All casesNo. 065
Hyundai Motor Group Puts AI on App Reviews and Crash Tests
Agents tag and draft app review replies; a crash engineer built his own test assistant
- Case no.
- 065
- Primary source
- Hyundai Motor Group journal article and press releases
- Source published
- 2026-08-12

1. Why this case
On 2026-08-12, Hyundai Motor Group held an AX showcase for the media at its Yangjae headquarters in Seoul, titled “Beyond AI, Transforming the Way We Work.” The same day, the group’s own journal profiled six AI projects across the value chain, from new-vehicle development to manufacturing, dealer service, and customer channels, each told through the employee who led it.
Two of the six show the whole change: what the work looked like before, how it runs now, who does which part, and what came of it.
- App review responses. Reviews of the Hyundai, Kia, and Genesis apps used to be read, classified, tagged, and answered by staff, one by one, under a taxonomy of 17 languages, 63 categories, and 119 tags. A multi-agent system now classifies, tags, and drafts the reply; staff review the output, and cases the AI cannot judge go to a person. Average handling time per review fell from 35 minutes to 5.
- The Crash Safety AI Assistant. At the Namyang R&D Center, about 600 crash tests a year produced reports that each engineer filed in their own way. Finding a past test meant searching several systems or asking a colleague. A senior research engineer on the team built an assistant himself with an in-house coding agent. In four months it received about 2,000 questions, and the group says time spent finding related cases and reviewing analysis material dropped by about 90%.
The other four projects are covered briefly: a single VIN-recognition model now serving about 70 points in plants worldwide, a cart-sequencing optimizer for assembly lines, a manufacturing agent platform still in beta, and an LLM tool for dealer technicians.
The group frames all six as the payoff of earlier groundwork: shared collaboration tools from 2022, a group-wide data pipeline, and an internal generative AI platform, H Chat Pro, which had more than 30,000 users as of July 2026. All figures are the group’s own. Quotes from the Korean-language journal are translated; quotes from the English press release are as published.
2. Problem
2.1 App reviews: one person, one review, 119 tags
The journal calls customer app reviews “an important clue for service improvement,” and the volume is large: thousands of reviews a month reach the Hyundai, Kia, and Genesis apps from customers worldwide. Analyzing and answering them well “was not easy.”
The work was manual from end to end. The English release puts it this way: “Previously, employees were required to manually review, categorize and respond to customer feedback.” The journal adds the detail that made it hard. Staff had to run a classification scheme of 17 languages, 63 categories, and 119 tags, and the standards differed by language and country. Reply quality also varied with the knowledge and experience of whoever handled the review.
2.2 Crash test knowledge: filed by person, lost with people
About 600 crash tests a year are run at the group’s Namyang R&D Center. Each produces test results, images, and analysis files. According to the journal, every engineer organized this material differently, and it was stored in different places. To find past test material, engineers searched several internal systems or asked the person in charge directly, which took considerable time. When an engineer retired or moved, experience and know-how that had never been recorded systematically often disappeared with them, and new engineers struggled to build expertise.
The Korean release describes the same gap from the research side: crash-safety work combines physical tests, computer simulations, and past case analysis, but relevant material was spread across systems or depended on individual engineers’ experience.
2.3 The other four
- VIN check on the line. Production lines manage each vehicle’s spec and production data by its vehicle identification number (VIN), and confirming that the registered data matches the actual vehicle is a required step. Workers judged the match from system data alone or compared by eye. As processes grew more complex, mismatches occurred occasionally, with a risk of misassembly. A first AI version ran at a few Korean plants, but each plant and line had its own model and server, which made central management hard and maintenance heavy.
- Cart sequencing. Transfer carts feed parts to the assembly line, move in one direction, and can only change direction at set points. Rare unexpected moves, from equipment errors or system-logic limits, break the ordered queues. Restoring them meant stopping the line and recalculating cart routes, which took 10 to 50 minutes depending on the worker’s skill and the line’s complexity.
- Agents in manufacturing. Quality, production, equipment, and logistics each use different systems and data, and turning shop-floor know-how into an agent meant the person also needed separate development knowledge and skills.
