AI in manufacturing: which three things to start with
The first step in adopting AI in manufacturing is not picking a platform; it is picking one scenario whose value can be counted. This is where things most often go wrong: most of the budget goes to a platform or hardware, the data is never connected, definitions are never unified, and nobody ends up using it. The sound starting order is three tasks — choose a scenario whose value can be counted, open its data path, and settle the definitions involved with named owners. Only then talk about expanding.
Why the order cannot be reversed
Reverse it and three familiar outcomes appear: hardware bought before a scenario is chosen, sitting for three months with no first use case; a large platform built before the data is ready, producing incomplete readings nobody trusts; dashboards built before definitions are settled, showing several numbers for the same fact, until managers stop looking after two days.
Policy points the same way. The eight-department Action Plan on "AI + Manufacturing" issued in December 2025 by the Ministry of Industry and Information Technology and seven other departments sets targets for 2027: 3–5 general-purpose large models deeply applied in manufacturing, 1,000 advanced industrial agents and 500 typical scenarios, alongside explicit measures to run maturity assessments on manufacturers and to draw up industry application maps and transformation roadmaps — see the current situation first, then the path.
Task one: choose a scenario whose value can be counted
One criterion: how much money does this scenario come to in a month? Those that can be counted go first, even if they look less advanced in a proposal.
| Candidate scenario | Countable value | Where the data sits |
|---|---|---|
| Unplanned downtime and repair | Yes: downtime hours, lost output, repair cost | Equipment register, inspection logs, repair orders |
| Overdue and slow-moving stock | Yes: tied-up cash, price erosion | ERP inventory, inbound and outbound records |
| Rework and customer complaints | Yes: material, labour, overhead, compensation | Quality records, rework orders, complaint log |
| Delivery delay | Partly: overtime, expedited freight, penalties | Orders, scheduling, supplier arrivals |
| Energy anomalies | Yes: electricity cost and peak structure | Meters, energy system, meter readings |
Task two: open the data path
At the start you do not need every system connected, only the two or three sources this scenario depends on. At least one of three routes must work, and it belongs in the plan:
| Route | Typical form | Main obstacle | Upkeep |
|---|---|---|---|
| Official system API | Open interfaces shipped with ERP, MES or finance systems | Vendor enablement, licensing and fees | Low |
| Read-only database account | An account that can query but not modify | Requires IT or vendor approval; must not affect the production database | Medium |
| Scheduled export directory | The system drops scheduled reports into a folder | Field names change; headers need checking | Medium; easiest to start |
Where none of the three works is simply out of scope for this round. Stating that access is read-only, writes nothing back and controls no production equipment is what lets IT and vendors approve it.
Task three: settle the definitions
"Inaccurate data" in manufacturing is rarely a broken system. It is usually one fact defined differently in different tables, or a manual entry step that gets skipped. At the start you only need a definition register for this one scenario, with six fields: metric, definition, source system, table and field, refresh time, owner. Definitions are settled by the company, not the vendor — the vendor registers and reconciles them. Leave this out of the contract and it will be argued about later.
How far peers have gone
Adoption is faster than most estimates suggest. IDC's 2025 survey of Chinese industrial enterprises found the share applying large models and agents rose from 9.6% in 2024 to 47.5% in 2025, and the share applying them simultaneously across R&D, manufacturing and supply chain jumped from 1.7% to 35%. The State Council Information Office reported in its first-half 2026 industrial briefing that AI adoption among manufacturers above a designated size has passed 30%.
Client-site numbers are appearing too. Midea's Meiyun Zhishu launched an agent-factory solution in January 2026 covering R&D, manufacturing, supply chain, marketing and management, with 158 core scenarios deployed. Midea states that equipment and mould agents for failure prediction and process optimisation lift overall equipment effectiveness by 30% and double inspection efficiency, while supply-chain agents shorten end-to-end delivery cycles by 39% and cut inventory turnover days by 30%.
The gap, then, is not whether to adopt but whether a pilot reaches daily operation: industry estimates suggest nearly 80% of enterprises have run an agent pilot but only about one in seven reached production, held back by data access, definitions and operational ownership.
Three pragmatic moves for smaller manufacturers
The agent matrices published by large manufacturers cannot be copied quickly. For smaller plants the practical route is simpler: pick one workshop or one line as a demonstration rather than the whole plant; build the numbers the finance function can read directly, so management sees value early; and hand operations to a party that will carry it long term instead of staffing a team to track model updates and interface changes. All three share the same traits: small in scale, verifiable, sustainable.
Four common mistakes
- Treating AI as one-off procurement. It is closer to a running routine: definitions, interfaces and business change, so someone owns it continuously.
- Demanding full coverage at the start. Full-process rollouts stall before acceptance; start with one countable scenario and expand.
- Acceptance criteria written vaguely. "Good results" in a contract means nothing. Specify test scope, definitions and how pass or fail is judged.
- Never asking who runs it. When an interface changes in year two, someone must own it — name that party in the plan.
On choosing a data route or a delivery model, see our enterprise AI agent deployment and custom delivery pages. If you are unsure which scenario should go first in your plant, send us your current systems and the one thing you most want visibility on (contact), and we will run a free on-site assessment.
In one line: choose one scenario whose value can be counted, get its data accurate and its definitions fixed, then talk about step two. Success in manufacturing AI is decided by these three concrete tasks, not by the model.
FAQ
- Where should a manufacturer start with AI?
- With one scenario whose value can be counted — downtime, slow-moving stock, rework or delivery delay. Only those give a pass-or-fail measure, and only those let you show savings afterwards. Building a platform or buying hardware first usually ends with nobody using it.
- Should we build a data platform first?
- Not as a first step. A platform is the product of large-scale reuse, not a precondition. To start, open one data path for one scenario — an official API, a read-only database account or a scheduled export directory — and get those numbers right. Revisit platform work once the second or third scenario runs steadily.
- Can a plant without an IT team adopt AI?
- Yes, if three things are written into the plan: who opens the data path, who settles the definitions, and who runs it afterwards. None of these requires an in-house development team, but each requires a named counterpart. In public schedules, the real delays come from data-access approval and inconsistent definitions, not from the model.
- How soon do results show?
- In public schedules the first scenario runs in two phases: site survey and data access (about 2–3 weeks), then definition registration, metrics and alerts, and the move to daily operation (about 4–6 weeks). What becomes visible first is steadier numbers and earlier detection of anomalies; countable savings usually appear one to two months later. Proposals promising instant results tend to skip data access and definitions.