Direct answer: For a 2027 AI project, do not start with “how much subsidy can we get”; start with the business problem, baseline and acceptable results. Policy support is a partial offset after the project stands, not the reason to start. Prepare five evidence sets: decision, investment, data/security, POC/ROI, and acceptance/exit.
This report is for owners, CIO/IT, finance, procurement and project leads at manufacturers/B2B companies, and discusses how to set up, document and accept a policy-supported AI project without “building a project for the subsidy”. Information is current to 2026-10-30. It is not a filing agent, subsidy amount, eligibility or legal opinion; national policy gives direction, while specific eligibility, materials, amounts and timing follow the local authority's notice for the year.
Policy targets are direction, not orders or subsidy promises. A project without a baseline and owner should not be created for an application. Acceptance evidence must be reproducible—not demos, screenshots or “expected benefits”.
What are policy targets—and what are they not?
National policy sets 2027 adoption and scenario targets, but does not mean an individual company will get a subsidy, pass review or earn a return; local eligibility, amounts and timing follow the current notice.
The State Council's “AI+” action opinion sets a target of deep integration with six key areas by 2027 and more than 70% penetration for new-generation intelligent terminals and agents, rising above 90% by 2030. The MIIT and seven other departments' “AI + Manufacturing” plan calls for 1,000 high-level industrial agents, 100 high-quality datasets, 500 typical scenarios and 1,000 benchmark enterprises by 2027. These are industrial direction and targets, not an order, subsidy amount or acceptance conclusion for any single company.
| Policy signal | What it says | What it does not say |
|---|---|---|
| >70% penetration by 2027 | Intelligent terminals/agents are a direction | Your project automatically qualifies |
| 1,000 industrial agents/500 scenarios | Manufacturing scenarios are a priority | Your scenario will be selected or funded at a fixed amount |
| 1,000 benchmark enterprises | Benchmarks are a policy tool | Applying guarantees benchmark status |
| Agent governance/security | Permissions, audit and responsibility become gates | Buying a platform alone meets compliance |
Targets are quoted from public documents; eligibility, materials, amounts, timing and review criteria follow the local authority's current notice. This article does not determine eligibility.
Which three gates should be passed before applying?
Pass the business/evidence gate, investment/compliance gate and ROI/stop-loss gate; if any lacks a baseline, owner or acceptable result, do not rush a project for an application.
A policy-supported project usually has to answer three questions: what business problem it solves, where the money goes, and how effectiveness is proven. Write these three gates into an internal business case before deciding whether to apply; do not start from an application template and reverse-engineer a project.
| Gate | What to write | If missing |
|---|---|---|
| Business/evidence | Problem, baseline (time/cost/defect rate), owner, process, acceptance metrics | Do not start; “policy encourages it” is not a requirement |
| Investment/compliance | Budget lines, contracts/invoices, data authorisation, permissions/security, supplier duties | Fix contracts/authorisation; hardware stacking is not an AI project |
| ROI/stop-loss | Cost per task, payback, stage gates, exit/decommission plan | Narrow or pause; do not submit |
The three gates are an editorial framework, not an authority's review criteria; confirm against the local notice and project facts.
How should investment evidence be kept?
Contracts, invoices, quotes/BOQ, timesheets, cloud/API bills, data/integration, training and operations must map to the project; a single total invoice cannot explain where the money went.
Policy support/audit usually asks whether the money was actually spent, whether it was used for the project, and whether it matches the application. Break investment into reconcilable lines and keep contracts, invoices, acceptance records, timesheets/allocation, cloud/API bills, data/integration service records and training records. Do not merge hardware, software, services, cloud resources and internal labour into one number; purchased and self-developed items have different evidence requirements.
| Line | What to keep | Common problem |
|---|---|---|
| Software/platform | Contract, licence/subscription, invoice, account/usage | Counting multi-year subscription at once |
| Hardware/compute | Purchase contract, invoice, asset ID, deployment/usage record | Buying GPUs that sit idle |
| Data/integration | Interface/data service contract, effort, acceptance record | Only “system integration fee” |
| Cloud/API | Bill, usage, project/environment tags | Personal account/no project attribution |
| Internal labour | Timesheet, roles, tasks, allocation rule | No record, filled in later |
| Training/operations | Training sign-in/materials, operations contract, on-call record | Booking training as software |
Evidence requirements follow the authority/audit rules; this article gives no tax, accounting or subsidy-amount opinion.
