AutoMeet: a proof-of-concept study of genAI to automate meetings in automotive engineering

Max Radyschevski; Simon Baeuerle; Ulrike Pado

REVIEW 2 major objections 6 minor 6 references

AutoMeet: a proof-of-concept study of genAI to automate meetings in automotive engineering

T0 review · 2 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A genAI meeting pipeline tested with automotive engineers is estimated to raise documented-meeting coverage from 38.5% to 54.7% and recover about 10.5% of working time per employee, provided privacy safeguards are in place.

desk verdict A clear-eyed PoC study whose qualitative findings are worth taking seriously, but the 10.5% savings headline is an unvalidated upper bound, not an estimate. read the letter →

arxiv 2507.16054 v1 pith:ZPGRNX3T submitted 2025-07-21 cs.CL

Simon Baeuerle , Max Radyschevski , Ulrike Pado This is my paper · ORCID

classification cs.CL

keywords ethicsknowledgemanagementgenAImeetingnotessummarizationRetrieval-AugmentedGenerationuseracceptanceautomotiveengineering

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that generative AI can automate the entire meeting documentation workflow in an engineering department, and that the main obstacle is organizational trust and data protection rather than model quality. To support this, the authors built two proof-of-concept tools—an automated meeting summarizer and a searchable chatbot—and gathered survey feedback from development engineers who tried them on real meetings. The quantitative headline is that a fully integrated AutoMeet pipeline would raise the share of meetings with created minutes from about 38.5% to 54.7% and save about 10.5% of each employee's working time. A sympathetic reader would care because this is one of the few user-grounded estimates of what enterprise AI meeting tools could actually deliver, with acceptance depending on features users explicitly asked for.

What carries the argument

The load-bearing mechanism is the AutoMeet end-to-end pipeline, defined as the sequence of meeting recording, automated transcription, LLM-based summarization of both the transcript and supplementary documents, a manual privacy-filter and editing step, storage of approved minutes with metadata in a central place, and a chatbot that uses retrieval-augmented generation to answer queries from the stored documents. The pipeline carries the argument because the survey questions about cancellable meetings, documentation coverage, and privacy requirements were anchored to demonstrations of these tools; without the pipeline functioning end to end, the savings and acceptance estimates would have nothing to attach to.

What would settle it

A controlled field deployment would settle the quantitative claim: track a department's actual meeting hours and the rate of created minutes before and after full AutoMeet integration, and compare the observed change with the predicted rise from 38.5% to 54.7% and the predicted 10.5% time saving. If the measured reduction in meeting time is close to zero, or documented minutes do not rise, the central estimate fails.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that an end-to-end AutoMeet pipeline—recording meetings, transcribing them, summarizing with a prompted large language model, applying a manual privacy filter, and making the minutes searchable through a retrieval-augmented-generation chatbot—is technically feasible today and would be accepted by engineers if the right privacy controls exist. In the survey, engineers estimated that full pipeline use would raise documented meetings from 38.5% to 54.7% and that about 20.6% of their weekly meetings could have been cancelled, leading to an estimated 10.5% saving of working time per employee. The paper also reports that technical transcription and summarization quality were not the main concern: 33 of 35 respondents said they would use such a tool if data-security concerns were addressed, and the most-valued features were easy start/stop of recording, automated removal of personal data, manual approval of minutes, and deletion of recordings after processing.

Load-bearing premise

The load-bearing premise is that engineers' self-reported estimates of cancellable meetings—about 20.6% on average—and of time saved reflect real behavior after full AutoMeet use, including the assumption that cancelled meetings convert into productive working time.

