企业对AI代理自主生产变更信任度降至56%

亿邦AI 2026-10-09 10:16

亿邦AI 2026/10/09 10:16

邦小白快读

全文速览

总1：AI代理自主生产变更的信任度明显下降。

1. 支持无人工审核生产变更或计划落地的企业占比从7月的75%降至8月的56%；单独统计核心决策群体，也从88%降至61%。

2. 当前27%的企业已放开低风险场景的无审核变更权限，35%明确不会放开，20%计划12个月内搭建能力。

总2：自动化评估并不完全可靠，存在真实故障风险。

1. 61%做过部署前评估的企业曾遇到AI通过内部测试、上线后引发客户故障，其中22%遇到过不止一次。

2. 仅9%的企业认为现有自动化评估结果可以直接使用；最大顾虑是结果无法匹配真实场景（27%），其次是缺乏可解释性（24%）、评估偏差（16%）、数据隐私（13%）。

总3：部署AI代理需要重视实时监控。

1. 仅29%的自主AI部署企业将实时质量检查作为主要监控，36%依赖交易日志，无法实时判断输出正误。

2. 即使已放开无审核权限的企业，也只有26%做实时输出质量检查；形式正确但内容有误的输出大概率不会告警。

总1：AI代理自主变更对品牌体验构成质量风险。

1. 61%的企业遇到过AI通过内部测试后上线引发面向客户故障的情况，品牌若依赖无人工审核的自主变更，可能导致消费者直接面对错误内容，伤害品牌信任。

2. 当前仅9%企业认为自动化评估结果足够可靠，品牌在涉及用户交互的环节不宜把审批权完全交给AI。

总2：品牌商应重点投入人工审核与可观测性来守护品控。

1. 未来一年可靠性投入增速最高的领域是人工审核工作流，30%受访者首选；生产可观测性工具以26%紧随其后。

2. 目前部署自主AI代理的企业中，只有29%把实时自动化质量检查作为主要监控手段，多数靠日志追踪，品牌需要补齐实时内容校验能力。

总3：工具选择与用户信任建设趋势。

1. OpenAI开发者平台企业渗透率达59%，Confident AI为36%，Braintrust为23%，Anthropic为16%，LangSmith为13%，品牌评估AI系统时可参考主流工具。

2. 62%企业计划12个月内新增或替换评估平台，说明品牌方正在加强AI审核能力，以适应更审慎的AI应用趋势。

总1：AI自主生产变更趋势收紧，谨慎部署成为主流。

1. 允许或计划放开无人工审核生产变更的企业占比从75%降到56%，核心决策者支持度从88%降到61%，说明卖家在选品、运营等环节引入AI时不宜激进。

2. 明确保留人工审核的企业比例从20%翻倍至42%，人工介入仍然是现阶段保障质量的关键。

总2：自动化评估工具市场快速增长，潜藏商业机会。

1. OpenAI原生评估工具的企业使用率从7月31%升至8月59%；62%受访企业计划12个月内新增、替换或补充评估工具，其中35%计划3个月内完成。

2. Confident AI、Braintrust、Anthropic、LangSmith等平台均有渗透，卖家可关注与这些工具相关的服务或代运营合作。

总3：风险提示与应对措施。

1. 61%有过评估经验的企业遇到AI测试通过但上线故障，只有9%信任自动化评估结果，卖家上线AI功能前要做真实场景测试。

2. 监控存在明显缺口，仅29%企业做实时质量检查；卖家应优先投入人工审核和可观测性，避免面向客户的AI事故。

总1：生产环节可探索AI代理的低风险场景自主变更。

1. 目前27%受访企业已允许低风险场景无人工审核的AI代理执行生产变更，20%计划12个月内搭建能力，35%明确不放开；工厂可参照此比例评估自身风险承受度。

2. 剔除未部署自主AI代理的样本后，32%已放开低风险无审核权限，24%正在搭建，42%保留人工审核。

总2：生产监控和数据校验仍是薄弱环节，是数字化转型的机会点。

1. 仅29%的自主AI部署企业将实时自动化质量检查作为主要监控，36%依赖交易追踪日志，16%只做API网关基础设施追踪，无法判断输出内容是否正确。

2. 实时输出质量检查在已放开权限企业中占比也仅26%，形式正确但内容错误不会被告警，制造企业在引入AI时应优先建设内容正确性校验。

总3：自动化评估工具需求增加，可对接供应商。

