KPI2KVI: A Multi Agent Workflow for Calculating Key Value Indicators from Service Descriptions

# KPI2KVI: A Multi Agent Workflow for Calculating Key Value Indicators from Service Descriptions
Source: [https://arxiv.org/html/2605.22825](https://arxiv.org/html/2605.22825)
###### Abstract

Key Value Indicators \(KVIs\) provide a decision oriented view of a service by summarizing how operational performance translates into stakeholder value, risk, and outcomes\. However, in many domains KVIs are difficult to compute in practice because they require selecting relevant KVI categories, defining measurable Key Performance Indicators \(KPIs\), collecting KPI values, and applying consistent calculation logic, all of which is typically performed manually and inconsistently from unstructured service documentation\. This paper presents KPI2KVI, a tool that transforms a natural language service description into computed KVI estimates by orchestrating a deterministic multi agent workflow powered by Large Language Models \(LLMs\) that \(i\) elicits missing service context, \(ii\) extracts and finalizes relevant KVI categories from a taxonomy, \(iii\) generates service specific KPIs with units and descriptions, \(iv\) collects KPI values through an interactive dialogue and also supports intelligent estimation for KPI values that are unavailable, and \(v\) computes interval valued KVI outputs \(minimum, exact, maximum\) with traceable explanations for each KVI code\. Simulations with representative service descriptions demonstrate that KPI2KVI consistently produces a complete end to end mapping from description to KVI intervals and provides transparent calculation narratives that support post hoc auditing and interactive advisory queries\.

## IIntroduction

The Sixth Generation vision increasingly frames networks as critical societal infrastructure, expected not only to deliver advanced capabilities but also to contribute to long\-term objectives such as sustainability, inclusion, resilience, and trust\[[22](https://arxiv.org/html/2605.22825#bib.bib11),[21](https://arxiv.org/html/2605.22825#bib.bib7),[20](https://arxiv.org/html/2605.22825#bib.bib16),[5](https://arxiv.org/html/2605.22825#bib.bib22)\]\. As these expectations shape research agendas, governance, and procurement, stakeholders need ways to*demonstrate*,*compare*, and*audit*the value impact of services \(not just their technical performance\) across heterogeneous contexts and lifecycle stages\[[14](https://arxiv.org/html/2605.22825#bib.bib1),[5](https://arxiv.org/html/2605.22825#bib.bib22)\]\. Key Value Indicators were proposed to make such value outcomes measurable and actionable, but in practice their computation is challenging: value effects are indirect and multi\-stakeholder, relevant evidence is often incomplete or only available via proxies \(measurements, certifications, surveys\), and early\-stage designs must still provide credible estimates with transparent assumptions\[[22](https://arxiv.org/html/2605.22825#bib.bib11)\]\. Without reproducible calculation workflows that link service descriptions to evidence\-backed Key Value Indicators, value assessment risks becoming ad\-hoc, hard to optimize, and prone to inconsistency or “value\-washing”\[[21](https://arxiv.org/html/2605.22825#bib.bib7)\]\.

Existing approaches broadly fall into two strands\. First, concept and governance frameworks motivate Key Value Indicators and propose value\-driven assessment processes \(e\.g\., eliciting stakeholders and values, defining indicator candidates, and staging evaluation by maturity\), which is valuable for shared language and decision\-making, but typically stops short of prescribing fully specified, end\-to\-end computation pipelines from service descriptions to concrete, reproducible indicator values\. Second, operational approaches embed Key Value Indicators into orchestration and optimization \(e\.g\., ranking alternatives or trading off performance and value objectives\), demonstrating that value\-aware decisions are possible when indicators are computable, but often do so for a small, pre\-selected Key Value Indicator set and under strong assumptions about the availability, meaning, and provenance of required inputs\. In practice, however, the hard part is often*operationalization*under realistic constraints: deciding which Key Value Indicators are actually relevant for a new heterogeneous service, turning narrative requirements into a measurement plan, and then computing results from mixed evidence sources where some inputs are missing, approximate, or only available via proxies\. These issues motivate end\-to\-end traceability and uncertainty\-aware outputs with clear, user\-facing rationales\.

To address these gaps, this paper proposes KPI2KVI, a Large Language Model\-powered multi\-agent workflow that computes Key Value Indicators from a service description in a general, reproducible, and traceable way\. KPI2KVI uses specialized Large Language Model\-based agents to \(i\) conduct a guided interview that elicits service intent, context, stakeholders, and potential value impacts, \(ii\) map the service to a controlled Key Value Indicator taxonomy and finalize the Key Value Indicator scope with human\-in\-the\-loop refinement, \(iii\) generate a compact, service\-specific Key Performance Indicator evidence plan for the selected Key Value Indicators and collect/structure the resulting measurements with provenance, and \(iv\) compute each Key Value Indicator with explicit\{exact,min,max\}\\\{\\text\{exact\},\\text\{min\},\\text\{max\}\\\}bounds and a short rationale that cites the precise Key Performance Indicators inputs and assumptions used\. By combining Large Language Model semantic understanding with a deterministic staged pipeline and persistent structured artifacts, KPI2KVI systematically bridges stakeholder value expectations to measurable evidence and makes uncertainty, assumptions, and computation steps explicit and auditable\.

The rest of this paper is organized as follows\. Section II reviews some related work on Key Value Indicators concepts, frameworks, and optimization\-based operationalizations\. Section III presents the KPI2KVI workflow and architecture in detail\. Section IV evaluates the approach through simulations\. Section V concludes and outlines directions for future work\.

