The main finding: policy signals for manufacturing AI in 2026 are strong, but companies are still at the “many starts, few at scale” stage. The State Council has set a target of more than 70% penetration for new-generation intelligent terminals and agents by 2027. The Ministry of Industry and Information Technology (MIIT) manufacturing action plan calls for 1,000 industrial agents and 500 typical scenarios. Enterprise surveys, however, show that 80% of organizations have started deploying agentic AI, only 17% use it enterprise-wide, and 69% see a significant gap between AI spending and financial results.
National Bureau of Statistics (NBS) data released on 19 October 2026 showed manufacturing investment down 2.1% year on year in the first three quarters, while high-tech manufacturing investment rose 4.0%, information services investment rose 20.2%, and industrial robot output rose 29.1%. That points to 2027 AI budgets concentrating on high-value scenarios rather than expanding across the board. This article updates the 2027 budget picture, scenario priorities, and the Q4 project-approval decision on that basis.
This is an industry analysis and public-source study for manufacturing owners, CIOs and IT leaders, and procurement managers. Research cut-off: 25 October 2026. The scope is mainly China's manufacturing sector and enterprise AI. It cites policy targets, institutional research, and public market data. It does not present forecasts as facts that have already happened, and it does not treat cross-industry data as if it were manufacturing-specific.
A policy target is not an order. The number of agents is not business value. A 2027 budget should be built around “scenario closure + integration capability,” not around “how many agents we bought.” Prove one or two scenarios first, then talk about platform and scale.
What actually changed in manufacturing AI in 2026?
The shift is from point tools to agents and closed process loops. Policy has set quantified 2027 targets, while companies are still expanding from pilots; the share truly running AI enterprise-wide remains low.
| Date | Policy / report | Key points relevant to this cycle | What it means for manufacturing |
|---|---|---|---|
| 26 August 2025 | State Council, “Opinions on Deepening the Implementation of the ‘AI+’ Action” (Guofa [2025] No. 11) | By 2027, penetration of new-generation intelligent terminals and agents exceeds 70%; by 2030, exceeds 90% | Policy makes agents the direct vehicle for application; procurement windows will move earlier |
| 25 December 2025 | MIIT and seven other departments, “Implementation Opinions on the ‘AI + Manufacturing’ Special Action” | By 2027, launch 1,000 high-level industrial agents, 100 high-quality datasets, and 500 typical application scenarios; select 1,000 benchmark enterprises | Manufacturing is a priority sector; scenarios, datasets, and benchmark enterprises are the keywords |
| 8 May 2026 | Cyberspace Administration of China and two other departments, “Implementation Opinions on the Standardized Application and Innovative Development of Agents” | Defines agents, product guidelines, and security and governance requirements; names intelligent manufacturing and other scenarios | Beyond “it runs,” permissions, audit, and responsibility boundaries will become procurement thresholds |
| 25 June 2026 | China Academy of Information and Communications Technology (CAICT), “Enterprise Agent Technology and Application Research Report (2026)” | Enterprise agents need model adaptation, memory, tool calling, planning, multi-agent collaboration, and private deployment; bottlenecks include data silos, hallucination, unclear human-machine responsibility, and insufficient audit | Scaling is engineering and governance, not “connecting a few more models” |
Policy targets come from public documents. The CAICT report is referenced through its public summary and table of contents; its original charts are not reproduced here. Policy targets set direction; they are not an order or a return promise for any single company.
The timeline makes one thing clear: the 2026 question for manufacturing AI is not “whether to do it,” but “where to start, how to reserve budget, and how to accept the result.” Policy has pushed agents into the foreground, but companies still lack scenarios, data, integration, and operations.
What do the first three quarters of 2026 signal for 2027 budgets?