1. OpenAI评估工具使用率从31%升至59%，但自动化评估信任度仅9%，主要顾虑是场景匹配度、可解释性和偏差。

2. 未来一年人工审核工作流是投入增长最快的可靠性领域（30%），生产可观测性以26%排第二，工厂可通过采购或合作补齐这些能力。

总1：客户核心痛点来自自动化评估不可靠。

1. 仅9%企业认为自动化评估结果足够可靠，27%认为结果无法匹配真实场景，24%受困于可解释性，16%遇到评估偏差，13%担心数据隐私，12%认为成熟度不足。

2. 61%做过部署前评估的客户遇到AI通过测试但上线后引发客户故障，其中22%不止一次。

总2：解决方案需求集中在人工审核、可观测性和实时质量检查。

1. 客户未来一年可靠性投入增速最高的领域依次是人工审核工作流（30%）、生产可观测性工具（26%）、自动化评估流水线（21%）。

2. 监控缺口明显：仅29%客户采用实时自动化质量检查，36%依赖日志追踪，服务商可提供实时输出正确性校验和可解释性增强方案。

总3：评估平台市场快速变动，存在切入机会。

1. OpenAI开发者平台企业渗透率59%（主要平台占40%），Confident AI渗透率36%（主要10%），Braintrust 23%（主要12%），Anthropic 16%，LangSmith 13%。

2. 62%客户计划12个月内新增、替换或补充评估平台，其中35%计划3个月内调整；服务商可以帮助客户铺设自动化评估流水线，并设计无人工审批的权限边界。

总1：企业评估平台使用现状和调整意愿。

1. 主要评估平台中OpenAI占40%，Confident AI占10%，Braintrust占12%，Anthropic占10%，LangSmith占2%；企业渗透率方面OpenAI达59%，Confident AI 36%，Braintrust 23%，Anthropic 16%，LangSmith 13%。

2. 62%企业计划12个月内新增、替换评估平台或补充工具，其中35%计划3个月内完成，平台存在增长与替换机会。

总2：企业核心诉求是提升评估可靠性和可解释性。

1. 自动化评估的核心顾虑中，27%无法匹配真实场景，24%缺乏可解释性，16%遇到偏差或不一致，13%担心数据隐私，平台需要针对性优化。

2. 61%企业经历过AI通过内部测试后上线故障，平台可提供更接近生产环境的评估测试和实时质量检查能力。

总3：平台应帮助客户管理无人工审核的权限边界。

1. 企业正在明确区分使用自动化评估工具和将评估结果作为生产变更唯一审批依据两种场景，平台可推出分级审批、风险场景识别等功能。

2. 当前仅29%自主AI部署企业将实时质量检查作为主要监控，平台上可加强输出内容正确性校验，避免形式合规但内容错误的输出通过。

总1：产业新动向：企业从激进放开AI代理权限转向加强评估。

1. 允许或计划无人工审核生产变更的企业占比从7月75%降至8月56%，核心决策者从88%降至61%，显示AI代理自主权限进入收缩期。

2. 同期OpenAI原生评估工具使用率从31%大幅升至59%，说明企业将资金和精力转向评估能力建设。

总2：新问题：自动化评估信任低、生产监控存在结构性缺口。

1. 仅9%受访者认为自动化评估结果足够可靠，主要顾虑为场景匹配（27%）、可解释性（24%）、偏差不一致（16%）、隐私（13%）、成熟度不足（12%）。

2. 部署自主AI代理的企业中，实时自动化质量检查仅占29%，36%依赖交易日志，16%依赖API网关，无法判断输出内容正确性。

总3：治理启示：人工监督仍是AI生产变更的必要环节。

1. 42%的已部署自主代理企业计划保留人工审核，投入增速首位是人工审核工作流（30%），显示人在回路仍是主要风控模式。

2. 调查时间早于Anthropic CEO公开发文呼吁放缓前沿AI开发的节点，后续行业高管表态可能进一步强化审慎趋势，监管可关注无人工审核变更的授权边界与问责机制。

返回默认

声明：快读内容全程由AI生成，请注意甄别信息。如您发现问题，请发送邮件至 run@ebrun.com 。

我是品牌商卖家工厂服务商平台商研究者帮我再读一遍。

Quick Summary

Summary 1: Trust in AI agents' autonomous production changes has declined noticeably.