Editorial extensions

If this is right

If a fully integrated AutoMeet pipeline is deployed, the share of meetings with minutes is expected to rise from about 38.5% to 54.7%, giving engineers searchable records for more than half of their meetings.
Meeting-related working time could fall by about 10.5% per employee, a saving that scales with hourly rates and headcount.
User acceptance turns on data-protection controls—easy start/stop recording, automatic personal-data removal, manual approval before storage, and deletion of recordings—not on further improvements to transcription or summarization quality.
The pipeline is model-agnostic: newer LLMs can replace the current summarizer without changing the rest of the system, so future model improvements flow directly into better minutes.
Briefing users about possible model errors such as hallucinations raises acceptance and their willingness to correct small mistakes, making AI-literacy training part of any deployment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A consequence the paper leaves implicit is that the same survey-and-demo method could be transferred to other knowledge-intensive sectors with heavy meeting loads, though the specific percentages would need to be re-measured.
The 10.5% saving is a self-reported estimate, not a measured time-use outcome; a deployment study that tracks calendars and actual meeting hours before and after integration would show whether cancelled meetings become productive time or simply shift work to other channels.
Because manual approval and deletion were rated as must-haves, the design implication is that the system should be framed as a human-controlled documentation aid, with recording used only when needed and minutes owned by users, rather than as an autonomous recorder.
The RAG-based chatbot could function as a project memory whose answers cite source and timestamp; a natural next test would ask whether engineers act on those answers in decisions, not just whether they find them credible.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Read it. The survey data are genuinely new and the authors are transparent about what they did. Forty-two automotive R&D engineers give concrete numbers: 51.2% working time in meetings, 38.5% minutes coverage, 20.6% of meetings could have been cancelled with full AutoMeet use. The privacy feature ratings (easy start/stop 1.76, automated personal data removal 1.87) and the qualitative finding that organizational and data-protection factors, not transcription quality, determine acceptance are useful for anyone building or buying meeting AI.

The technical pipeline is standard—Whisper, GPT-4o, LangChain, RAG—and the authors admit it. The PoC tools are demos, not deployable systems; Section 5 says so explicitly. That honesty is good, but it also means the quantitative headline overreaches. The 10.5% savings in Section 4.1 is the product of two survey averages, with no uncertainty propagation, no individual-level linkage, and no evidence that cancelled meetings convert into productive time. The stress-test note is right: cancelling a meeting still leaves the information-transfer task; someone has to read the minutes. The paper never subtracts that replacement time. So 10.5% is better described as an optimistic upper bound.

The qualitative central claim survives the soft spot. The survey design does have mild circularity: participants saw the demo before estimating cancellable meetings, which could bias them toward optimism, but that affects the size of the number, not the direction of the qualitative conclusions.

Other nits: sampling frame is thin (how were the 42 chosen? from how many departments? response rate?), the raw survey data are not released, and one key readiness question has n=17. All fixable.

Bottom line: this deserves a serious referee—the empirical data are new, the reporting is transparent, and the limitations are acknowledged. The revision should reframe the savings as an expectation with explicit error bars, describe the sampling frame, and release the instrument and anonymized responses. The qualitative findings will hold. For my own work, I would not cite the 10.5% number, but I would cite the survey results on privacy feature priorities and the minutes-coverage estimates.

Yes, send it to peer review. The core is honest and useful.

Referee Report

2 major / 6 minor

Summary. The paper presents AutoMeet, a proof-of-concept pipeline for automating meeting documentation in automotive engineering departments. The system records meetings, transcribes them with Whisper, generates short and long summaries with GPT-4o, applies a manual privacy and quality filter, and offers a retrieval-augmented chatbot over the resulting minutes. The authors deployed the two central PoC components in automotive R&D departments and collected survey responses from 42 engineers. Based on the survey, they report that meetings carry roughly 40% of information flow, that a fully integrated pipeline could raise the documented-meeting share from 38.5% to 54.7%, that 20.6% of meetings could be cancelled, and that this implies approximately 10.5% saved working time per employee. The paper's qualitative central claim is that organizational and data-protection factors, not technical summarization quality, determine user acceptance.

Significance. The study's main strength is its real-world, user-centered evaluation: it implements actual PoC tools, reports per-question response counts and standard deviations, and derives a concrete feature list for future systems. The qualitative finding that privacy safeguards, manual approval, and organizational trust dominate acceptance is valuable and actionable. However, the headline quantitative benefit of 10.5% was directly critical reading of the manuscript: it assumes that cancellable meetings convert fully into productive time, ignores the time needed to consume the generated minutes, and is computed from self-reported hypothetical estimates without uncertainty propagation. If reframed as an optimistic upper bound under stated assumptions, the qualitative contribution remains solid, but the current presentation overstates the evidence for realized savings.

major comments (2)