## IILiterature Review

In Sixth Generation, a*service*is an end\-to\-end capability for users or verticals realized by chaining functions across heterogeneous domains \(edge/cloud, terrestrial/non\-terrestrial\) and governed through Service Level Agreements or intent\-based abstractions\[[3](https://arxiv.org/html/2605.22825#bib.bib14),[22](https://arxiv.org/html/2605.22825#bib.bib11),[4](https://arxiv.org/html/2605.22825#bib.bib21),[5](https://arxiv.org/html/2605.22825#bib.bib22)\]\. A service request therefore combines functional goals with workload characteristics and stringent*requirements*on latency, throughput/data rate, reliability/availability, coverage, positioning, privacy, and security, often context\-dependent\[[3](https://arxiv.org/html/2605.22825#bib.bib14),[22](https://arxiv.org/html/2605.22825#bib.bib11),[19](https://arxiv.org/html/2605.22825#bib.bib17),[18](https://arxiv.org/html/2605.22825#bib.bib18),[8](https://arxiv.org/html/2605.22825#bib.bib19),[9](https://arxiv.org/html/2605.22825#bib.bib20),[7](https://arxiv.org/html/2605.22825#bib.bib23)\]\. In intent\-based formulations, services may be decomposed into tasks and mapped to intent categories to make requirements machine\-actionable and comparable\[[3](https://arxiv.org/html/2605.22825#bib.bib14)\]\. Key Performance Indicators are technical, measurable quantities estimating performance \(e\.g\., delay, throughput, packet loss\), typically specified as desired values with tolerable thresholds\[[16](https://arxiv.org/html/2605.22825#bib.bib4),[22](https://arxiv.org/html/2605.22825#bib.bib11)\]\. Key Value Indicators complement Key Performance Indicators by estimating enabled \(or harmed\) societal values such as sustainability, inclusion, privacy/confidentiality, and trust\[[22](https://arxiv.org/html/2605.22825#bib.bib11),[21](https://arxiv.org/html/2605.22825#bib.bib7)\]\. Because many value dimensions are not directly observable at run time, Key Value Indicators are often realized via sensor\-based measurements, periodic certification/audits, and compositions over lower\-level indicators\[[14](https://arxiv.org/html/2605.22825#bib.bib1)\]\. This Key Performance Indicator–Key Value Indicator split motivates values\-driven service design and evaluation beyond performance\-only engineering\.

Several works defined Key Value Indicators and provided high\-level workflows rather than fully specified calculation pipelines\. Atzoriet al\.\[[1](https://arxiv.org/html/2605.22825#bib.bib8)\]proposed EthicNet/Value of Service, where stakeholders expressed Key Value Indicator requirement profiles and Key Value Indicators should be monitored and composed end\-to\-end, but composition operators remained open\. Pintoret al\.\[[14](https://arxiv.org/html/2605.22825#bib.bib1)\]\(and\[[13](https://arxiv.org/html/2605.22825#bib.bib5)\]\) systematized architectural formalization \(sensor\-based vs\. certified Key Value Indicators, metadata, and parent–child structuring\)\. Wikströmet al\.\[[21](https://arxiv.org/html/2605.22825#bib.bib7)\]offered a five\-step framework from scenario/value elicitation to Key Value Indicator formulation and staged assessment, while their white paper defined Key Value Indicators as the*scale of effect*of a use case and linked Key Values→\\rightarrowKey Value Indicators→\\rightarrowenablers→\\rightarrowKey Performance Indicators, explicitly noting the unknown “exchange rate” between usage and societal value\[[22](https://arxiv.org/html/2605.22825#bib.bib11)\]\. Ziegleret al\.\[[23](https://arxiv.org/html/2605.22825#bib.bib9)\]provided qualitative Key Performance Indicator\-to\-value impact mapping, and Osmanet al\.\[[11](https://arxiv.org/html/2605.22825#bib.bib3)\]emphasized Key Performance Indicator proxies and business\-model\-driven prioritization\. Other Sixth Generation discussions highlighted ecosystem/governance drivers without prescribing calculation rules\[[17](https://arxiv.org/html/2605.22825#bib.bib10),[2](https://arxiv.org/html/2605.22825#bib.bib13),[15](https://arxiv.org/html/2605.22825#bib.bib12)\], while enterprise Business Intelligence work focused on Key Performance Indicator aggregation/visualization rather than Key Performance Indicator→\\rightarrowKey Value Indicator translation\[[6](https://arxiv.org/html/2605.22825#bib.bib2)\]\. Across these contributions, assessment was often staged by maturity: early Technology Readiness Levels relied on expert/qualitative evidence, later Technology Readiness Levels relied on measurement and more objective signals\[[22](https://arxiv.org/html/2605.22825#bib.bib11)\]\.

A second group operationalized Key Value Indicators through explicit calculation embedded in orchestration/optimization\. De Trizioet al\.\[[3](https://arxiv.org/html/2605.22825#bib.bib14)\]modeled intent\-mapped service provisioning as a many\-to\-many matching problem, combining Key Performance Indicator constraints \(deadline, throughput\) with Key Value Indicator\-related constraints \(budget, risk appetite\) and ranking providers by entropy\-weighted Technique for Order Preference by Similarity to Ideal Solution over cost and cyber risk\. Sciddurloet al\.\[[16](https://arxiv.org/html/2605.22825#bib.bib4)\]defined Key Performance Indicator vectors for services/resources, computed Key Value Indicator components \(environmental sustainability, trustworthiness, inclusiveness\) via formulas, and optimized Key Performance Indicator–Key Value Indicator trade\-offs via a bi\-objective model solved by an exactϵ\\epsilon\-constraint method; they also proposed translating natural\-language requests into intents enriched with Key Performance Indicators and Key Value Indicators\. Mertenset al\.\[[10](https://arxiv.org/html/2605.22825#bib.bib6)\]proposed Sustainable Development Goal\-indexed Key Value Indicators where “objective” service Key Value Indicators could be scored via International Organization for Standardization\-standards coverage ratios and combined with user preference profiles\. Methodologically, these works relied on normalization and aggregation \(e\.g\., relative\-closeness ranking in\[[3](https://arxiv.org/html/2605.22825#bib.bib14)\], weighted\-sum Key Value Indicator aggregation and Pareto optimization in\[[16](https://arxiv.org/html/2605.22825#bib.bib4)\]\)\.