Manufacturing investment fell 2.1% year on year, but high-tech manufacturing investment rose 4.0% and information services investment rose 20.2%. The 2027 budget will be more concentrated, not a broad expansion.
| Observation | Latest 2026 data | Year-on-year / basis / source | What it means for the 2027 budget |
|---|---|---|---|
| Investment structure | Manufacturing investment down; high-tech manufacturing and information services investment up | Manufacturing −2.1%; high-tech manufacturing +4.0%; information services +20.2%; NBS, 19 October 2026 | Total contraction, structural divergence; budgets concentrate on quantifiable scenarios |
| Production | Value added grew in manufacturing, high-tech manufacturing, and equipment manufacturing | Manufacturing +5.7%; high-tech manufacturing +14.3%; equipment manufacturing +10.2%; NBS, 19 October 2026 | Production is still growing; AI must tie to quality, efficiency, and delivery |
| Equipment and automation | Output of industrial robots, 3D printing equipment, and lithium batteries grew | Industrial robots +29.1%; 3D printing +62.5%; lithium batteries +39.5%; NBS, 19 October 2026 | Automation and equipment investment still have momentum; projects bundled with line upgrades are easier to approve |
| Smart factory supply | Advanced-level and excellence-level smart factories and AI scenarios have reached scale | 13,000+ advanced-level and 500+ excellence-level; 20 industries and 1,000+ scenarios; MIIT press conference / Xinhua, 21 October 2026 | Many benchmarks, hard to replicate; 2027 will favor replicable scenarios |
| Enterprise adoption | Large-model and agent adoption is rising; the industrial AI market is viewed positively | 9.6% in 2024 → 47.5% in 2025; 2026 addressable market RMB 116.3 billion; 36Kr Research / IDC, 16 October 2026 (research estimate) | Demand is spreading, but definitions differ; not every company needs a platform |
NBS and MIIT figures are official actuals. The 47.5% from 36Kr Research / IDC is the share of industrial enterprises applying large models and agents in 2025, and RMB 116.3 billion is a research estimate, not official statistics. It is not the same survey basis as IDC's “17% enterprise-wide use” cited above, and the two cannot be added together or substituted for each other.
The most important word in this dataset is “divergence,” not “AI is hot.” Overall fixed-asset investment fell 7.2% year on year in the first three quarters, and manufacturing investment fell 2.1%, but high-tech manufacturing investment rose 4.0% and information services investment rose 20.2%. Capital has not disappeared; it is concentrating in high-tech manufacturing and digital services. The 2027 AI budget will follow the same logic: projects that can clearly account for quality, efficiency, delivery, or cost benefits are more likely to get funded.
At its 21 October 2026 press conference, MIIT also reported that China has built more than 13,000 advanced-level and more than 500 excellence-level smart factories, forming more than 1,000 application scenarios across 20 industries including steel, automobiles, and pharmaceuticals, and that it is drafting a new-generation smart manufacturing policy. Stronger supply does not make procurement easier. Integration, data, permissions, and operations remain bottlenecks on the company side.
For the 2027 budget, the direct implication is three shifts: from buying a platform to buying a closed scenario; from one-off hardware and software spending to integration, data, and operations; and from “how many agents went live” to “task closure rate and cost per task.” Q4 2026 is the project-approval window for the 2027 budget, but “spend this year's budget” is not a project rationale. Build a baseline and a POC on one or two scenarios first, then decide whether to expand next year.
Where do policy targets and company reality diverge?
Policy targets describe national supply and scenario counts, not any single company's ROI. IDC research shows 80% have started deployment but only 17% use AI enterprise-wide; Gartner predicts more than 40% of projects will be canceled.
| Dimension | Public data / judgement | Basis and timing | What it means for procurement |
|---|---|---|---|
| Policy targets | By 2027, penetration of intelligent terminals and agents exceeds 70%; 1,000 industrial agents and 500 scenarios | State Council, August 2025; MIIT, December 2025; national scope | Budgets will be more influenced by policy and benchmark cases, but not every company needs a large platform |
| Enterprise adoption | 80% of organizations have started deploying agentic AI; only 17% use it enterprise-wide | IDC 2026 survey, cross-industry | Most companies are still piloting; scenario closure is more realistic than enterprise-wide rollout |
| Return on investment | 69% see a significant gap between AI spending and measurable financial results; 73% see excessive AI cost as a major risk | IDC 2026 survey, cross-industry | Define the financial basis at project approval; do not report only “how many agents went live” |
| Project survival | By the end of 2027, more than 40% of agentic AI projects will be canceled because of cost, unclear value, and insufficient risk control | Gartner forecast, June 2025, cross-industry | This is a forecast, not a statistic, but it is enough to justify starting small |
| Buyer dissatisfaction | 34% are dissatisfied with limited integration, opaque models and data, or vendor lock-in; 31% are dissatisfied with overpromising or unclear pricing | IDC 2026 C-suite survey, cross-industry | Selection should shift from “many features” to “can integrate, can accept, can exit” |
IDC and Gartner data come from cross-industry surveys or forecasts and cannot be equated directly with manufacturing as a whole. Policy targets describe the national manufacturing direction. Keep the original basis when citing them, and avoid stitching data with different definitions into a single “manufacturing AI market size.”