1. The share of companies supporting production changes or planned rollouts without human review fell from 75% in July to 56% in August; among core decision-makers alone, it dropped from 88% to 61%.

2. Currently, 27% of companies have enabled approval-free changes in low-risk scenarios, 35% explicitly will not enable them, and 20% plan to build this capability within 12 months.

Summary 2: Automated evaluation is not fully reliable and carries real risk of production failures.

1. Of companies that conducted pre-deployment evaluations, 61% have encountered AI that passed internal tests but caused customer-facing incidents after launch; 22% experienced this more than once.

2. Only 9% of companies believe existing automated evaluation results can be used directly; the biggest concern is that results do not match real-world scenarios (27%), followed by lack of explainability (24%), evaluation bias (16%), and data privacy (13%).

Summary 3: Deploying AI agents requires attention to real-time monitoring.

1. Only 29% of companies with autonomous AI deployments use real-time quality checks as their primary monitoring method, while 36% rely on transaction logs and cannot judge output correctness in real time.

2. Even among companies that have enabled approval-free changes, only 26% perform real-time output quality checks; outputs that are well-formed but wrong in content will likely not trigger alerts.

Summary 1: Autonomous AI changes pose quality risks to brand experience.

1. 61% of companies have experienced AI that passed internal tests but caused customer-facing failures after launch. If brands rely on autonomous changes without human review, consumers may be directly exposed to incorrect content, eroding brand trust.

2. Only 9% of companies currently consider automated evaluation results reliable enough; brands should not fully delegate approval authority to AI in customer-interaction touchpoints.

Summary 2: Brands should prioritize human review and observability to protect quality control.

1. Over the next year, the reliability area with the highest investment growth is human review workflows, selected by 30% of respondents; production observability tools follow at 26%.

2. Among companies deploying autonomous AI agents, only 29% use real-time automated quality checks as their primary monitoring method; most rely on log tracking. Brands need to close this gap with real-time content validation capabilities.

Summary 3: Tool selection and consumer trust-building trends.

1. OpenAI's developer platform has reached 59% enterprise penetration, Confident AI 36%, Braintrust 23%, Anthropic 16%, and LangSmith 13%. Brands evaluating AI systems can reference these mainstream tools.

2. 62% of companies plan to add or replace evaluation platforms within 12 months, indicating that brands are strengthening AI review capabilities to align with a more cautious AI adoption trend.

Summary 1: The trend toward autonomous AI production changes is tightening, and cautious deployment is becoming mainstream.

1. The share of companies allowing or planning production changes without human review fell from 75% to 56%, and support among core decision-makers dropped from 88% to 61%, suggesting sellers should not be overly aggressive in introducing AI to product selection, operations, or similar processes.

2. The share of companies explicitly retaining human review doubled from 20% to 42%, underscoring that human involvement remains essential to quality assurance at this stage.

Summary 2: The automated evaluation tool market is growing rapidly, creating business opportunities.

1. Usage of OpenAI's native evaluation tools among enterprises rose from 31% in July to 59% in August; 62% of surveyed companies plan to add, replace, or supplement evaluation tools within 12 months, with 35% planning to do so within 3 months.

2. Confident AI, Braintrust, Anthropic, and LangSmith all have meaningful penetration. Sellers may find opportunities in services or agency partnerships related to these tools.

Summary 3: Risk warnings and countermeasures.

1. Among companies with evaluation experience, 61% have encountered AI that passed tests but failed after launch, and only 9% trust automated evaluation results. Sellers should run real-world scenario tests before launching AI features.

2. Monitoring has clear gaps: only 29% of companies perform real-time quality checks. Sellers should prioritize human review and observability to avoid customer-facing AI incidents.

Summary 1: Production environments can explore AI-agent autonomy in low-risk scenarios.

1. Currently, 27% of surveyed companies already allow AI agents to execute production changes without human review in low-risk scenarios; 20% plan to build this capability within 12 months, and 35% explicitly will not allow it. Factories can gauge their own risk tolerance against these proportions.

2. Excluding companies without autonomous AI agents deployed, 32% have enabled approval-free access in low-risk scenarios, 24% are building it, and 42% retain human review.