[§4.1] The 10.5% savings estimate is computed as the product of the group mean meeting-time share (51.22%, n=41, StdDev 18.90) and the group mean cancellable-meeting share (20.57%, n=35, StdDev 14.13). This product of means is not the mean of individual products, and the two survey items have different response counts and large dispersions. The paper gives no confidence interval or other uncertainty measure for the central quantitative claim, so the reader cannot assess whether the reported saving is distinguishable from zero or from a much smaller value. The authors should either compute individual-level savings if the paired responses are available, or report a proper uncertainty interval for the product of means.
[§4.1 and §5] Even if the two percentages were accepted as population means, 'meetings that could have been cancelled' is a count-based estimate, not a time-based estimate, and cancelling a meeting does not eliminate the information-transfer task. With AutoMeet, participants would still need to read or search the generated minutes, so the actual saving is at most meeting-time share × cancellable-meeting share × (1 − replacement-time fraction). The paper neither measures nor assumes this replacement fraction, and §5 explicitly states that the tooling is a proof-of-concept not deployed for daily use. The 10.5% figure is therefore an optimistic upper bound under the additional assumption that cancelled meetings convert fully into productive time, and the survey design (estimates provided after viewing the demo) may introduce optimism bias. This should be stated explicitly wherever the figure appears.

minor comments (6)

[§4.1, Figure 6] The text says '35 out of 36 participants responded that information would be more transparent,' but the figure reports n=37 with Avg 0.03, which implies 36 of 37 answered 'yes.' Please reconcile the reported counts.
[§4.1, Figure 5] The comparison of 'current' 38.50% (n=40) with 'AutoMeet' 54.72% (n=36) uses different response counts; reporting a paired analysis on the subset of respondents who answered both questions would strengthen the claim that the proportion of documented meetings increases.
[§2, references] The reference list contains two entries, [FFQ22a] and [FFQ22b], that appear to be the same paper; the duplicate should be merged or the two entries should be clearly differentiated.
[Abstract and author line] The author line contains 'und' (German for 'and') and should use the English conjunction for consistency with the rest of the manuscript.
[§3.1 and Figure 1] The text and figure captions contain typos such as 'Manuel meeting minutes' (should be 'Manual'), and Figure 6 has an extra space in 'T ake a quick look.'
[§6] The concluding sentence 'highlight features which a necessary for a successful implementation' is missing the verb 'are'; please correct the grammar.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the savings estimate is transparent arithmetic on survey means, and the paper contains no self-citations or fitted parameters relabeled as predictions.

full rationale

The paper's central quantitative claim, the estimated 10.5% of saved working time, is explicitly computed from two survey averages: 51.22% meeting-time share and 20.57% cancellable meetings. This is not a hidden reduction or a fitted parameter called a prediction; the paper states the calculation directly, and the result is exactly the product of the reported means. The survey-based nature of the inputs and the lack of uncertainty propagation are validity concerns, not circularity. There are no self-citations in the reference list, no uniqueness theorem imported from the authors' prior work, and no ansatz smuggled in via citation. The qualitative conclusions about organizational and privacy concerns are derived from open-ended survey responses and are independent of the quantitative estimate. The proof-of-concept limitations acknowledged in Section 5 are honest scope statements rather than circularity. Overall, the derivation chain is self-contained: claims are either direct survey findings or transparent arithmetic on those findings.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central quantified result (10.5% working-time savings) is computed as the product of survey means that are treated as predictive measurements even though they are self-reported expectations (Section 4.1). The paper also assumes transcription and LLM summary quality are sufficient for the workflow despite WER 0.23 to 0.65 and a mean summary rating of 2.58 on a 1 to 6 scale, and it assumes future LLMs will match or exceed GPT-4o quality. No new scientific entities are introduced; AutoMeet is a software architecture, not an invented physical or formal entity.

free parameters (3)

meeting_time_share = 51.22% (survey mean, StdDev 18.90, n=41)
Self-reported share of working time spent in meetings; multiplied into the 10.5% savings estimate without propagating its standard deviation.
cancellable_meeting_share = 20.57% (survey mean, StdDev 14.13, n=35)
Self-reported hypothetical estimate of meetings that could be cancelled under full AutoMeet usage; the other factor in the 10.5% savings estimate. No validation against actual deployment.
future_minutes_coverage = 54.7% (survey mean, StdDev 24.78, n=36)
Estimated future share of meetings with minutes under AutoMeet, used to argue an improvement over the current 38.5% share.

