AI agents can now read emails, query systems and draft work, but most manufacturing pilots still stall. Five use cases that are delivering real value in 2026, the architecture that keeps them safe, and the guardrails that decide success.
A sales engineer at a mid-sized contract manufacturer starts Monday with 37 unread RFQs. Some are complete drawing packs; some are a single PDF and a line saying "please quote 5k". By the time she has worked out which ones are worth quoting, which material grades are realistic and which customers have paid late before, half the week has gone. The work that wins business, the engineering judgement, gets squeezed into Thursday.
That bottleneck, skilled people spending most of their time gathering, checking and re-typing information, is exactly where AI agents are earning their keep in 2026. It is also where most pilots go wrong.
What an agent is, and what it is not
A chatbot answers questions. An agent does work: it reads inputs such as emails, drawings and spreadsheets, uses tools such as ERP lookups, document search and calculators, decides what to do next, and produces an output for a person to approve. The difference matters, because once software can act, the questions change from "is the answer good?" to "what is it allowed to touch, and who checks it?".
We covered the foundations in using ChatGPT carefully and AI governance for engineering firms. Agents raise the stakes on both.
Five use cases that are working now
- RFQ intake and triage. The agent reads incoming enquiries, extracts part numbers, materials, quantities and due dates, checks the customer's history in the CRM, flags missing information and drafts a reply asking for it. Engineers receive a clean, ranked queue instead of an inbox. We described the wider quoting workflow in automating RFQs and quotes.
- Non-conformance reports. From an operator's photo and a few words, the agent drafts the NCR, pulls the drawing revision and inspection plan, suggests likely causes from past NCRs on similar parts, and routes it to the right engineer. Time to raise a complete NCR typically drops from 30–40 minutes to under ten.
- Supplier follow-up. The agent monitors open purchase orders, chases suppliers for confirmations and delivery updates in plain email, and escalates only the exceptions to the buyer.
- Maintenance triage. Operators describe a fault in their own words. The agent searches manuals, work-order history and the condition-monitoring data, then proposes likely causes, parts and a draft work order.
- Engineering knowledge retrieval. Asked "what draft did we use on the last textured PC/ABS housing, and why?", the agent searches design reviews, DFM reports and lessons learned, with sources. This builds directly on the engineering knowledge assistant approach.
The architecture that keeps agents safe
The pilots that succeed share the same design principles:
- Read before write. Start with agents that read and draft, never ones that update the ERP, release orders or email customers on their own. Add write permissions one action at a time, with evidence.
- A human approval gate on anything external or irreversible. Customer replies, supplier commitments, quality dispositions and anything with a dollar value are approved by a person until the error rate is proven.
- Least-privilege access. The agent gets its own account, with access to the folders and systems the task needs and nothing else. Never let it inherit an administrator's permissions.
- Grounded answers with sources. Every claim about a part, a price or a specification must link back to the record it came from. If the agent cannot cite it, it should say so.
- Full audit logs. Record what the agent read, what it decided and what a person approved. When something goes wrong, and occasionally it will, you need to see why.
Why most pilots still stall
- The process was a mess before the agent arrived. An agent that automates a broken process produces broken output faster. Map and simplify first; the principle is the same as Lean before automation.
- The data is not where people think it is. The real knowledge lives in email threads, personal spreadsheets and people's heads. Budget time to find and clean the sources.
- Nobody owns it. An agent is a digital team member. Someone has to review its work, tune its instructions and decide when it has earned more autonomy.
- Success was never defined. "Save time" is not a measure. "Cut RFQ triage from three days to four hours, with no missed due dates" is.
A realistic first project
Choose one workflow that is high-volume, rule-heavy, text-based and low-risk if a draft is wrong: RFQ triage and NCR drafting are the most common starting points. Run it in shadow mode for two to four weeks, with the agent drafting and people doing the work as usual, and compare the two every day. Move to assisted mode only when the drafts are consistently right. Most well-scoped pilots prove or disprove their value within six to eight weeks, for a modest cost compared with the hours they recover.
The bigger picture
Agents will not replace your engineers, buyers or quality team. The firms getting real value are the ones using agents to take the gathering, checking and re-typing off skilled people, so that their time goes on judgement, customers and improvement. That is the same goal continuous improvement has always had, which is why we see kaizen in the age of AI as an evolution, not a replacement.
The question to ask is not "where can we use AI agents?" It is "where are our best people spending hours on work that does not need their judgement?" Start there.