- Dealer diagnostics. Technicians checked fault codes and then searched manuals, technical documents, and past cases one by one. Diagnosis was slow, its consistency depended on the technician’s experience, and new models and features took time to reach the service floor.
Confirmed facts: thousands of app reviews a month across Hyundai, Kia, and Genesis apps; manual review, classification, tagging, and reply; 17 languages, 63 categories, 119 tags; standards differ by language and country; reply quality varied by staff; about 600 crash tests a year at Namyang; material organized and stored differently per engineer; search across systems or by asking; know-how lost on departure; VIN check by eye or from system data, per-plant models and servers in the first version; cart queue recovery 10 to 50 minutes with the line stopped; technicians searched manuals and past cases by hand.
3. Solution
3.1 App reviews: agents classify, tag, and draft; people review
The project was led by a manager on the Customer Channel Service Quality Team, Kim Seul-chan (김슬찬). It started from a hypothesis the journal states as a question: can AI classify reviews as accurately as a person, or better? In the proof of concept, classification accuracy reached 85%, above the 80% target.
The production system uses a multi-agent structure. One flow takes a review through classification, tag generation, and reply drafting. The English release lists what the system does:
- “Automatically categorizes reviews by topic and sentiment”
- “Generates appropriate response tags based on review content”
- “Drafts reply text that addresses specific customer concerns”
- “Flags escalation-worthy issues for human review”
According to the Korean release, the AI draws on a structured knowledge database to analyze the review and decide the direction of the response before writing it. A staff member does the final review before the reply goes to the customer. The English release: “Today, personnel can review AI-generated outputs before responding to customers.”
Two mechanisms sit around that flow. First, replies confirmed after staff review are fed back into the AI’s learning, so classification and reply quality keep improving. Second, the system has three-stage guardrails, and cases that are hard for the AI to judge go to a person for final review, which the journal calls a human-in-the-loop setup. The stages themselves are not described.
The Korean release says about 85% or more of all customer reviews were being processed with AI at the time, and that the group planned to automate the entire review response process from September 2026. The English release states the same plan: “The Group plans to automate the entire customer review response process starting in September 2026.”
The solution handles up to 17 languages. The group plans to expand it to global markets, naming Europe first, then North America, Asia-Pacific, and South America. It connects to existing infrastructure rather than requiring a new system. Kim described the next scope: “Going forward, we plan to optimize the whole customer channel with AI infrastructure, not only app review responses but also evaluation of contact-center staff’s service quality, automatic classification of company-wide voice of the customer, and an in-app FAQ chatbot.”
3.2 The Crash Safety AI Assistant: built by the engineer who needed it
Lee Shin-sun (이신선), a senior research engineer on the Safety Performance Advanced Development Team, started from a question the journal quotes: “Couldn’t we solve this problem with the data we already have?” After seeing what generative AI could do, he began building a solution himself with a coding agent the company supports internally.
He did not start with a large system. The first goal was one place to search scattered test results, images, and analysis files together. After that he kept adding fixes for the inconveniences engineers ran into most often.
Three upgrades followed to make the agent perform better:
- Slides as images. Every slide of the existing PowerPoint test reports was converted to an image, so the AI reads text, tables, and graphs together in one multimodal pass.
- A pre-generated wiki. To avoid the gaps of ordinary search, he generated 700 comprehensive analysis reports in advance, one per keyword, in wiki form.
- A 30-page prompt. He wrote his own crash-safety expertise into a 30-page prompt, so the AI would understand context and answer at the level of an experienced crash-safety engineer.
The result has two core functions. An engineer asks a question in natural language and gets an answer built from the internal database, with links to the original reports. And the assistant reads a report and organizes its key questions and answers automatically. The AI also drafts new reports, so the engineer in charge concentrates on reviewing and supplementing them. Test images that are hard to read because of flare or reflections are corrected with AI image generation and kept as usable data.
The Korean release adds what engineers can now look at for similar crashes beyond the documents themselves: the test conditions, vehicle structure, causes of injury, and countermeasures. The assistant runs in a web browser with no installation, so anyone in crash safety can use it. The Korean release dates its introduction at Namyang to the end of 2025.
Lee continues to check the accuracy of answers and the evidence behind them as he develops the system further.