Despite rapid progress, gaps remained for computing*service\-relevant*Key Value Indicators in a general, reproducible way\. Vision/framework works clarified Key Value Indicator concepts and governance, but often did not specify how to select relevant Key Value Indicators for a new service nor how to compute them from available evidence without extensive manual modeling\[[1](https://arxiv.org/html/2605.22825#bib.bib8),[21](https://arxiv.org/html/2605.22825#bib.bib7),[14](https://arxiv.org/html/2605.22825#bib.bib1)\]\. Qualitative mappings and vertical matrices communicated priorities but were difficult to audit, compare, or optimize because scales and aggregation rules were coarse or subjective\[[23](https://arxiv.org/html/2605.22825#bib.bib9),[15](https://arxiv.org/html/2605.22825#bib.bib12)\]\. Optimization\-oriented work, in contrast, typically fixed a small Key Value Indicator set and assumed access to non\-trivial inputs \(e\.g\., carbon factors, attack likelihoods, certification mappings\) that might be unavailable or ambiguous at design time\[[3](https://arxiv.org/html/2605.22825#bib.bib14),[16](https://arxiv.org/html/2605.22825#bib.bib4),[10](https://arxiv.org/html/2605.22825#bib.bib6)\]\. Across strands, there was limited support for missing/uncertain evidence, end\-to\-end traceability from service description to computed indicators, and systematic bridging from stakeholder value expectations to measurable Key Performance Indicator evidence\[[22](https://arxiv.org/html/2605.22825#bib.bib11),[11](https://arxiv.org/html/2605.22825#bib.bib3)\]\. Moreover, many approaches provided neither uncertainty bounds \(e\.g\., intervals\) nor concise, user\-facing rationales connecting results to underlying evidence\. These limitations motivated more operational and traceable Key Performance Indicator\-to\-Key Value Indicator computation workflows that remained usable under uncertainty and heterogeneous services\.

## IIIApproach

This section presents KPI2KVI, a multi\-agent workflow that calculates Key Value Indicators for a given service\. We first define the nine\-stage workflow and the roles of its Large Language Model\-based agents, then describe the system architecture and orchestration logic that implement this workflow, and finally walk through a concrete example that shows how a small set of Key Value Indicators can be calculated for a cloud\-based telemedicine video consultation service\.

![Refer to caption](https://arxiv.org/html/2605.22825v1/x1.png)Figure 1:KPI2KVI pipeline: from service interview to Key Value Indicator category selection, Key Performance Indicator generation and value collection, per\-Key Value Indicator calculation with explicit bounds, and post\-hoc advising, with shared memory storing reusable artifacts across stages\.### III\-AWorkflow and Agent Responsibilities

The end\-to\-end workflow is a deterministic pipeline composed of nine steps, alternating between conversational steps \(to elicit information and allow corrections\) and structured steps \(to produce machine\-readable artifacts\)\. The progression of stages and the shared memory that carries artifacts across them are shown in Fig\.[1](https://arxiv.org/html/2605.22825#S3.F1)\.

#### Key Value Indicator Category Selection

The workflow starts with the*inspector*\(Step 1\), whoseinputsare the initial service description and any subsequent clarifications\. Itsprocessis a guided interview that elicits service intent, operational context, stakeholders, and potential value impacts \(e\.g\., privacy/security expectations, sustainability concerns, accessibility constraints\), and itsoutputis an interview transcript stored in memory and used as the primary evidence for downstream selection\. Next, the*kvi category extractor*\(Step 2\) takes asinputsthe full inspector transcript and the global Key Value Indicator taxonomy,processesthem by mapping the service to the most relevant Key Value Indicator categories using the taxonomy as a controlled vocabulary, andoutputsa ranked structured list of candidate categories for refinement\. The taxonomy is extracted from the KVIs proposed in\[[12](https://arxiv.org/html/2605.22825#bib.bib15)\]\. The*kvi category evaluator*\(Step 3\) then uses asinputsthe extracted categories and taxonomy context,processesthem via a short conversational loop that justifies the proposal and invites the user to add/remove/replace categories based on domain knowledge and stakeholder priorities, andoutputsa refinement transcript capturing the user’s decisions\. Finally, the*kvi category finalizer*\(Step 4\) takes asinputsthe extractor output and evaluator transcript,processesthem by consolidating the final category set \(resolving duplicates and ensuring consistent IDs\), andoutputsa structured finalized list of categories that defines the Key Value Indicator scope for subsequent Key Performance Indicator generation and Key Value Indicator calculation; this contract is critical for auditability because it fixes*which*values will be computed before numerical evidence is collected\.