Changzhou has a more specific set of local policy instruments: a city-level three-year action plan with 2028 targets, and a separate RMB 20 million-a-year “four vouchers” program in the Changzhou Economic Development Zone. Their eligible recipients, subsidy definitions, and application conditions differ. See the Insights index for our guide to Changzhou's manufacturing AI subsidies.
Where is the money going?
First to software and agents that plug directly into business processes, then to integration, data, and operations. A platform subscription is only part of the first-year investment.
IDC's “China AI-Enhanced Enterprise ERP Market Share, 2025,” published in May 2026, put China's AI-enabled ERP market at USD 315.7 million in 2025, up 96.1% year on year. The figure counts only new revenue generated by AI products, AI modules, and AI capabilities in ERP scenarios. It excludes traditional ERP revenue, and it is not the total manufacturing AI market.
| Budget line | Typical share | What it covers | Advice for the 2027 budget |
|---|---|---|---|
| Platform / software | 10–30% | Agent platform, ERP/MES AI modules, subscriptions or licenses | POC before buying; do not buy all modules at once |
| Model / compute | 10–30% | API, private inference, GPU or dedicated resources | Estimate from real call volume; do not buy cards for demo peaks |
| Integration / data | 30–50% | ERP/MES/OA interfaces, master data, permissions, evaluation sets | Budget separately; no project approval without an interface inventory |
| Operations | 15–30% per year | Model upgrades, monitoring, security, on-call support | Plan over three years; at least one or two administrators or managed support |
Shares are editorial estimates for manufacturing AI projects, not official statistics or a quote; they vary with the number of systems, data quality, and compliance requirements. The high growth of AI-enabled ERP shows that AI investment in that segment is being released; it does not mean every manufacturer should privatize AI or buy a platform immediately.
The lesson for manufacturing is not “buy ERP AI modules now.” It is that AI is moving from peripheral tools into core business systems. Once it enters ERP/MES, integration, permissions, master data, and audit become the main body of the budget. A project that buys a platform but not integration is unlikely to close the loop.
Which manufacturing scenarios should go first, and which should wait?
Quality traceability and inspection, scheduling and reporting assistance, equipment maintenance, and process knowledge Q&A are ready first. Cross-plant autonomy, core transaction write-back, and full-process transformation without a data foundation should wait.
| Scenario | Data foundation | Value | Risk | Recommendation |
|---|---|---|---|---|
| Quality traceability / inspection | Inspection records, batches, equipment parameters | Fewer missed inspections; faster traceability | Misjudgement; responsibility boundary | Start with assisted decisions and keep human review |
| Scheduling / reporting assistance | ERP work orders, MES reporting, BOM and process data | Less duplicate entry; faster planning response | Write-back errors affect production | Read-only or advisory first, then small-scope write-back |
| Equipment maintenance | Inspection, repair, and sensor data | Less downtime; longer equipment life | Missing data; false alarms | Start with knowledge Q&A and work-order assistance, then prediction |
| Process knowledge Q&A | Process documents, SOPs, historical issues | Faster onboarding; fewer repeat questions | Version confusion; permissions | Start with a 200–500-item evaluation set and permissions |
| Cross-plant autonomy / core transaction write-back | Multi-plant master data, permissions, audit | High theoretical value | Organization, compliance, and incident costs are high | Wait until single-plant closure is stable |
Scenario ranking is an editorial judgement, not an industry consensus. The same scenario varies widely by industry and data foundation. Validate with real tasks and an evaluation set before expanding.
For manufacturing execution, start with the ERP/MES/OA integration capabilities of the Lingyi AI Manufacturing Platform (LingMOE). For office collaboration, approvals, and documents, see the LingWork Office Agent. For a private execution engine with model independence and audit, see the LingHarness AI Foundation. Products are the vehicle; confirm the scenario and data boundary first.
How should companies of different sizes budget for 2027?