How should technical/data/security evidence be kept?
Architecture, data source/authorisation, permissions, logs, evaluation set, supplier contracts and change records must reconcile; agent projects also need decision authority, human takeover and decommission/recall plans.
An AI project does not end with “buying a model”. Application/acceptance materials should explain the system boundary, where data comes from, who may use it, how it is evaluated and who is responsible for errors. For agents/automated decisions, policy documents explicitly mention decision authority, behaviour control, supply-chain security and tiered governance; procurement and acceptance should turn these into documents and tests, not just “security and compliance”.
| Evidence | What to write | Acceptance test |
|---|---|---|
| Architecture | Model/tool/data/interface/deployment boundary | Architecture diagram, environment list |
| Data authorisation | Source, consent/authorisation, purpose, retention, cross-border | Authorisation file, data list |
| Permissions/logs | Roles, least privilege, audit, alerts | Permission matrix, log sample |
| Evaluation set | Independent holdout, defect/question types, metrics | Evaluation report, recalculation script |
| Human takeover | Who takes over, when, escalation path | Drill record, ticket |
| Supplier/exit | Data/model/prompt ownership, stop/export/delete | Contract, exit test |
Agent governance requirements can reference the CAC and two other departments' “Implementation Opinions on Regulated Application and Innovative Development of AI Agents”; applicability depends on the scenario, local requirements and contract.
How should acceptance/audit be self-checked?
Compare actual results with the same baseline; verify investment evidence, technical evidence, POC/ROI, asset list and exit plan; do not accept with demos, screenshots or “expected benefits”.
Before acceptance, run a “mock audit”: put the business case, contracts/invoices, technical documents, evaluation report, go-live record, training/operations records and stop-loss lines together and check whether they support each other. If actual results are below baseline or gate, write down the reason, remediation or exit—do not use “policy requirements” as a reason to keep spending.
| Self-check | Pass standard | Evidence |
|---|---|---|
| Baseline/results | Same definition, differences explainable | Baseline sheet, acceptance report |
| Investment | Contracts/invoices/bills/timesheets map to project | Evidence list, asset table |
| Technology | Architecture/data/permissions/logs/evaluation reproducible | Documents, test records |
| POC/ROI | Stage gates, cost per task, payback | POC report, financial model |
| Assets/exit | Accounts/data/model/prompts can be handed over/stopped | Asset list, exit drill |
| Remediation | Missed items have owner/deadline/retest | Remediation sheet, retest record |
The self-check is an editorial suggestion, not a substitute for authority review, audit or legal advice; prepare materials according to the current notice.
Which projects are not suitable for an application?
Projects without a business owner, baseline, data/permissions, acceptance/exit plan, or those created mainly for a subsidy, should not be submitted; fix the basics or narrow the scope first.
Policy support is not “free budget”. If the project does not stand on its own, more application materials only increase acceptance/audit risk. The situations below should be paused or reworked rather than packaged as an “AI project”.
| Situation | Risk | Do first |
|---|---|---|
| No business owner | No use/maintenance after launch | Assign owner and business metrics |
| No baseline | Cannot prove impact | Add baseline/historical data |
| Hardware/platform only | Cannot explain AI use/output | Add scenario/integration/acceptance |
| Data/permissions unclear | Compliance/security risk | Add authorisation/permissions/logs |
| No acceptance/exit | Audit/decommission risk | Write gates/exit plan |
| Mainly for subsidy | Project does not stand; hard to explain later | Return to the business problem; do not apply if needed |
The “not suitable” list is an editorial suggestion, not an authority conclusion; confirm against the local notice and company facts.
Who should read this, who should not, and what decision does it support?