assumptions (3)

domain assumption Transcription quality is adequate for producing usable minutes despite Word Error Rates of 0.23 to 0.65.
Invoked implicitly in Section 3.1 and Section 5, where transcription errors such as 'centering' for 'sintering' are reported as user-visible drawbacks; the pipeline's value depends on transcription being good enough.
domain assumption GPT-4o summary quality is representative of what future LLMs will deliver, so current limitations are treated as a user-acceptance issue.
Stated in Section 5: 'future LLM models can be used as drop-in replacement without any changes to the remaining pipeline.'
domain assumption The 42 survey respondents represent the broader automotive R&D engineer population.
Sampling was voluntary and the response base is not described (how many were invited, which departments, drop-out rates); per-question n drops from 42 to 17 for tool-usage readiness.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AutoMeet: a proof-of-concept study of genAI to automate meetings in automotive engineering." pith.science (2026). https://pith.science/paper/ZPGRNX3T

@misc{pith2026250716054,
  author       = {Pith},
  title        = {Pith review of: AutoMeet: a proof-of-concept study of genAI to automate meetings in automotive engineering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZPGRNX3T}},
  note         = {Machine review of arXiv:2507.16054}
}

read the original abstract

In large organisations, knowledge is mainly shared in meetings, which takes up significant amounts of work time. Additionally, frequent in-person meetings produce inconsistent documentation -- official minutes, personal notes, presentations may or may not exist. Shared information therefore becomes hard to retrieve outside of the meeting, necessitating lengthy updates and high-frequency meeting schedules. Generative Artificial Intelligence (genAI) models like Large Language Models (LLMs) exhibit an impressive performance on spoken and written language processing. This motivates a practical usage of genAI for knowledge management in engineering departments: using genAI for transcribing meetings and integrating heterogeneous additional information sources into an easily usable format for ad-hoc searches. We implement an end-to-end pipeline to automate the entire meeting documentation workflow in a proof-of-concept state: meetings are recorded and minutes are created by genAI. These are further made easily searchable through a chatbot interface. The core of our work is to test this genAI-based software tooling in a real-world engineering department and collect extensive survey data on both ethical and technical aspects. Direct feedback from this real-world setup points out both opportunities and risks: a) users agree that the effort for meetings could be significantly reduced with the help of genAI models, b) technical aspects are largely solved already, c) organizational aspects are crucial for a successful ethical usage of such a system.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

6 extracted references · 4 canonical work pages

[1]

Additionally, frequent in-person meetings produce inconsistent documentation – official minutes, personal notes, presentations may or may not exist

AutoMeet: a proof-of-concept study of genAI to automate meetings in automotive engineering Simon Baeuerle1,2, Max Radyschevski2,3 und Ulrike Pado3 Abstract: In large organisations, knowledge is mainly shared in meetings, which takes up significant amounts of work time. Additionally, frequent in-person meetings produce inconsistent documentation – official...

arXiv 2025
[2]

we need this as soon as possible

The featureEasy starting/stopping of recordingwas rated 1.76 on average (StdDev 1.48, n=38),Automatedremovalofpersonaldata wasrated1.87(StdDev1.45,n=38), Obligatory manual approval of minutes before integration to databasewas rated 2.58 (StdDev 1.66, n=38), Deletion of recording after processingwas rated 2.58 (StdDev 1.73, n=38). Privacy feature Importanc...

work page 2021
[6]

[STW25] Schneider, F.; Turchi, M.; Waibel, A.: Policies and Evaluation for Online Meeting Summarization, 2025, arXiv: 2502.03111 [cs.CL], url: https: //arxiv.org/abs/2502.03111. [Tk24] Tkalac Verčič, A.; Verčič, D.; Čož, S.; Špoljarić, A.: A systematic review of digitalinternalcommunication.PublicRelationsReview50/1,S.102400,2024, issn: 0363-8111, url: ht...

work page Pith review arXiv 2025
[2023]

6246–6261, 2023,url: https: //aclanthology.org/2023.findings-emnlp.413/

Association for Computational Linguistics, Singapore, S. 6246–6261, 2023,url: https: //aclanthology.org/2023.findings-emnlp.413/. [Lü24] Lünendonk®-Studie: Generative AI – Von der Innovation bis zur Marktreife. Wo stehen Unternehmen im deutschsprachigen Raum bei der Nutzung von Generative AI?, Techn. Ber., Lünendonk & Hossenfelder GmbH,

work page 2023
[2024]