Summary 2: Production monitoring and data validation remain weak links, and represent an opportunity for digital transformation.

1. Only 29% of companies with autonomous AI deployments use real-time automated quality checks as their primary monitoring; 36% rely on transaction trace logs, and 16% use only API gateway infrastructure tracking, which cannot determine whether output content is correct.

2. Among companies that have already enabled approval-free access, only 26% perform real-time output quality checks. Well-formed but incorrect content will not trigger alerts, so manufacturers should prioritize content-correctness validation when introducing AI.

Summary 3: Demand for automated evaluation tools is increasing, creating opportunities to work with suppliers.

1. OpenAI evaluation tool usage rose from 31% to 59%, but trust in automated evaluation is only 9%; main concerns are scenario fit, explainability, and bias.

2. Over the next year, human review workflows are the fastest-growing reliability investment area (30%), followed by production observability at 26%. Factories can acquire or partner to build these capabilities.

Summary 1: Clients' core pain point is unreliable automated evaluation.

1. Only 9% of companies consider automated evaluation results reliable enough. Among concerns, 27% say results do not match real-world scenarios, 24% struggle with explainability, 16% encounter evaluation bias, 13% worry about data privacy, and 12% cite insufficient maturity.

2. Of clients who conducted pre-deployment evaluations, 61% encountered AI that passed tests but caused customer-facing failures after launch; 22% experienced this more than once.

Summary 2: Solution demand is concentrated on human review, observability, and real-time quality checks.

1. Over the next year, the fastest-growing reliability investment areas are human review workflows (30%), production observability tools (26%), and automated evaluation pipelines (21%).

2. The monitoring gap is clear: only 29% of clients use real-time automated quality checks, and 36% rely on log tracking. Service providers can offer real-time output-correctness validation and explainability enhancement solutions.

Summary 3: The evaluation platform market is shifting rapidly, leaving room for entry.

1. OpenAI's developer platform has 59% enterprise penetration (40% as primary platform), Confident AI 36% (10% primary), Braintrust 23% (12% primary), Anthropic 16%, and LangSmith 13%.

2. 62% of clients plan to add, replace, or supplement evaluation platforms within 12 months, with 35% planning changes within 3 months. Service providers can help clients build automated evaluation pipelines and design approval-free permission boundaries.

Summary 1: Current usage of enterprise evaluation platforms and willingness to adjust.

1. Among primary evaluation platforms, OpenAI accounts for 40%, Confident AI 10%, Braintrust 12%, Anthropic 10%, and LangSmith 2%. In terms of enterprise penetration, OpenAI reaches 59%, Confident AI 36%, Braintrust 23%, Anthropic 16%, and LangSmith 13%.

2. 62% of companies plan to add, replace, or supplement evaluation platforms within 12 months, with 35% planning to do so within 3 months. This creates growth and switching opportunities for platforms.

Summary 2: Enterprises' core demand is improving evaluation reliability and explainability.

1. Among concerns about automated evaluation, 27% cite failure to match real-world scenarios, 24% cite lack of explainability, 16% cite bias or inconsistency, and 13% cite data privacy. Platforms need to address these issues.

2. 61% of companies have experienced AI passing internal tests but failing after launch. Platforms can offer evaluation environments closer to production and real-time quality-check capabilities.

Summary 3: Platforms should help customers manage boundaries around approval-free authority.

1. Companies are increasingly distinguishing between using automated evaluation tools and using evaluation results as the sole approval basis for production changes. Platforms can introduce features such as tiered approvals and risk-scenario identification.

2. Only 29% of autonomous AI deployers currently use real-time quality checks as their primary monitoring method. Platforms can strengthen output-content validation to prevent well-formed but semantically erroneous outputs from passing through.

Summary 1: Industry shift: companies are moving from aggressively granting AI agents autonomy toward strengthening evaluation.

1. The share of companies allowing or planning production changes without human review fell from 75% in July to 56% in August; among core decision-makers, it fell from 88% to 61%, indicating a contraction in AI-agent autonomy.

2. Over the same period, usage of OpenAI's native evaluation tools jumped from 31% to 59%, showing that companies are redirecting investment and attention toward evaluation capability building.

Summary 2: Emerging problems: trust in automated evaluation is low, and production monitoring has structural gaps.