3.3 The other four
- AI Automation Recognition Service. Kwon Oh-hyun (권오현), a manager on the Vision AI Analytics Team, rebuilt the VIN-recognition service in two moves. He trained one model on data from all plants, built to run on any infrastructure, so there was a single point of management. He moved inference to the AWS cloud and built a new API that handles large numbers of simultaneous requests without delay. Cameras on the line read the VIN, and the system compares the vehicle with its registered data in real time. A mismatch can be checked immediately, and processing status per process is monitored in real time.
- Transfer Cart Sequencing Optimization Technology. Park Cheol-woong (박철웅), a senior manager on the Optimization Analytics Team, treated queue recovery as a pathfinding problem and combined an existing search algorithm with reinforcement learning, so the AI finds the optimal move sequence instead of calculating every case. In the side-complete process, which joins the left and right body panels, it was validated with 28 carts, more than twice the 12 used in development, a setting where a sequencing problem has millions of possible cases. One-way moves and branch points at entry and exit were modelled. In an environment simulating the assembly line at Kia’s Hwaseong Plant 2, it found the optimal route within the time limit with 99% probability, taking 3 seconds on average and 7 at most.
- E-FOREST: POLARIS. Lim Yong-taek (임용택), a senior manager on the Manufacturing AI Technology Development Team, built a manufacturing agent ecosystem on seven core technologies, including agent orchestration, Text2SQL, and ontology-based knowledge graphs. Teams in quality, production, equipment, and logistics build their own agents, validate them, and run them as services, with multimodal analysis of images and video. He separates a pilot from production: “These days it isn’t hard to build a pilot agent quickly with AI tools, but developing a service fit for a production environment that hundreds of people use every day is a different level of work.” The platform has to meet internal security standards and connect to “thousands of legacy databases.” It passed Alpha 0.5 in October 2025 and entered Beta 0.7 in June 2026. The plan is Official 1.0 first at the new Ulsan EV plant in the second half of 2026, then all group plants by the second half of 2027.
- Maintenance Support AI Service. Park Seung-hyuk (박승혁), a senior manager on the Global Passenger Car Service Team, began preparing an LLM-based tool before the generative AI boom took off. A technician enters a fault code or a symptom in plain language, such as an oil smell or a grinding noise from the engine, or asks for a procedure such as replacing a high-voltage battery. The AI answers with the cause and countermeasure, drawing on manuals and data. It is linked to the group’s global database of vehicle information, past repair records, and real-time fault data, and can connect to the diagnostic tool’s vehicle-control functions. Diagnosis data built up by dealers and technicians goes back into training.
3.4 The groundwork
The group traces these projects to earlier changes. In 2018, Executive Chair Euisun Chung declared the group would become “a company more IT-like than IT companies themselves.” The Korean release dates the digital transformation program from 2019. From 2022 the group rolled out Jira, Dooray, Confluence, and Microsoft 365. R&D units working on national core technologies such as autonomous driving and hydrogen, which security rules had kept off external collaboration tools, joined the same platform after the government set guidelines. The Global One Data Pipeline standardized and connected data across R&D, production, quality, and sales.
H Chat Pro gives employees secure access to external models including ChatGPT, Gemini, and Claude. It has been open to all general employees since 2025. More employees with no software background are now building their own AI agents with it.
Confirmed facts: PoC hypothesis = AI classifies as accurately as a person or better; 85% accuracy vs 80% target; multi-agent flow classify → tag → draft; structured knowledge database; staff final review before sending; confirmed replies fed back; three-stage guardrails with hard cases to a person; about 85% or more of reviews processed with AI (Korean release); full automation planned from September 2026; up to 17 languages, expansion with Europe named first; crash assistant built by a senior research engineer with an in-house coding agent; unified search first; slides to images, 700 keyword wiki reports, 30-page prompt; Q&A with source links, automatic Q&A extraction, AI-drafted reports reviewed by the engineer, image correction; browser access; introduced end of 2025; single VIN model on all-plant data, inference on AWS; cart sequencing by search plus reinforcement learning; POLARIS Alpha 0.5 (Oct 2025), Beta 0.7 (Jun 2026), 1.0 at Ulsan EV plant (H2 2026), all plants (H2 2027 target); maintenance tool linked to the global database.
