Finding a company that builds factory AI agents: how to screen them
Five screens decide which company can actually build a factory AI agent: a referenceable similar site, a clear data-access plan, who owns definitions, who operates it afterwards, and how the engagement ends. Get through them and the list shortens by itself — only then is price worth comparing. Supply has expanded fast, with wide quality variation: many vendors stop at wrapping a model API and showing a demo, and lack system integration and long-term operations.
Who builds factory agents today
| Type | Visible strengths | What to watch | Usual role |
|---|---|---|---|
| Platform vendors | Models, compute, full product lines, flagship cases | Platform and programme-level delivery; may not take a single-plant scenario | Compute, models or platform |
| Industry integrators | Know the processes, can connect ERP/MES, familiar with tendering and acceptance | Model capability often external; subcontracting risk | Prime contractor |
| Vertical specialist teams | Close to the site, responsive, flexible, willing to start with one scenario | Long-term delivery and operations need verification | Scenario implementation |
Policy targets are specific: the eight-department Action Plan on "AI + Manufacturing" issued in December 2025 targets 1,000 advanced industrial agents, 100 high-quality industrial datasets and 500 typical scenarios by 2027. That is both an opportunity and a warning — a wave of new firms will label themselves "factory agent" vendors, so screening matters more than searching.
The five screens
- Referenceable similar sites. Ask for three, and on the calls ask how data access was handled, who set the definitions, and whether it is still running in year two. Vague scenarios and no site numbers mean hold off.
- Data-access plan. Official API, read-only account or scheduled export: state which works and which does not. Where none works is out of scope this round.
- Definition ownership. The plant decides business definitions; the vendor registers and reconciles. Ask to see what a definition register looks like — metric, definition, source, field, refresh time, owner.
- Operational ownership. Model updates, interface changes and server upkeep, named in the contract with a contact.
- Exit terms. How data is exported, accounts returned, the system shut down. Otherwise you are tied to one vendor.
Four things to check on a site visit
Once the shortlist is down to two or three, arrange a site visit and check four things: data access actually running, demonstrated live in your environment rather than shown on a recording; a definition register genuinely in use, since an empty template proves nothing; how exceptions are reviewed, by whom, how often, and how false positives are handled; and the named operations contact with response times that appear in the contract. If all four can be verified on the spot, the vendor has been doing this for a while. If any is deferred to "we will confirm later", write it down and keep asking.
Three warning signals
- Clearly templated content. The same article with city and company names swapped describes a script, not your site.
- Architecture diagrams with no site numbers. Tidy layers, but no table names and no owner for definitions.
- Signing before a site visit. Surveying is where a workable plan starts; demanding a deposit before coming to the plant shifts risk onto the buyer.
Verifying from public records
There are more routes than most buyers expect, and they are free:
- Tender and award notice platforms. Search by company name for project names, values and dates, and see whether the firm primes or subcontracts.
- Client-site figures in press coverage. Midea's Meiyun Zhishu, for example, published its agent-factory solution in January 2026 covering the whole value chain with 158 core scenarios — the kind of scenario count that can serve as a benchmark.
- Policy documents and scenario lists. Shenzhen's "AI + Advanced Manufacturing" action plan for 2026–2027, issued in February 2026, and five batches of nearly 300 published "city + AI" scenarios help with orientation and comparison.
Is the vendor really in manufacturing, or newly arrived?
Three checkable clues. First, the original business: firms from industrial software, equipment integration or automation move into plant projects naturally, while those from consumer internet or general-purpose models usually need a first project to find their way. Second, the industry and age of the cases: one client still citing the work two years later says more than ten new logos in one year. Third, how much plant vocabulary appears in the proposal — inspection sheets, work orders, batches, changeovers, overall equipment effectiveness (OEE) used correctly indicates someone who has been inside a plant. All three are visible from public information without the vendor's cooperation.
On judging whether a plan is workable at all, see our custom delivery and enterprise AI agent deployment pages. If you have a shortlist, send us the company names and scenarios (contact) and we will return a written view, at no cost.
In one line: screen factory agent vendors by the sites they have built and the numbers from those sites, then by the data-access plan, then by who operates it. Once all three hold, price is worth discussing.
FAQ
- How do vendors in this space divide up?
- Roughly three groups: platform vendors with models and compute, delivering at platform or programme level; industry integrators who know the processes and can connect existing systems; and small specialist teams close to the shop floor. In practice a smaller plant often sees an integrator prime, a specialist build the scenario and a platform vendor supply compute or models. Establish who signs and who answers for results first.
- How do I know a company has really done factory projects?
- Three routes together: search tender and award notice platforms by company name for project names, values and dates; ask for three referenceable similar projects and, when calling, ask how data access was done, who set the definitions and whether it is still running in year two; and check public reporting for client-site numbers such as downtime, inventory or how much earlier anomalies were caught. If none of the three is available, hold off.
- What should the first conversation settle?
- Three things: the data-access route for this round (official API, read-only account or scheduled export — which works), what the plant must provide in staff and permissions, and who owns model updates and interface changes during operation. These can be asked in the first meeting; vague answers here usually mean a vague plan later.
- Why does operational ownership matter so much?
- Because the operating phase is where cost and risk concentrate: model versions update, interfaces change, servers fail and definitions shift. Where public projects fall out of use in year two, it is usually not a broken system but an unassigned responsibility. Writing it into the contract with a named contact is worth more than trimming a few points off the price.