1. Only 9% of respondents believe automated evaluation results are reliable enough; main concerns are scenario fit (27%), explainability (24%), bias or inconsistency (16%), privacy (13%), and insufficient maturity (12%).

2. Among companies deploying autonomous AI agents, real-time automated quality checks account for only 29%, while 36% rely on transaction logs and 16% on API gateways, making it impossible to judge output-content correctness.

Summary 3: Governance implications: human oversight remains a necessary part of AI production changes.

1. 42% of companies with autonomous agents deployed plan to retain human review, and human review workflows rank first in investment growth (30%), showing human-in-the-loop remains the primary risk-control model.

2. The survey was completed before Anthropic's CEO publicly called for slowing frontier AI development; subsequent executive statements may reinforce the cautious trend. Regulators should pay attention to authorization boundaries and accountability mechanisms for approval-free production changes.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

VB Intelligence2026年8月发布的Agent可靠性与评估专项调查显示，允许AI代理不经人工审核自主执行生产变更，或计划在一年内落地该机制的受访企业占比，从7月调查的75%降至56%。

两次调查样本均为自主报名的行业从业者及小组成员，结果无法直接推导全企业市场的整体变化。8月调查共回收140份有效回复，其中53%的受访者为企业AI采购最终决策者，7月调查中该群体占比为44%。即便单独统计这一核心决策群体，支持无人工审核生产变更或计划搭建相关能力的占比也从7月的88%降至8月的61%，排除了样本结构差异导致数据波动的可能。

本次调查开展时间早于9月12日Anthropic首席执行官达里奥·阿莫迪公开发文呼吁放缓前沿AI开发节奏、强化安全监督的节点，也早于OpenAI首席执行官萨姆·奥尔特曼、Google DeepMind首席执行官杰米斯·哈萨比斯等行业高管公开支持更审慎发展路径的表态，相关态度转变与上述行业公开呼吁不存在直接关联，目前也无明确数据指向企业收紧AI自主权限的具体触发因素。

企业对自主部署的谨慎态度，和持续攀升的AI评估工具投入形成明显反差。数据显示，OpenAI原生评估工具的企业使用率从7月的31%大幅升至8月的59%。参与调查的企业中，61%已在部署前开展AI效果评估的受访者提到，过去12个月内曾遇到AI代理或大模型功能通过内部测试，上线后却引发面向客户故障的情况，其中22%的受访者遇到过不止一次同类故障，7月同口径统计的故障发生率为53%。

当前企业对自动化评估的信任仍处低位，仅9%的受访者认为现有自动化评估结果足够可靠可直接使用。针对自动化评估的核心顾虑中，27%的受访者面临的核心问题是评估结果无法匹配真实场景的实际表现，24%的受访者受困于评估过程缺乏可解释性，16%的受访者遇到过评估存在偏差或结果不一致的问题，13%的受访者顾虑数据泄露或隐私风险，12%的受访者认为相关工具成熟度不足。

经历过AI通过内部测试却引发线上故障的企业，推进无人工审核自主部署的意愿并未出现明显下滑。数据显示，经历过同类故障的企业中，59%已放开特定场景的无审核变更权限或正在搭建相关能力，未经历过故障的企业中该比例为55%，二者不存在统计意义上的明显差距。

从企业当前的AI部署规则看，27%的受访企业已经允许部分低风险场景下的AI代理不经人工验证执行生产变更，35%的企业明确在可预见未来不会放开无人工审核的变更权限，20%的企业计划在12个月内搭建支持自主变更的系统能力，16%的企业尚未部署任何自主AI代理，剩余2%的受访者未明确表态。剔除未部署自主AI代理的样本后，32%的企业已对特定低风险场景放开无审核变更权限，24%的企业正在搭建相关能力，42%的企业会在可预见未来保留生产变更的人工审核环节，这一占比较7月的20%实现翻倍增长。

投入优先级层面，30%的受访者将人工审核工作流列为未来一年可靠性相关投入增长最快的领域，26%的受访者选择生产可观测性工具，21%的受访者选择自动化评估流水线，11%的受访者选择安全与政策评估方向，剩余11%的受访者所在企业不会增加AI可靠性相关预算。将安全与政策评估列为投入增长最快领域的受访者占比从7月的16%降至11%，该数据仅反映增速优先级变化，不代表相关领域实际投入下滑。该投入结构和7月调查结果基本持平，并未出现资金从自动化评估领域向人工环节大规模转移的迹象。