4. Impact
All figures are the group’s own, published 2026-08-12. No baseline measurements are published.
| Project | Metric | Figure | Basis |
|---|---|---|---|
| App review responses | Average handling time per review | 35 min → 5 min | Journal; releases say “about 5 minutes” |
| App review responses | PoC classification accuracy | 85% (target 80%) | Proof of concept |
| App review responses | Share of reviews processed with AI | about 85% or more | Korean release, as of 2026-08-12 |
| App review responses | Annual cost saving | about KRW 260 million | Expected, not yet realized |
| Crash Safety AI Assistant | Questions asked | about 2,000 in 4 months | Journal |
| Crash Safety AI Assistant | Time to find related cases and review analysis material | about 90% less | “compared with the previous workflow” |
| VIN recognition | Annual cost saving | about KRW 5.24 billion (about $3.9 million) | Group-wide, from fewer unnecessary stoppages |
| VIN recognition | Deployment | about 70 points | Journal: operating points in plants; Korean release: production bases; English release: manufacturing processes |
| Cart sequencing | Unnecessary work stoppage time | about 86% less | Applied on a real line |
| Cart sequencing | Annual cost saving | about KRW 40 million | Per single process |
| Cart sequencing | Optimal route found in time limit | 99%; 3 s average, 7 s max | Simulated Kia Hwaseong Plant 2 line |
| Maintenance support | Maintenance response time | down 42% or more | Journal; releases say “about 42%” |
| H Chat Pro | Users | 30,000+, about 80% of Hyundai Motor and Kia general and research staff | July 2026 |
On the review work, both releases say reply quality also improved, without a figure. The journal credits the 35-to-5 drop to automating every step from classification to the reply, not just positive-or-negative sorting. Staff time freed by fewer repetitive tasks goes to planning and analysis, in the English release’s words “customer experience analysis and service improvement.”
For crash safety, Lee described the change: “In the past we often relied on experienced engineers’ past experience to make judgments, but now we can verify quickly based on data. That is the most practical change we feel in the work.” The freed time goes to reviewing analysis results, improving designs, and planning test strategy.
The maintenance tool started with US dealers in December 2025 and expanded from June 2026 to Europe, Asia-Pacific, Latin America, Africa, and the Middle East. The plan is to serve more than 6,000 dealers and 40,000 technicians by the end of 2026. The journal also reports higher initial diagnosis accuracy and rising customer satisfaction, without figures.
Eunsook Jin, President and Head of ICT Management Division, set the frame at the event: “AI is undoubtedly an important technology, but ultimately it is a tool that companies must utilize. What matters is not the technology itself, but how quickly an organization can understand new technologies, including AI, apply them to its work and business, and internalize them as a source of competitive advantage.”
5. Insight
In review handling, the AI took over the steps and people kept the check. Before, one person read, classified, tagged, and answered each review. Now agents do all four, and staff review the output before it is sent, with hard cases routed to them by the guardrails. The plan to automate the entire process from September 2026 suggests the next step: staff reviewing exceptions instead of every reply. The group has not reported whether that step happened.
The taxonomy was the hard part, and it became the test. A scheme of 17 languages, 63 categories, and 119 tags is what made the manual work slow and uneven. The project’s first question was whether AI could apply that scheme as well as people, and it set an accuracy bar (80%) before building. The production design then feeds staff-confirmed replies back in, so each review a person approves also trains the system.
The crash assistant was built by the person whose knowledge it holds. A senior research engineer, not an IT team, built it with an in-house coding agent and wrote his own expertise into a 30-page prompt. The work that made it useful was preparation: turning slides into images so tables and graphs could be read, and writing 700 keyword reports in advance. What used to leave with a retiring engineer now sits in the wiki, the prompt, and the linked reports.
Scaling a working model was an operations problem. The first VIN service worked but ran on a separate model and server per plant and line. The fix was one model trained on all plants’ data and one cloud inference service, and the service now runs at about 70 points worldwide.