Under 50 people, buy SaaS and APIs first. From 50 to 500, build a hybrid by scenario. Above 500, consider private deployment and multiple plants. Work backwards from scenario benefits; do not buy a platform first.
| Company size | 2027 focus | First-year budget (editorial estimate) | Deployment advice | What not to do |
|---|---|---|---|---|
| Under 50 people | 1–2 high-frequency scenarios; start with a knowledge base or production-reporting assistance | RMB 50,000–200,000 | SaaS/API, pay as you go | Build a GPU cluster; buy a full platform |
| 50–500 people | 3–5 scenarios; read-only ERP/MES integration; establish a data owner | RMB 300,000–1,000,000 | Hybrid: sensitive data local, non-sensitive tasks in the cloud | Replace ERP/MES in one go; skip the pilot and roll out |
| 500+ people / multiple plants | Platform, multi-plant replication, permissions and audit, model governance | RMB 1–3 million+ | Private or hybrid, phased by plant | Measure by “number of agents”; roll out without an operations team |
Budgets are editorial estimates, excluding buildings, equipment replacement, and a full ERP rebuild; they are not quotes. For a more detailed private deployment TCO, see “What private AI deployment really costs in 2026.” For ERP/MES/OA integration budgets, see “What ERP, MES, and OA integration really costs in 2026.” For implementation and acceptance, see the Insights index.
How has vendor selection changed in 2026?
Buyers are moving from comparing models and features to comparing integration, data boundaries, acceptance, and exit. The top three buyer complaints in IDC's research are limited integration, opaque models and data, and vendor lock-in.
| Evaluation area | What to ask | Red flag |
|---|---|---|
| Scenario closure | Which real tasks, how many evaluation cases, and what counts as a pass | Only demonstrates carefully prepared questions and will not share the evaluation method |
| System integration | How ERP/MES/OA connect, whether write-back is possible, and what happens with legacy systems that have no API | Avoids the interface rework effort, or pushes all integration to the customer |
| Data boundary | Whether offline or intranet deployment is possible, where logs are stored, and whether external APIs are called | “Private” deployment, but key components still depend on the public internet |
| Acceptance and SLA | Task success rate, P95 latency, fault response, and support hours | Only promises “good results,” with no measurable metrics |
| Exit and migration | Whether data, models, prompts, tools, and the knowledge base can be exported; whether source code is delivered | Only renewal is possible, not migration; data cannot be taken out |
| Billing transparency | How platform, model, integration, changes, and renewal are priced | Low-price lead, with integration and changes used to claw back the difference |
This table is a procurement question list, not an assessment of any vendor. Specific terms should be reviewed by legal and business teams together.
Which three metrics matter most in 2027?
Task closure rate, cost per task, and time from project approval to production. Not the number of agents, token consumption, or demo quality; the last two only help explain the first three.
| Metric | Definition | Why it matters | Data source |
|---|---|---|---|
| Task closure rate | Tasks completed without human intervention ÷ total tasks | Separates “sounds right” from “actually done” | Work-order or process system + agent trace |
| Cost per task | (Model + tools + labor + operations) ÷ successful tasks | Prevents “smarter but losing more money” | Bills, work hours, operations allocation |
| Time from approval to production | Time from project approval to stable launch | Reflects the real friction in integration, data, and organization | Project plan, POC and pilot records |
| Adoption rate | Weekly active target users ÷ total target users | If nobody uses it, technical metrics are meaningless | Product analytics, logins, work orders |
| Human takeover rate | Tasks taken over by humans ÷ total tasks | Degree of automation and the cost of risk fallback | Agent event stream + human work orders |
The first three are the core 2027 metrics recommended in this article; the last two help explain them. Metric definitions and acceptance thresholds should be adjusted by scenario; they are not an industry standard.
What should you do now, and what should you not do?
Run a POC on one or two high-value scenarios, and reserve budget for integration, data, and operations. Do not buy a full platform for a subsidy, a benchmark slot, or a demo, and do not start by spreading agents everywhere.
Do: assign a business owner and a data owner; choose one or two real tasks from work orders, production reporting, quality inspection, or equipment maintenance; prepare 200–500 evaluation cases; integrate read-only first and write back on a small scale later; put integration, data, and operations into a three-year budget; review quarterly using task closure rate, cost per task, and time to launch.
Do not: buy a platform without a scenario; build a knowledge base without a data owner; accept a project without an evaluation set; use “number of agents” as a KPI; rush a project for a policy subsidy; treat a demo as production; or treat cross-industry survey data as your own ROI.