It fits manufacturers preparing 2027 AI/smart-manufacturing applications, acceptance or audits; it does not fit teams without a business owner, seeking only a subsidy, or expecting a “guaranteed approval/fixed amount”.
Who should read: owners, CIO/IT, finance, procurement and project leads at manufacturers/B2B companies; companies preparing AI/smart-manufacturing/digital applications or explaining project investment and results to customers/auditors; teams doing agent/automation projects that need permissions/acceptance evidence. Who should not: teams without a clear business problem, data/process owner, or only buying hardware to access a subsidy; teams expecting a fixed subsidy amount, guaranteed approval or a replacement for the authority's notice; projects needing formal audit/legal advice. Decisions supported: whether to apply, what evidence to prepare, how to split investment, what technical/data/security materials to keep, how to self-check acceptance/audit, and when not to apply.
FAQ
Does the 70% adoption target mean we must adopt AI in 2027?
No. The 70% is a directional target for key areas, not a requirement for every company. Whether and what to adopt depends on the business problem, baseline, data/process and ROI; without these, fix the basics first.
Is “expected benefit” the most important application material?
No. Expected benefit can be included, but it needs assumptions, a baseline and a calculation method; acceptance/audit looks more at actual results, investment evidence, technical evidence and reproducible data. “Expected benefit” alone rarely supports acceptance.
Does buying hardware/compute count as an AI project?
Hardware/compute is an investment line, not an AI application. The application/acceptance should explain which scenario it supports, how the model/system is used and how output is measured; hardware without application and acceptance is easy to challenge.
What extra evidence do agent projects need?
At minimum decision authority, human takeover, behaviour logs, permissions/audit, supply-chain security and a decommission/recall plan. Policy documents set agent-governance requirements; applicability depends on the scenario, local requirements and contract.
What if the project misses the application target?
Handle it through acceptance/stop-loss: document the gap, cause, remediation or narrow/pause/exit; do not use “policy requirements” or “a little more investment” to hide underperformance. Whether it affects the subsidy/acceptance depends on the authority's rules.
How do national policy and local notices relate?
National policy gives direction and priorities; local notices give specific eligibility, materials, amounts, timing and review criteria. Follow the local notice for the year; this article does not replace it or guarantee eligibility/amount.
Bottom line
For a 2027 AI project application/acceptance, start with a business problem, baseline and owner, then discuss policy support. Keep decision, investment, data/security, POC/ROI and acceptance/exit evidence complete; narrow or exit when gates are missed—do not package a project for a subsidy.
Method and limits
This article is public policy, statistical data and editorial method, not eligibility, amount, audit or legal advice for any company, project, region or application. It uses no unpublished application materials, review opinions or subsidy data; the three gates, evidence tables, self-check and “not suitable” list are editorial, and specific conditions, materials, amounts, timing and review criteria follow the local authority's current notice. Policy/statistics change; the latest official text governs.
Sources
Information is current to 2026-10-30; each official/primary page governs its own scope and version.
- State Council, “Opinions on Deepening the Implementation of the ‘AI+’ Action” (Guofa [2025] No. 11, 26 August 2025): gov.cn. 2027/2030 adoption targets, key areas and “AI+” direction.
- MIIT and seven other departments, “Implementation Opinions on the ‘AI + Manufacturing’ Special Action” (MIIT Lianke [2025] No. 279, 25 December 2025): nda.gov.cn. 2027 targets for industrial agents, datasets, typical scenarios and benchmark enterprises.
- CAC and two other departments, “Implementation Opinions on Regulated Application and Innovative Development of AI Agents” (8 May 2026): cac.gov.cn. Agent definition, decision authority, behaviour control, supply-chain security and tiered governance.
- National Bureau of Statistics, “Industrial profits of industrial enterprises above designated size rose 15.5% in January–September 2026” (27 October 2026): stats.gov.cn. Industrial/manufacturing profit and sector divergence, as background on 2027 investment capacity.
- National Bureau of Statistics, “National industrial capacity utilisation was 73.1% in Q3 2026” (19 October 2026): stats.gov.cn. Manufacturing 73.3%, automotive 66.2%, etc.