Transactions of the Association for Computational Linguistics 11/, S

[Re23] Rennard, V.; Shang, G.; Hunter, J.; Vazirgiannis, M.: Abstractive Meeting Summarization: A Survey. Transactions of the Association for Computational Linguistics 11/, S. 861–884, 2023,url: https://aclanthology.org/2023.tacl- 1.49/. [Ri22] Riedl, R.: On the stress potential of videoconferencing: definition and root causes of Zoom fatigue. Electronic ...

doi:10.1007/s12525-021-00501-3 2023
[2025]

In (Dernoncourt, F.; Preoțiuc-Pietro, D.; Shimorina, A., Hrsg.): Proceedingsofthe2024ConferenceonEmpiricalMethodsinNaturalLanguage Processing: Industry Track

[Ki24] Kirstein, F.; Ruas, T.; Kratel, R.; Gipp, B.: Tell me what I need to know: Exploring LLM-based (Personalized) Abstractive Multi-Source Meeting Sum- marization. In (Dernoncourt, F.; Preoțiuc-Pietro, D.; Shimorina, A., Hrsg.): Proceedingsofthe2024ConferenceonEmpiricalMethodsinNaturalLanguage Processing: Industry Track. Association for Computational L...

work page 2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Additionally, frequent in-person meetings produce inconsistent documentation – official minutes, personal notes, presentations may or may not exist

AutoMeet: a proof-of-concept study of genAI to automate meetings in automotive engineering Simon Baeuerle1,2, Max Radyschevski2,3 und Ulrike Pado3 Abstract: In large organisations, knowledge is mainly shared in meetings, which takes up significant amounts of work time. Additionally, frequent in-person meetings produce inconsistent documentation – official...

arXiv 2025

[2] [2]

we need this as soon as possible

The featureEasy starting/stopping of recordingwas rated 1.76 on average (StdDev 1.48, n=38),Automatedremovalofpersonaldata wasrated1.87(StdDev1.45,n=38), Obligatory manual approval of minutes before integration to databasewas rated 2.58 (StdDev 1.66, n=38), Deletion of recording after processingwas rated 2.58 (StdDev 1.73, n=38). Privacy feature Importanc...

work page 2021

[3] [6]

[STW25] Schneider, F.; Turchi, M.; Waibel, A.: Policies and Evaluation for Online Meeting Summarization, 2025, arXiv: 2502.03111 [cs.CL], url: https: //arxiv.org/abs/2502.03111. [Tk24] Tkalac Verčič, A.; Verčič, D.; Čož, S.; Špoljarić, A.: A systematic review of digitalinternalcommunication.PublicRelationsReview50/1,S.102400,2024, issn: 0363-8111, url: ht...

work page Pith review arXiv 2025

[4] [2023]

6246–6261, 2023,url: https: //aclanthology.org/2023.findings-emnlp.413/

Association for Computational Linguistics, Singapore, S. 6246–6261, 2023,url: https: //aclanthology.org/2023.findings-emnlp.413/. [Lü24] Lünendonk®-Studie: Generative AI – Von der Innovation bis zur Marktreife. Wo stehen Unternehmen im deutschsprachigen Raum bei der Nutzung von Generative AI?, Techn. Ber., Lünendonk & Hossenfelder GmbH,

work page 2023

[5] [2024]

Transactions of the Association for Computational Linguistics 11/, S

[Re23] Rennard, V.; Shang, G.; Hunter, J.; Vazirgiannis, M.: Abstractive Meeting Summarization: A Survey. Transactions of the Association for Computational Linguistics 11/, S. 861–884, 2023,url: https://aclanthology.org/2023.tacl- 1.49/. [Ri22] Riedl, R.: On the stress potential of videoconferencing: definition and root causes of Zoom fatigue. Electronic ...

doi:10.1007/s12525-021-00501-3 2023

[6] [2025]

In (Dernoncourt, F.; Preoțiuc-Pietro, D.; Shimorina, A., Hrsg.): Proceedingsofthe2024ConferenceonEmpiricalMethodsinNaturalLanguage Processing: Industry Track

[Ki24] Kirstein, F.; Ruas, T.; Kratel, R.; Gipp, B.: Tell me what I need to know: Exploring LLM-based (Personalized) Abstractive Multi-Source Meeting Sum- marization. In (Dernoncourt, F.; Preoțiuc-Pietro, D.; Shimorina, A., Hrsg.): Proceedingsofthe2024ConferenceonEmpiricalMethodsinNaturalLanguage Processing: Industry Track. Association for Computational L...

work page 2024