Generative AI in manufacturing SMEs: two places an AI-trained intern could start

Industry AI playbooks

Most generative AI advice aimed at manufacturers is written for companies with an IT department. Small and medium-sized manufacturers rarely have one. What they have is a handful of people who know how everything works, a filing system built over decades, and no spare capacity to run an experiment that might not pay off.

So the useful question is not "what could AI do for manufacturing". It is narrower: which specific tasks, in a company of this size, are drafting tasks? Because drafting is what generative AI is good at, and drafting is where a mistake is caught by the person who reviews it rather than by a customer.

Two candidates come up repeatedly. Both are described below as possibilities to evaluate, not as results to expect.

First, the four skills an AI-trained intern brings

Before the applications, the underlying capability. APRO interns complete a four-week, project-based practicum in generative-AI ways of working, covering four areas:

• Document drafting and translation support. Preparing business documents in English and Japanese, materials for overseas customers, and quotation documents with generative AI.
• Automating routine tasks. Combining AI tools with spreadsheets to handle repetitive administrative work and data organisation.
• Using AI development tools. Working with tools such as Claude Code to build small tools and prototypes suited to a specific task.
• Organising records and knowledge. Converting meeting and discussion records into structured document drafts that are easier to search and share.

Note what is absent from that list: engineering judgement, commercial judgement, and any claim to domain knowledge. Those belong to the company.

Application one: documenting what experienced engineers know

Every small manufacturer has the same exposure. A small number of people carry knowledge that exists nowhere in writing, and the written procedures that do exist were often produced under time pressure years ago.

The academic framing for this is the SECI model of organisational knowledge creation, set out by Ikujiro Nonaka and Hirotaka Takeuchi in The Knowledge-Creating Company (1995). Its externalisation quadrant describes exactly this problem: converting tacit knowledge, the kind held in an experienced person's hands and habits, into explicit knowledge that can be shared. The obstacle has never been the concept. It is that externalisation is slow, and the people whose knowledge is being captured are the same people who are busiest.

One possible application: recording structured interviews with experienced engineers, then using generative AI to draft manuals and procedure documents from those records. The intern runs the interviews and the drafting. Review and approval of the content is always carried out by the company's own engineers.

What this could change is the cost of the first draft. What it cannot change is who decides whether the draft is correct. A procedure document that has not been checked by the engineer whose knowledge it encodes is not documentation. It is a transcript with formatting.

The bottleneck was never writing the manual. It was finding a week in which the person who knows the answer could write it.

Application two: groundwork for quotation and cost estimation

Quoting is a judgement task sitting on top of a retrieval task. Before anyone can price a job, somebody has to find the comparable past jobs, pull the relevant drawings, and assemble what is known about materials, tolerances, and the last time something similar was made.

That retrieval and assembly layer is where the hours go, and it is the part that is legible to a generative AI tool.

One possible application: organising information from past job records and technical drawings as groundwork for preparing quotations. Judgement and the final estimate rest with the company's own staff.

This is worth being precise about, because it is the application where an overclaim would be most expensive. Nothing here suggests that generative AI should produce a price, verify a tolerance, or decide whether a job is worth taking. Those are commercial and engineering decisions with money and liability attached to them. The candidate contribution is the preparation of material a human estimator then works from.

The safeguards that make this viable

The two applications above are only sensible inside a set of constraints. In an APRO practicum these are fixed, not negotiable:

The third of those is the one that carries the most weight, and it is worth stating plainly rather than burying in terms. Generative AI produces confident text regardless of whether the underlying content is right. In a manufacturing context, confident and wrong is a worse failure mode than obviously incomplete, because it survives a skim. Human approval is not a formality attached to the end of the process. It is the process.

What a four-week practicum actually looks like

APRO interns complete a four-week, project-based practicum, remote-first. During it they work on real projects, task completion is checked every week, and a pitch and communication block is mandatory. On completion, each intern receives a certificate recording the content of the practicum and what they achieved.

For a host company evaluating this, the honest framing is that four weeks is enough to produce drafts and a documented method, and not enough to deliver a finished internal system. The output to expect is a starting point that a company's own staff can carry forward, plus a clearer view of whether the approach fits the way the company actually works.

Where to start if you are considering it

Pick the narrower of the two applications. Documenting one process, from one engineer, is a better first trial than an attempt to capture a department, because the review loop closes fast enough for someone to tell you whether the output is any good while it is still cheap to change course.

Then decide the review path before the drafting starts. Who checks the draft, against what, and how long they have. If that question does not have an answer, the drafting will finish and the documents will sit unapproved, which is the most common way these efforts stall.

---

Generative AI in a small manufacturer is not a systems project. It is a drafting assistant applied to two or three tasks where the first draft is the expensive part and the review is already someone's job.

→ Companies considering hosting an intern can book a call or read the host-company overview.