#### KPI Generation

Given the finalized Key Value Indicator scope, KPI2KVI next produces the measurement evidence needed for computation\. The*kpi generator*\(Step 5\) takes asinputsthe inspector transcript \(service context\) and the finalized Key Value Indicator categories \(value scope\),processesthem by proposing a compact set of service\-specific Key Performance Indicators that can serve as measurable evidence for the selected value indicators \(including name, description, and unit for each Key Performance Indicator\), andoutputsa structured Key Performance Indicator list stored in memory and presented to the user as the measurement plan\. The*kpi collector*\(Step 6\) then uses asinputsthis Key Performance Indicator list and the ongoing user chat,processesthem by collecting one Key Performance Indicator value at a time while explicitly supporting missing evidence \(the user may provide a value or delegate it to the system, which is recorded as an assumption rather than an observation\), andoutputsa collection transcript capturing raw values, units, and whether each value was user\-provided or system\-decided\. Finally, the*kpi structurer*\(Step 7\) takes asinputsthe Key Performance Indicator list and the collector transcript,processesthem by converting free\-form conversation into a machine\-readable table of Key Performance Indicator values with provenance flags and normalized representation \(e\.g\., consistent numeric parsing and unit alignment\), andoutputsa structured Key Performance Indicator table that becomes the single source of truth for downstream Key Value Indicator calculations\.

#### KVI Calculation

Given the structured Key Performance Indicator table and finalized Key Value Indicator scope, KPI2KVI computes the target indicators and makes the results explainable\. The*kvi calculator*\(Step 8\) takes asinputs\(i\) one target Key Value Indicator definition, including a step\-by\-step calculation narrative \(a reasoning chain\) describing how to derive the Key Value Indicator and \(ii\) the structured Key Performance Indicator table,processesthem by producing an estimated value with explicit bounds and a short rationale that links the result to the exact Key Performance Indicators and assumptions used, andoutputsper\-Key Value Indicator calculation artifacts\{exact,min,max,rationale\}\\\{\\text\{exact\},\\text\{min\},\\text\{max\},\\text\{rationale\}\\\}stored in memory\. The orchestrator loops this step over all Key Value Indicators implied by the finalized categories and stores results so later stages can cite them precisely\. Finally, the*kvi advisor*\(Step 9\) takes asinputsa consolidated advisor context containing the finalized categories, the structured Key Performance Indicator table, and all per\-Key Value Indicator calculation artifacts,processesthem by answering user questions and explaining trade\-offs with traceable justifications, andoutputsuser\-facing explanations and follow\-up guidance \(e\.g\., which Key Performance Indicators would most reduce uncertainty if measured more precisely\)\.

### III\-BArchitecture

The KPI2KVI workflow is implemented as a two\-tier system with a chat interface and a backend workflow controller\. The frontend provides a conversational User Interface and consumes a Server\-Sent Events stream to incrementally render intermediate status updates and agent outputs\. The backend exposes a streaming Application Programming Interface endpoint and delegates each user turn to a workflow engine that selects and executes the appropriate step of the workflow, persists the session state, and emits progress and content events for the UI\. The runtime architecture is shown in Fig\.[2](https://arxiv.org/html/2605.22825#S3.F2)\. The workflow controller \(orchestrator\) is the central component that realizes the nine\-stage process described above\. It discovers agent modules from a registry, executes them through an Large Language Model interface, and maintains cross\-agent memory\. Concretely, the orchestrator implements a staged state machine where each session stores \(i\) the complete chat history, \(ii\) the name of the*current*agent responsible for the next user message, and \(iii\) a key–value store of artifacts accumulated so far\. Artifacts include both human\-readable text \(assistant messages shown to the user\) and structured objects \(e\.g\., lists of selected Key Value Indicator category IDs, a Key Performance Indicator list with units/descriptions, and a table of collected Key Performance Indicator values with provenance\)\. This design is what enables the system to advance from one workflow step to the next without losing context, while still keeping each step’s output explicit and auditable\.

Operationally, the orchestrator performs three actions on every user turn\. First, it*builds context*: it constructs the next prompt by combining the new user message with the relevant subset of stored artifacts \(e\.g\., inspector transcript for category extraction; finalized categories for Key Performance Indicator generation; structured Key Performance Indicator table for Key Value Indicator calculation\)\. Second, it*runs and routes*: it executes the current agent, inspects the response for completion cues \(for conversational steps\), and when a step is complete it automatically triggers the next structured step\(s\) without requiring another user message\. Third, it*stores and streams*: it writes outputs back to memory under explicit keys and streams progress/content events so the frontend can render multi\-agent turns coherently\. Finally, the architecture separates*domain knowledge*from*interaction logic*: the orchestrator loads the shared Key Value Indicator taxonomy and canonical Key Value Indicator definitions from data assets and passes them into relevant agents, ensuring stable IDs/codes and enabling reuse of intermediate structured artifacts across runs\.

![Refer to caption](https://arxiv.org/html/2605.22825v1/x2.png)Figure 2:TheKPI2KVI architecture: a streaming frontend/backend setup with a workflow controller that discovers agents, executes them via an LLM provider, and manages cross\-agent state and reusable structured artifacts\.
### III\-CIllustrative Example and Addressing Literature Gaps

Consider a simple service: a cloud\-based video consultation platform that enables remote medical appointments\. The service description typically includes performance intent \(stable video/audio, low delay, high availability\) and operational constraints \(sensitive personal data, multi\-tenant deployment, and varying client devices\)\[[4](https://arxiv.org/html/2605.22825#bib.bib21),[19](https://arxiv.org/html/2605.22825#bib.bib17),[8](https://arxiv.org/html/2605.22825#bib.bib19),[7](https://arxiv.org/html/2605.22825#bib.bib23)\]\. In KPI2KVI, the inspector first elicits missing context \(e\.g\., what data are processed and stored, who can access recordings, which regulations apply, and the expected user base\)\. Based on this transcript, the category extractor selects a single relevant Key Value Indicator category, for example*User Trust, Perception, & Requirement Compliance*, which in the KPI2KVI taxonomy maps to Key Value Indicators such asPUC\-UPCA\(share of identified user privacy concerns addressed, %\),PUC\-USCA\(share of identified user security concerns addressed, %\), andRPS\-DDSS\(user\-reported perceived security of the service, %\)\.

