AI for Manufacturing: What Works in a Plant With 20 to 500 People
Most AI projects in manufacturing stall because the model has no clean feed from the ERP, the MES or the shop floor. Vascoh builds the integrations and workflows that connect AI to the systems a plant already runs.
of manufacturers already generate measurable value from AI, while only 20% of use cases are scaled across sites or the enterprise.
Source: Deloitte, AI in Manufacturing 2026of manufacturers cite quality as a function where AI has high impact, ahead of production (57%) and logistics and supply chain (49%).
Source: Deloitte, AI in Manufacturing 2026manufacturing firms operated in the U.S. in 2022, and about 98% of them are small businesses with fewer than 500 employees.
Source: National Association of Manufacturers, Facts About Manufacturing (2026)How is AI being used in manufacturing?
The use cases with the clearest payoff are narrow ones: reading supplier documents, predicting late orders, flagging quality drift, and drafting quotes from past jobs. Deloitte's 2026 research ranks quality (62%) and production (57%) as the functions where manufacturers see high AI impact, with logistics and supply chain at 49%.
The pattern behind a working deployment is the same each time. A model needs structured input from a system of record, a defined decision to make, and a place to write the result back. Without the write-back, the output lives in a chat window and nobody acts on it.
- Document intake: purchase orders, certificates of conformance and packing lists parsed into ERP records
- Quality: inspection images and SPC readings screened for drift before a lot ships
- Planning: late-order risk scored from routing, material availability and machine history
- Quoting: similar past jobs retrieved to draft a price and lead time for an estimator to review
Why do most manufacturers get stuck after the pilot?
Deloitte reports that 84% of manufacturers already get measurable value from AI, yet only 20% of use cases are scaled beyond a single site or team. The gap is mostly plumbing. The pilot ran on an exported spreadsheet, and the production version has to read from the ERP, respect user permissions and survive a schema change.
Deloitte also lists high cost (43%), technical expertise gaps (35%) and resistance to change (35%) as the top barriers. Small plants usually have no data engineer, so the integration layer needs to be simple enough that the maintenance cost stays near zero.
Change management deserves a line in the plan. Operators and planners accept tools that save them effort and reject ones that add steps. Involve them in choosing the first workflow and let them veto outputs that look wrong.
What data does a plant need before it adopts AI?
Less than most vendors claim, but it has to be reachable. For most plants that means item masters, bills of materials, routings, work orders with actual hours, and purchase and receipt history. These tables exist in Epicor, Acumatica, NetSuite, Fishbowl, JobBOSS, Infor SyteLine and similar ERPs, and most expose a REST or SOAP API, an ODBC connection or scheduled CSV export.
Where the ERP is old, the practical route is a read-only replica or nightly extract into a small database, with the AI workflow reading from there. That keeps the production ERP untouched and gives you an audit trail of what the model saw.
A worked example: a fabrication shop receives supplier PDFs in several layouts. An extraction workflow reads each PDF, matches the vendor and part numbers against the ERP, and drafts the receipt. Anything below a confidence threshold lands in a review queue. The ERP stays the system of record, and the model only proposes entries.
Where does custom software fit next to off-the-shelf AI tools?
Off-the-shelf tools handle generic tasks well: transcription, summarizing, and general document reading. They struggle with plant-specific logic, such as a customer who requires a revision-level check on every drawing, or a routing that changes with material thickness.
Custom integration code carries that logic. It maps the model output to your part numbers, applies your tolerance rules, and creates the ERP transaction. The firm count matters here: the National Association of Manufacturers counts 239,265 U.S. manufacturing firms, about 98% of them small, so most plants are better served by a narrow build than a broad platform.
How do you keep AI output safe on the shop floor?
Treat every model output as a proposal. A person approves anything that changes a work order, a purchase order or a customer commitment, and the system logs the input, the prompt version, the output and the approver.
Set a rollback path for each workflow. If the parsing step starts misreading a supplier's new invoice layout, the workflow should fall back to manual entry and alert someone rather than writing bad data into the ERP.
Privacy and IP matter too. Keep drawings, pricing and customer data inside your own cloud account or a vendor tier with contractual no-training terms, and restrict which documents each workflow can read.
- Human approval on any record that moves money or inventory
- Versioned prompts and logged inputs for audit
- Confidence thresholds that route low-confidence items to a queue
How a project runs
From first call to working system.
Pick one decision
Start with a single repeated decision that costs real hours, such as keying supplier invoices or answering order-status calls. Define the input, the output and who approves it.
Connect and build
Vascoh builds the read and write integrations against your ERP or MES, plus the model workflow, with approvals and logging in place from the first release.
Run and measure
The workflow runs beside the manual process first. You compare outputs, tune thresholds, then switch over once the error rate is acceptable to the people who own the process.
Questions
Common questions
How is AI being used in manufacturing?
Common uses are document intake, quality screening, late-order prediction, demand and maintenance forecasting, and drafting quotes from past jobs. Each one needs a data feed from the ERP, MES or a sensor system.
Which AI is best for manufacturing?
No single model wins. Quality inspection uses computer vision models, forecasting uses statistical or machine learning models, and document or language tasks use large language models. The deciding factor is integration with your ERP and data quality.
Can a small manufacturer afford AI?
Yes, if the scope is one narrow workflow connected to an existing ERP. Plants that try to buy a full platform first usually stall. Starting with one decision and one integration keeps cost and risk contained.
What is the 30% rule for AI?
It is an informal heuristic that suggests automating about 70% of a task and keeping humans on the 30% that needs judgment. It is a rule of thumb and has no standards body behind it. For manufacturing, the safer approach is human approval on anything that changes inventory or commitments.
Do I need to replace my ERP to use AI?
No. Most ERPs expose APIs or exports that an AI workflow can read from and write to. Replacement is only worth discussing when the ERP has no usable interface at all.
Related
Related problems.
AI Agent for Manufacturing: Automation That Reads and Writes Your ERP
An agent that cannot touch your ERP or MES is a chatbot.
AI Production Scheduling for Small and Mid-Size Plants
A planner rebuilding the schedule in Excel every morning is the bottleneck nobody put on the capacity chart.
Manufacturing Software Integration: ERP, MES, EDI and Accounting
A typical plant runs an ERP, an accounting package, a quality system, spreadsheets and customer portals that do not talk to each other.
MES ERP Integration: Shop Floor Data Into Your ERP
When the shop floor reports to one system and finance plans in another, work orders close late and costs are guesses.
Manufacturing Software Development Company for Plants Under 500 People
Your ERP covers 80% of the plant, and the other 20% runs on spreadsheets, whiteboards and email.
More in Manufacturing.
Contact
Tell us what needs to talk to what.
Describe the systems and the manual work, and we will tell you what is realistic to build and what is not.