Limits. All figures are the group’s own and none has a stated baseline. The two 85% figures for review handling measure different things: classification accuracy in the proof of concept, and the share of reviews processed with AI. The KRW 260 million saving is an expectation. The three guardrail stages are not described. What “about 70” counts differs across the three sources. The cart optimizer’s 99% and 3-second figures come from a simulated line; the 86% and KRW 40 million come from one real process. POLARIS has no reported results yet.
One-line takeaway: In Hyundai Motor Group’s two clearest projects, the team that owned the work set the test, AI took over classification, search, and drafting, and people kept the review, with what they approve feeding back into the system.
6. Implementation hypotheses (not in the source)
6.1 Review routing: taxonomy check, confidence, and escalation topics
Basis: A multi-agent flow classifies, tags, and drafts; there are “three-stage guardrails” and cases hard for the AI go to a person; the English release says the system “Flags escalation-worthy issues for human review”; about 85% or more of reviews were processed with AI, and full automation was planned from September 2026.
A likely reading of the three stages follows the three agent steps. After classification, a first check would confirm that the category and tags exist in the 63-category, 119-tag scheme for that language and country, and that the classifier’s confidence clears a threshold; low-confidence reviews go to staff. A second check would look at the content for topics that should not get an automatic reply, such as safety complaints, legal threats, or refund and warranty disputes, and mark them for escalation. A third would check the draft itself: correct language, no promises about repairs or compensation, and consistency with the knowledge base entry it drew on. The “about 85% or more” figure would then be the share that passes all three, and the September plan would mean sending those replies without a per-review approval while the rest stay in the staff queue.
6.2 Feedback from approved replies
Basis: “Replies confirmed after staff review” are fed back into AI learning; the Korean release says the AI uses a structured knowledge database to decide the response direction; the PoC tested whether AI could classify as accurately as people, against an 80% target.
The cheapest way to use approved replies is probably retrieval rather than retraining: each approved reply, with its category, tags, language, and any staff edits, is stored and pulled in as an example for the next review of the same type. Staff edits to the draft would be the most useful signal, since they mark where the draft was wrong. The PoC accuracy was most likely measured against reviews that staff had already classified by hand, and the same comparison, AI label against confirmed label, would let the team keep tracking accuracy per category and language after launch.
6.3 Crash assistant: wiki first, slides as evidence
Basis: Report slides were converted to images for multimodal reading; 700 keyword reports were pre-generated “to avoid information missed by ordinary search”; answers come with links to the original reports; a 30-page prompt carries the engineer’s expertise; employees build agents with models available through H Chat Pro.
The assistant most likely answers in two layers. The 700 wiki reports, each summarizing all tests that touch one keyword, would be searched first, because they already gather results that a keyword search over individual slide decks would scatter or miss. The relevant slide images from the original reports would then be retrieved and read to support the answer, and their report links returned so the engineer can open the source. The 30-page prompt would sit as fixed instructions on every call. The wiki would need a rebuild step when new test reports arrive, likely a batch job that re-reads new decks, updates the affected keyword reports, and runs the automatic question-and-answer extraction on each new report. Model calls probably go through the group’s internal platform, since H Chat Pro is where employees reach external models in a secured environment.
7. Sources
- Hyundai Motor Group, HMG Journal — 현대차그룹이 AI를 경쟁력으로 만드는 방법 (How Hyundai Motor Group turns AI into competitiveness)
https://www.hyundaimotorgroup.com/ko/story/hyundai-motor-group-ai-competitiveness
2026-08-12 (accessed 2026-11-17). Korean-language article; quotes from it are translated from Korean. - Hyundai Motor Group, press release (Korean) — 현대차그룹, 전사 DX·AX 여정과 성과 공개… 선제적 DX 완료로 전 사업 영역 AI 혁신
https://www.hyundaimotorgroup.com/ko/news/hyundai-motor-group-dx-ax-ai-innovation
2026-08-12 (accessed 2026-11-17) - Hyundai Motor Group, press release (English) — Hyundai Motor Group Accelerates AI Transformation Across Its Business, Advancing Toward the Physical AI Era
https://www.hyundaimotorgroup.com/en/news/hyundai-motor-group-accelerates-ai-transformation-across-its-business-advancing-toward-the-physical-ai-era
2026-08-12 (accessed 2026-11-17)