The Key Performance Indicator generator then proposes a small evidence set sufficient to compute these Key Value Indicators\. For this example, consider three Key Performance Indicators:NpN\_\{p\}= number of privacy concerns collected during requirements elicitation \(count\),ApA\_\{p\}= number of those privacy concerns addressed by implemented controls \(count\), andrsr\_\{s\}= average user perceived\-security score on a 1–5 Likert scale from a short pilot survey \(dimensionless\)\. Suppose the Key Performance Indicator collection yieldsNp=10N\_\{p\}=10, and because some controls are still under implementation the collector recordsAp∈\[7,9\]A\_\{p\}\\in\[7,9\]\(delegated estimate\)\. From a pilot with limited respondents, assumers∈\[3\.8,4\.4\]r\_\{s\}\\in\[3\.8,4\.4\]with a nominal meanrs=4\.1r\_\{s\}=4\.1\. These values are stored in the structured Key Performance Indicator table with provenance \(the interval\-valued entries explicitly reflecting uncertainty\)\.

The Key Value Indicator calculator then produces transparent formulas and interval results\. ForPUC\-UPCA, a natural operationalization is the percentage of addressed privacy concerns:

PUC\-UPCA=100⋅ApNp\.\\texttt\{PUC\-UPCA\}=100\\cdot\\frac\{A\_\{p\}\}\{N\_\{p\}\}\.WithNp=10N\_\{p\}=10andAp∈\[7,9\]A\_\{p\}\\in\[7,9\], the result isPUC\-UPCAmin=70\\texttt\{PUC\-UPCA\}\_\{\\min\}=70,PUC\-UPCAmax=90\\texttt\{PUC\-UPCA\}\_\{\\max\}=90, andPUC\-UPCAexact=80\\texttt\{PUC\-UPCA\}\_\{\\text\{exact\}\}=80\(using the midpoint\)\. For perceived securityRPS\-DDSS, the calculator can map the Likert score to a 0–100 scale:

RPS\-DDSS=100⋅rs−14\.\\texttt\{RPS\-DDSS\}=100\\cdot\\frac\{r\_\{s\}\-1\}\{4\}\.Withrs∈\[3\.8,4\.4\]r\_\{s\}\\in\[3\.8,4\.4\], we obtainRPS\-DDSSmin=70\\texttt\{RPS\-DDSS\}\_\{\\min\}=70,RPS\-DDSSmax=85\\texttt\{RPS\-DDSS\}\_\{\\max\}=85, andRPS\-DDSSexact=77\.5\\texttt\{RPS\-DDSS\}\_\{\\text\{exact\}\}=77\.5\. Importantly, the advisor can later explain that the uncertainty comes from delegated control\-coverage assumptions and limited survey evidence, and can recommend which Key Performance Indicators \(e\.g\., tighteningApA\_\{p\}via control verification or increasing the survey sample forrsr\_\{s\}\) would most reduce the Key Value Indicator bounds\. Overall, the example demonstrates a concrete chain from service description to bounded, explainable Key Value Indicators outputs grounded in an explicit Key Performance Indicator evidence set\.

### III\-DAddressing Literature Gaps

KPI2KVI directly addresses the main gaps identified in Sec\. II by making Key Performance Indicator\-to\-Key Value Indicator computation operational, traceable, and usable under uncertainty, while keeping the workflow largely automatic through Large Language Models\. First, instead of assuming a fixed, small Key Value Indicator set or requiring extensive manual modeling, Large Language Models perform taxonomy\-grounded category extraction and refinement, and the inspector/evaluator loops explicitly incorporate human feedback before a finalized category contract is recorded\. Second, it systematically bridges stakeholder value expectations to measurable evidence by using Large Language Models to propose a compact, service\-specific Key Performance Indicator set and by structuring collected values into a single machine\-readable table that can be reused and compared across runs\. Third, computation and explanation are supported by the Key Value Indicator definitions and the calculator’s step\-by\-step reasoning: for each Key Value Indicator, the system produces bounded results with short rationales tied to the exact Key Performance Indicators and assumptions, and the advisor can answer questions by citing stored artifacts\. Together, these mechanisms improve end\-to\-end traceability from the service description to computed Key Value Indicators, and make uncertainty and human\-in\-the\-loop corrections explicit rather than implicit\.