If the budget is limited, the priority order is: scenario and evaluation set → data and permissions → read-only integration → pilot → write-back and production → scale. Choose the platform and model after the second step, not before.
FAQ
What does the 70% penetration target for intelligent terminals and agents by 2027 mean?
It is a stage target in the State Council's “Opinions on Deepening the Implementation of the ‘AI+’ Action” for the penetration of new-generation intelligent terminals and agents. It does not mean manufacturing AI procurement or any single company's orders will grow by the same amount. For companies, it means procurement timing and budgets will be more influenced by policy and benchmark cases.
Which AI scenarios are most mature in manufacturing now?
Quality traceability and inspection, scheduling and reporting assistance, equipment maintenance, and process knowledge Q&A are relatively mature. Cross-plant autonomy, core transaction write-back, and full-process transformation without a data foundation should wait.
Should small and mid-sized companies adopt AI in 2027?
Yes, if they have a clear scenario, a data owner, and one or two administrators. Start with SaaS or APIs on one or two scenarios. Do not build a GPU cluster or buy a full platform at the start.
What is a typical first-year budget for manufacturing AI?
Depending on size and whether deployment is private, first-year budgets commonly run RMB 50,000–200,000, RMB 300,000–1,000,000, or RMB 1–3 million or more. These are editorial estimates, excluding buildings, equipment replacement, and a full ERP rebuild; the actual figure depends on the scenario list and the POC.
Which metrics should we watch most in 2027?
Task closure rate, cost per task, time from project approval to production, adoption rate, and human takeover rate. Do not substitute the number of agents, token consumption, or demo quality for business results.
The keyword for manufacturing AI in 2026 is not “number of agents” but “scenario closure.” Policy has set clear 2027 targets; companies need to put budget into one or two real tasks, integration, data, and acceptance. Prove it first, then scale.
Sources and caveats
This article is public-source research, not first-hand interviews, surveys, or testing. Policy targets, institutional research, and market sizes come from different bases. State Council and MIIT figures are national policy targets; NBS figures are actual macro data for the first three quarters of 2026; IDC and Gartner figures are cross-industry surveys or forecasts; the IDC AI-enabled ERP figure is a segment market statistic; the CAICT report covers enterprise agents generally, not manufacturing specifically; the 47.5% and RMB 116.3 billion figures from 36Kr Research / IDC are research estimates, not official statistics. This article does not combine them into a “total manufacturing AI market size,” and it does not present forecasts as facts. The 47.5% and the earlier IDC 17% enterprise-wide use figure are not the same survey basis and cannot substitute for each other. Budget ranges, scenario ranking, and metrics are editorial judgements and should be adjusted to the company.
- State Council, “关于深入实施‘人工智能+’行动的意见” (Guofa [2025] No. 11, 26 August 2025). gov.cn
- MIIT and seven other departments, “‘人工智能+制造’专项行动实施意见” (MIIT Lianke [2025] No. 279, 25 December 2025). nda.gov.cn
- Cyberspace Administration of China and two other departments, “智能体规范应用与创新发展实施意见” (8 May 2026). cac.gov.cn
- IDC, “ERP 的下一站，是‘会执行的 AI’” (8 June 2026, citing 2025 AI-enabled ERP market data). idc.com
- IDC, “The Enterprise AI Deal Has Changed. Here’s What Vendors Need to Prove.” (20 October 2026). idc.com
- Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027” (25 June 2025). gartner.com
- CAICT, “企业级智能体技术与应用研究报告（2026年）” (25 June 2026). caict.ac.cn
- National Bureau of Statistics, “前三季度经济运行总体平稳 高质量发展取得新成效” (19 October 2026). stats.gov.cn. First-three-quarter data on manufacturing investment, high-tech manufacturing investment, information services investment, and industrial robot output.
- Xinhua, “前三季度工业和信息化发展释放稳中提质信号” (21 October 2026, MIIT press conference). news.cn. 13,000+ advanced-level and 500+ excellence-level smart factories; 20 industries and 1,000+ scenarios; information on the drafting of a new-generation smart manufacturing policy.
- 36Kr Research Institute, “2026年中国工业AI行业发展研究报告” (16 October 2026). 36kr.com. Cites IDC survey data showing the share of industrial enterprises applying large models and agents rising from 9.6% in 2024 to 47.5% in 2025; estimates a 2026 industrial AI addressable market of RMB 116.3 billion. Research estimate, not official statistics.