# An internal AI rollout needs a measured productivity denominator. LG pairs model publication with an employee service, moving assessment from benchmark scores to the economics of work actually completed. Canonical: https://kgcf.dsmlholdings.com/insights/lg-exaone-three-five-employee-rollout/ Published: 2026-10-08 Author: [DSML Holdings LLC](https://www.dsmlholdings.com/) Company: LG AI Research Event: 2024-12-09 EXAONE 3.5 release and official employee ChatEXAONE service start; counted together as one product operating milestone. ## Reported metrics - Model parameter configurations: 2.4B / 7.8B / 32B. Architecture, not users, revenue or processing savings. [Source 2](https://arxiv.org/abs/2412.04862v1) ## Reported evidence LG released EXAONE 3.5 in 2.4-billion, 7.8-billion and 32-billion parameter configurations and started ChatEXAONE service for employees on 9 December 2024. The pinned technical report says the models are available for research and directs commercial users to LG. The rollout is not a disclosed external sales contract or evidence that every employee gained a quantified productivity improvement. ## Investment interpretation An internal service gives LG a route to observe useful workflows and deployment constraints inside an actual organization. That can improve the model and reveal commercial opportunities. Its return nevertheless needs evidence of net work improvement after verification, infrastructure and support. Internal availability should not be valued as if it were a completed external revenue stream. ## Economic assessment Measure accepted work and resources consumed against a defined baseline. Time saved in drafting may be offset by review or by additional tasks generated. The cash benefit depends on whether released capacity is redeployed productively or a cost is actually avoided. Research distribution and internal adoption can have value without proving immediate monetization. ## The Work Baseline Productivity is meaningful only relative to a task and a counterfactual. Document drafting, retrieval and coding have different error costs, review requirements and completion criteria. An internal rollout should identify where the service reduces the full effort required for accepted work, not merely the time needed to produce an initial answer. If employees spend additional time correcting or checking output, the apparent saving can shrink. If they complete more valuable work with the released time, the benefit may exceed payroll reductions even though headcount remains unchanged. The assessment should preserve that distinction. LG's announcement identifies intended work uses and service availability, but does not quantify realized net savings. A capital appraisal should therefore request representative workflow measurements and adoption quality. It should also allow for valuable learning during the rollout, while ensuring learning is not mislabeled as already achieved operating cash improvement. Internal cost allocation should make infrastructure and support visible to the workflow owner. Otherwise use can appear free and grow without a meaningful resource constraint. A transparent charge or budget can help distinguish genuinely valuable tasks from activity that is attractive only because another part of the organization absorbs the expense. 1. [LG / model release and employee service start](https://www.lg.co.kr/media/release/28438) ## Model Choice as Resource Allocation The three sizes allow different resource and performance trade-offs. Choosing the smallest model solely for lower inference expense could increase correction or failure costs. Choosing the largest for every task could waste infrastructure on simple work. Routing should be based on the quality and reliability required by each workflow. A useful comparison includes context handling, latency, retries and review, using the same accepted outcome. Parameter count cannot establish these economics by itself. The technical report provides architecture and evaluation evidence, but internal tasks may differ from published benchmarks. A deployment strategy can allocate capabilities more efficiently by using different configurations where appropriate. That flexibility requires engineering and maintenance, which belong in the cost model. The business value lies in matching resources to productive work, not in treating the range of model sizes as proof that all enterprise needs are already covered. 1. [LG / model release and employee service start](https://www.lg.co.kr/media/release/28438) 2. [LG authors / EXAONE 3.5 technical report, original version](https://arxiv.org/abs/2412.04862v1) ## Internal Data and Useful Deployment Employee work often depends on documents and systems that a general model cannot access safely or accurately without integration. Retrieval, permission controls and the quality of source information therefore influence usefulness. A capable model can still produce poor results if it receives irrelevant or outdated material. Internal deployment also needs responsibility for access changes, monitoring and incident handling. These are continuing operating functions, not one-time features. The company can learn from its own environment while controlling some deployment conditions more closely than an external vendor might. However, an internal workflow may depend on systems or data unavailable to future customers. Commercial portability should be assessed separately. The rollout creates a practical learning environment, but does not establish that the same service can be supplied externally at the same cost or with the same integration quality. The investment case should identify which capabilities are generalizable and which remain organization-specific. 1. [LG / model release and employee service start](https://www.lg.co.kr/media/release/28438) ## Research Release and a Later Business Research availability can widen evaluation and generate feedback without granting unrestricted commercial use. The pinned report explicitly distinguishes commercial inquiries. A model's distribution strategy and its service revenue strategy should consequently be mapped separately. External customers could value licensing, managed deployment or integration, while internal users provide a different source of evidence. None of these routes should be inferred from download availability alone. The provider needs a clear maintenance promise, rights package and price for the intended commercial use. It also needs to understand how supporting external environments differs from serving its own employees. A broad research community can improve visibility but generate little cash if the paid proposition is unclear. The economic test is whether observed internal usefulness can become a transferable, contractually supported service with recurring contribution. This remains an analytical commercialization question rather than a result reported by the December announcement. 1. [LG / model release and employee service start](https://www.lg.co.kr/media/release/28438) 2. [LG authors / EXAONE 3.5 technical report, original version](https://arxiv.org/abs/2412.04862v1) ## China - DSML comparison Chinese models provide a technical comparison; this employee rollout is not a disclosed Chinese commercial route. ## Japan - DSML comparison Japanese workflows require separate language and document evaluation beyond the stated English-Korean focus. ## Other Asia - Reported connection The employee service is a Korean-company operating milestone, not proof of general Asian enterprise adoption. ## United States - DSML comparison English capability allows testing against US alternatives but does not establish US receipts. ## Europe - DSML comparison European deployment would require separate rights, data and support arrangements; research access alone is insufficient. ## Counterpoint Internal deployment can generate valuable knowledge and protect organizational capability even before direct revenue appears. Its financial appraisal should recognize that option without treating all time-saving claims as realized cash or assuming internal success automatically transfers to external customers. ## Underwriting questions 1. Which workflows show a measured reduction in total accepted-work effort? 2. How are model size, review and infrastructure costs allocated by task? 3. Which internal capabilities can be licensed or served externally under verified commercial terms? ## Primary sources 1. [LG / model release and employee service start](https://www.lg.co.kr/media/release/28438) (2024-12-09) 2. [LG authors / EXAONE 3.5 technical report, original version](https://arxiv.org/abs/2412.04862v1) (2024-12-06) ## Photograph context LG Twin Towers behind Yeouido Park in Seoul; archive corporate context, not an EXAONE launch or AI laboratory. Seoul Metropolitan Government / Wikimedia Commons, CC BY 4.0. Resized to WebP with 360x240 thumbnail; source watermark retained. No endorsement implied. [Photograph source](https://commons.wikimedia.org/wiki/File:Yeouido_Park_(%EC%97%AC%EC%9D%98%EB%8F%84%EA%B3%B5%EC%9B%90)_2013178_002_%EC%9B%90%EA%B2%BD2_resize.jpg) [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)