## IVSimulations

We evaluate KPI2KVI by testing four variants: \(1\)*monolithic LLM*\(DeepSeek\-R1\), where a single system prompt describes the end\-to\-end KPI2KVI logic and embeds the Key Value Indicator taxonomy; \(2\)*agentic SLM \(no taxonomy, no CoT\)*, where the KPI2KVI agents are implemented with an Small Language Model \(Gemini 2\.5 Flash Lite\) and have no access to the taxonomy and no explicit Chain\-of\-Thought calculation prompting; \(3\)*agentic SLM \(\+ taxonomy, no CoT\)*, identical to \(2\) but with access to the taxonomy artifacts; and \(4\)*KPI2KVI \(full\)*\. To produce the measurements, we execute each method end\-to\-end on a suite of service cases and systematically vary three experimental conditions: \(i\) the computational difficulty of Key Value Indicator derivation, \(ii\) the scope of requested Key Value Indicators, and \(iii\) the quality of taxonomy grounding available to the method\. Concretely, calculation complexity is controlled by selecting Key Value Indicator instances whose definitions require different formula depths and different numbers of Key Performance Indicators per Key Value Indicator; the Key Value Indicator scope is controlled by varying the number of requested Key Value Indicator categories; and taxonomy qualityq∈\[0,1\]q\\in\[0,1\]is controlled by applying a reproducible degradation procedure to the taxonomy artifacts \(e\.g\., removing a fraction1−q1\-qof entries or fields such as aliases, descriptions, and IDs\), thereby modulating how reliable taxonomy grounding is\. For each x\-axis point, we repeat each method 10 times \(different seeds and prompt paraphrases\); the plotted curves show the mean and the shaded regions show the variance across runs\.

![Refer to caption](https://arxiv.org/html/2605.22825v1/x3.png)Figure 3:Simulation results for the four method variants\. \(A\) KVI instability across repeated runs versus calculation complexity \(formula depth / \#KPIs per KVI\)\. \(B\) end\-to\-end success rate \(runs without verifier\-flagged errors\) versus the number of requested KVI categories\. \(C\) category\-selection accuracy \(F1\) versus taxonomy qualityqq\(coverage/correctness\)\. Curves show mean over 10 runs; shaded regions show variance\.Fig\.[3](https://arxiv.org/html/2605.22825#S4.F3)highlights complementary failure modes\. In Fig\.[3](https://arxiv.org/html/2605.22825#S4.F3)–A, increasing complexity amplifies run\-to\-run variation in computed Key Value Indicators\. KPI2KVI \(4\) remains the most stable because it structures the Key Performance Indicator table, applies explicit calculation steps with bounded outputs, and enables the verifier to catch inconsistencies\. The Small Language Model baselines without Chain\-of\-Thought \(2–3\) show higher instability as multi\-step arithmetic and unit handling become harder, while the monolithic LLM \(1\) is typically more stable than the Small Language Model variants due to higher model capacity but less stable than the fully staged pipeline because it lacks intermediate structured artifacts and per\-Key Value Indicator contracts\. In Fig\.[3](https://arxiv.org/html/2605.22825#S4.F3)–B, the success rate drops as more Key Value Indicator categories are requested, reflecting the difficulty of maintaining a consistent category scope, collecting sufficient Key Performance Indicator evidence, and completing all computations at scale\. Taxonomy access delays the collapse for the agentic Small Language Model \(3\) relative to \(2\), and the monolithic LLM \(1\) degrades more gracefully than the Small Language Model baselines but still fails earlier than KPI2KVI \(4\), which stays robust due to its staged contract \(finalized categories\), shared memory artifacts, and per\-Key Value Indicator computation loop\. Finally, Fig\.[3](https://arxiv.org/html/2605.22825#S4.F3)–C isolates taxonomy dependence: method \(2\) is largely insensitive toqq\(no taxonomy\), method \(1\) improves withqqbecause the taxonomy is embedded in its prompt, and method \(3\) benefits strongly from a high\-quality taxonomy but can be misled whenqqis low; KPI2KVI \(4\) consistently achieves the best category\-selection accuracy by combining taxonomy grounding with the inspector/evaluator refinement loop and explicit finalization of categories\.

## VConclusion

In this paper, we presented KPI2KVI, a multi\-agent workflow that calculated Key Value Indicators for a given service from its description by translating stakeholder intent into measurable Key Performance Indicators and then producing traceable, bounded Key Value Indicator results\. We defined a deterministic nine\-stage pipeline, specified the responsibilities of each Large Language Model\-based agent, and implemented an orchestrated architecture with shared memory that stored reusable artifacts such as the interview transcript, the finalized Key Value Indicator category contract, and a structured Key Performance Indicator value table\. Using taxonomy\-grounded category selection and a human\-in\-the\-loop refinement loop, the workflow fixed the Key Value Indicator scope before evidence collection, which improved auditability and reduced ambiguity in downstream calculations\. We then computed each Key Value Indicator with explicit formulas, interval bounds, and short rationales that linked outputs to the exact Key Performance Indicators and assumptions used\. In simulations, KPI2KVI achieved the best overall robustness across increasing calculation complexity, larger requested Key Value Indicator scopes, and varying taxonomy quality\. As future work, we planned to fine\-tune Small Language Models so they no longer required a taxonomy and explicit Chain\-of\-Thought prompting for reliable Key Value Indicator derivation, and to integrate Retrieval\-Augmented Generation to improve long\-horizon memory management and reuse of past artifacts across sessions and services\.

## Acknowledgment

The work in this paper was supported in part by the Federal Ministry of Research, Technology, and Space \(BMFTR\), Germany, through the Project 6GEM\+ under Grant 16KIS2411; and in part by the European Union through the 6G\-SANDBOX project \(Grant No\.101096328\) and the 6G\-Path project \(Grant No\. 101139172\)\.

