cloud Summit, pages 143–148. IEEE Computer
Society.
Shi, C., Liang, P., Wu, Y., Zhan, T., and Jin, Z. (2024).
Maximizing user experience with llmops-driven per-
sonalized recommendation systems. arXiv preprint
arXiv:2404.00903.
Tricco, A. C., Lillie, E., Zarin, W., O’Brien, K. K.,
Colquhoun, H., Levac, D., Moher, D., Peters, M. D.,
Horsley, T., Weeks, L., et al. (2018). Prisma extension
for scoping reviews (prisma-scr): checklist and expla-
nation. Annals of internal medicine, 169(7):467–473.
Venkatapathy, C. (2023). Genai for retail. Technical report,
Dell Technologies.
Wang, L. and Zhao, J. (2024). Strategic Blueprint for En-
terprise Analytics. Springer.
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
866