上线后的运行监控环节仍存在明显缺口。在部署自主AI代理的企业中，仅29%将实时自动化质量检查作为主要生产监控手段，直接校验AI输出内容的正确性。36%的企业主要依赖交易追踪日志，记录基础设施运行状态、token消耗、输入输出内容供后续调试，不对输出正误做实时判断。16%的企业主要通过API网关做基础设施层面的运行追踪，11%的受访者不了解所在企业的监控方式，8%的企业采用临时人工审核的监控方式。现有监控数据仅统计企业采用的主要监控手段，无法证明未选择实时质量检查的企业完全缺失其他质量控制机制，部分企业可能部署了补充自动化校验或其他未被问卷覆盖的审核流程。

即便是已经放开无人工审核生产变更权限的企业，也仅26%将实时输出质量检查作为主要监控手段，多数企业的监控重点仍聚焦系统是否正常运行，而非输出内容是否准确，形式合规、返回正常状态码但实际内容有误的AI输出，大概率不会触发系统告警。

评估工具市场的格局仍在快速变动。OpenAI开发者平台当前的企业渗透率达59%，40%的受访企业将其作为主要评估平台。Confident AI渗透率达36%，10%的受访企业将其作为主要平台。Braintrust渗透率达23%，12%的受访企业将其作为主要平台。Anthropic控制台及工作台渗透率达16%，10%的受访企业将其作为主要平台。LangSmith渗透率达13%，2%的受访企业将其作为主要平台。62%的受访企业计划在12个月内新增、替换现有评估平台或补充相关工具能力，其中35%的企业计划在3个月内完成相关调整。

整体来看，受访企业正在明确区分两类决策，一类是使用自动化评估工具检测AI系统运行表现，另一类是直接以自动化评估结果作为生产变更的唯一审批依据。当前企业AI落地的核心矛盾，已经从是否要搭建自动化评估能力，转向在无人工介入的场景下，可以给自动化评估结果授予多大范围的操作权限。

本文首发于亿邦动力官方网站

文章来源：亿邦动力

FAQ回顾

企业为什么开始收紧AI代理的自主生产变更权限？

VB Intelligence 2026年8月调查显示，允许AI代理不经人工审核执行生产变更或计划一年内落地的企业占比从7月的75%降至56%，企业AI采购最终决策者中支持比例也从88%降至61%。调查开展早于Anthropic、OpenAI等高管公开呼吁放缓AI开发的节点，目前无明确数据指向具体触发因素。

AI代理通过内部测试但上线后引发故障的情况有多普遍？

VB Intelligence调查显示，在已开展部署前AI效果评估的企业中，61%在过去12个月内遇到过AI代理或大模型功能通过内部测试、上线后却引发面向客户故障的情况，其中22%遇到不止一次；7月同口径故障发生率为53%。这说明内部测试与真实场景表现之间存在明显落差。

企业使用自动化AI评估工具的主要顾虑是什么？

调查显示，仅9%的企业认为现有自动化评估结果足够可靠可直接使用。核心顾虑包括：27%认为评估结果无法匹配真实场景实际表现，24%认为评估过程缺乏可解释性，16%遇到过评估偏差或结果不一致，13%担心数据泄露或隐私风险，12%认为工具成熟度不足。

目前主流的AI评估工具有哪些？

VB Intelligence 2026年8月调查显示，OpenAI开发者平台企业渗透率59%，其中40%的企业将其作为主要评估平台；Confident AI渗透率36%，主要平台占比10%；Braintrust渗透率23%，主要平台占比12%；Anthropic控制台及工作台渗透率16%，主要平台占比10%；LangSmith渗透率13%，主要平台占比2%。62%的企业计划12个月内新增、替换或补充评估工具。

企业如何监控已部署的自主AI代理？

在部署自主AI代理的企业中，仅29%将实时自动化质量检查作为主要生产监控手段；36%主要依赖交易追踪日志，16%通过API网关做基础设施追踪，11%不了解企业监控方式，8%采用临时人工审核。即使已放开无人工审核生产变更的企业，也只有26%将实时输出质量检查作为主要监控手段。