## References

- \[1\]L\. Atzori, C\. Campolo, A\. Iera,et al\.\(2023\-11\)Toward the EthicNet: Challenges and Enablers for Ethics\-Aware Networks\.IEEE Communications Magazine61\(11\),pp\. 192–198\.External Links:ISSN 1558\-1896,[Document](https://dx.doi.org/10.1109/MCOM.004.2300110)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p2.4),[§II](https://arxiv.org/html/2605.22825#S2.p4.1)\.
- \[2\]C\. Christophorou, I\. Ioannou, V\. Vassiliou,et al\.\(2023\-12\)ADROIT6G DAI\-Driven Open and Programmable Architecture for 6G Networks\.In2023 IEEE Globecom Workshops \(GC Wkshps\),pp\. 744–750\.External Links:[Document](https://dx.doi.org/10.1109/GCWkshps58843.2023.10465217)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p2.4)\.
- \[3\]F\. de Trizio, G\. Sciddurlo, I\. Cianci,et al\.\(2024\-12\)Optimizing Key Value Indicators in Intent\-Based Networks through Digital Twins aided service orchestration mechanisms\.Computer Communications228,pp\. 107977\.External Links:ISSN 0140\-3664,[Document](https://dx.doi.org/10.1016/j.comcom.2024.107977)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p1.1),[§II](https://arxiv.org/html/2605.22825#S2.p3.1),[§II](https://arxiv.org/html/2605.22825#S2.p4.1)\.
- \[4\]M\. Farhoudi, M\. Shokrnezhad, S\. Kianpisheh,et al\.\(2025\-06\)Deep Learning Based Service Composition in Integrated Aerial\-Terrestrial Networks\.In2025 IEEE 11th International Conference on Network Softwarization \(NetSoft\),pp\. 204–208\.Note:ISSN: 2693\-9789External Links:ISSN 2693\-9789,[Document](https://dx.doi.org/10.1109/NetSoft64993.2025.11080624)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p1.1),[§III\-C](https://arxiv.org/html/2605.22825#S3.SS3.p1.1)\.
- \[5\]M\. Farhoudi, M\. Shokrnezhad, and T\. Taleb\(2025\)Service Registration, Indexing, Discovery, and Selection: An Architectural Survey Toward a GenAI\-Driven Future\.IEEE Access13,pp\. 209680–209722\.External Links:ISSN 2169\-3536,[Document](https://dx.doi.org/10.1109/ACCESS.2025.3642631)Cited by:[§I](https://arxiv.org/html/2605.22825#S1.p1.1),[§II](https://arxiv.org/html/2605.22825#S2.p1.1)\.
- \[6\]V\. D\. Kolychev and A\. A\. Shebotinov\(2019\)Application of Business Intelligence instrumental tools for visualization of key performance indicators of an enterprise in telecommunications\.Scientific Visualization11\(1\)\.External Links:ISSN 2079\-3537,[Document](https://dx.doi.org/10.26583/sv.11.1.03)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p2.4)\.
- \[7\]H\. Mazandarani, M\. Farhoudi, M\. Shokrnezhad,et al\.\(2025\-08\)Adaptive Multiple Access and Service Placement for Generative Diffusion Models\.arXiv\.Note:arXiv:2508\.07978 \[cs\]Comment: This manuscript has been accepted for presentation at IEEE GLOBECOM 2025\. You can use this material personally\. Reprinting or republishing this material for the purpose of advertising or promotion, etc\., must adhere to IEEE policy\. The DOI will be supplied as soon as it becomes availableExternal Links:[Document](https://dx.doi.org/10.48550/arXiv.2508.07978)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p1.1),[§III\-C](https://arxiv.org/html/2605.22825#S3.SS3.p1.1)\.
- \[8\]H\. Mazandarani, M\. Shokrnezhad, and T\. Taleb\(2025\)A Novel Multiple Access Scheme for Heterogeneous Wireless Communications Using Symmetry\-Aware Continual Deep Reinforcement Learning\.IEEE Transactions on Machine Learning in Communications and Networking3,pp\. 353–368\.External Links:ISSN 2831\-316X,[Document](https://dx.doi.org/10.1109/TMLCN.2025.3546183)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p1.1),[§III\-C](https://arxiv.org/html/2605.22825#S3.SS3.p1.1)\.
- \[9\]H\. Mazandarani, M\. Shokrnezhad, and T\. Taleb\(2025\-05\)Semantic\-Aware Dynamic and Distributed Power Allocation: a Multi\-UAV Area Coverage Use Case\.In2025 IEEE International Conference on Machine Learning for Communication and Networking \(ICMLCN\),pp\. 1–6\.External Links:[Document](https://dx.doi.org/10.1109/ICMLCN64995.2025.11140421)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p1.1)\.
- \[10\]J\. S\. Mertens, L\. Galluccio, and A\. Lombardo\(2024\-09\)A deep dive into KVIs for ethics\-aware networks\.In2024 IEEE 35th International Symposium on Personal, Indoor and Mobile Radio Communications \(PIMRC\),pp\. 1–6\.Note:ISSN: 2166\-9589External Links:ISSN 2166\-9589,[Document](https://dx.doi.org/10.1109/PIMRC59610.2024.10817271)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p3.1),[§II](https://arxiv.org/html/2605.22825#S2.p4.1)\.
- \[11\]H\. Osman, J\. Bradford, and S\. Mitchell\(2024\-04\)Bridging the Gap Between 6G Technologies and Societal Values: A Comprehensive Analysis of Key Value Indicators \(KVIs\) and Business Models\.In2024 IEEE Wireless Communications and Networking Conference \(WCNC\),pp\. 1–5\.Note:ISSN: 1558\-2612External Links:ISSN 1558\-2612,[Document](https://dx.doi.org/10.1109/WCNC57260.2024.10570540)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p2.4),[§II](https://arxiv.org/html/2605.22825#S2.p4.1)\.
- \[12\]I\. Patsouras, A\. Charemis, I\. Wedikkara Gedara,et al\.\(2025\-04\)6G KVIs – SNS Projects Initial Survey Results 2025\.Technical reportZenodo\.External Links:[Document](https://dx.doi.org/10.5281/zenodo.15220946)Cited by:[§III\-A](https://arxiv.org/html/2605.22825#S3.SS1.SSS0.Px1.p1.1)\.
- \[13\]L\. Pintor, L\. Atzori, and A\. Iera\(2024\-09\)Building the Foundations of Ethical Networks: Integrating Key Value Indicators for Social, Economic, and Environmental Impact\.In2024 IEEE 35th International Symposium on Personal, Indoor and Mobile Radio Communications \(PIMRC\),pp\. 1–6\.Note:ISSN: 2166\-9589External Links:ISSN 2166\-9589,[Document](https://dx.doi.org/10.1109/PIMRC59610.2024.10817165)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p2.4)\.
- \[14\]L\. Pintor, L\. Atzori, and A\. Iera\(2025\-10\)Sustainability in telecommunication networks and Key Value Indicators: A survey\.Computer Networks271,pp\. 111466\.External Links:ISSN 1389\-1286,[Document](https://dx.doi.org/10.1016/j.comnet.2025.111466)Cited by:[§I](https://arxiv.org/html/2605.22825#S1.p1.1),[§II](https://arxiv.org/html/2605.22825#S2.p1.1),[§II](https://arxiv.org/html/2605.22825#S2.p2.4),[§II](https://arxiv.org/html/2605.22825#S2.p4.1)\.
- \[15\]A\. Pouttu\(2020\-06\)6G white paper on validation and trials for verticals towards 2030’s\.raportti,University of Oulu\(eng\)\.Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p2.4),[§II](https://arxiv.org/html/2605.22825#S2.p4.1)\.
- \[16\]G\. Sciddurlo, F\. de Trizio, G\. Piro,et al\.\(2025\-09\)A value\-driven system design framework for sustainable 6G networks\.Computer Networks269,pp\. 111477\.External Links:ISSN 1389\-1286,[Document](https://dx.doi.org/10.1016/j.comnet.2025.111477)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p1.1),[§II](https://arxiv.org/html/2605.22825#S2.p3.1),[§II](https://arxiv.org/html/2605.22825#S2.p4.1)\.
- \[17\]S\. Seppo Yrjölä, P\. Ahokangas, and M\. Matinmikko\-Blue\(2022\)Value Creation and Capture From Technology Innovation in the 6G Era\.IEEE Access10,pp\. 16299–16319\.External Links:ISSN 2169\-3536,[Document](https://dx.doi.org/10.1109/ACCESS.2022.3149590)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p2.4)\.
- \[18\]M\. Shokrnezhad and S\. Khorsandi\(2018\-03\)Joint power control and channel assignment in uplink IoT Networks: A non\-cooperative game and auction based approach\.Computer Communications118,pp\. 1–13\.External Links:ISSN 0140\-3664,[Document](https://dx.doi.org/10.1016/j.comcom.2018.01.002)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p1.1)\.
- \[19\]M\. Shokrnezhad and T\. Taleb\(2022\-12\)Near\-optimal Cloud\-Network Integrated Resource Allocation for Latency\-Sensitive B5G\.InGLOBECOM 2022 \- 2022 IEEE Global Communications Conference,pp\. 4498–4503\.Note:ISSN: 2576\-6813External Links:ISSN 2576\-6813,[Document](https://dx.doi.org/10.1109/GLOBECOM48099.2022.10001109)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p1.1),[§III\-C](https://arxiv.org/html/2605.22825#S3.SS3.p1.1)\.
- \[20\]Q\. Wang, A\. Diaz Zayas, L\. Cordeiro,et al\.\(2026\-05\)6G\-PATH Open Experimentation Platform for B5G Enablers and Diverse Vertical Applications\.In2025 IEEE International Conference on Communications Workshops \(ICC Workshops\),Glasgow, Scotland, UK,pp\. 1–6\.Cited by:[§I](https://arxiv.org/html/2605.22825#S1.p1.1)\.
- \[21\]G\. Wikström, N\. Bledow, M\. Matinmikko\-Blue,et al\.\(2024\-07\)Key value indicators: A framework for values\-driven next\-generation ICT solutions\.Telecommunications Policy48\(6\),pp\. 102778\.External Links:ISSN 0308\-5961,[Document](https://dx.doi.org/10.1016/j.telpol.2024.102778)Cited by:[§I](https://arxiv.org/html/2605.22825#S1.p1.1),[§II](https://arxiv.org/html/2605.22825#S2.p1.1),[§II](https://arxiv.org/html/2605.22825#S2.p2.4),[§II](https://arxiv.org/html/2605.22825#S2.p4.1)\.
- \[22\]G\. Wikström, A\. Schuler Scott, I\. Mesogiti,et al\.\(2022\-05\)What societal values will 6G address?\.Technical reportZenodo\.External Links:[Document](https://dx.doi.org/10.5281/zenodo.6557534)Cited by:[§I](https://arxiv.org/html/2605.22825#S1.p1.1),[§II](https://arxiv.org/html/2605.22825#S2.p1.1),[§II](https://arxiv.org/html/2605.22825#S2.p2.4),[§II](https://arxiv.org/html/2605.22825#S2.p4.1)\.
- \[23\]V\. Ziegler and S\. Yrjola\(2020\-03\)6G Indicators of Value and Performance\.In2020 2nd 6G Wireless Summit \(6G SUMMIT\),pp\. 1–5\.External Links:[Document](https://dx.doi.org/10.1109/6GSUMMIT49458.2020.9083885)Cited by:[§II](https://arxiv.org/html/2605.22825#S2.p2.4),[§II](https://arxiv.org/html/2605.22825#S2.p4.1)\.