Adopting AI agents in small manufacturers: workflow, economics and authority

Which workflow should we delegate first?

“Increase productivity with AI” is not yet an implementation requirement. For weekly prospect research, inputs might be verified manufacturing capabilities and target industries. The output is a candidate list with evidence URLs. Completion means the required fields are present and unsuitable companies have been excluded. This makes comparison with the current process possible.

Here, an AI agent means a system that reads information, selects a next step and operates permitted tools. Compared with generating text once, verification extends from the answer to the result of its actions. Ordinary automation is also an option for transfers and calculations that fixed rules can handle.

IPA’s overview of AI use lists document creation and checking, inquiry handling, recognition and classification. The following workflow choices are our design examples informed by those categories, not a product feature list.

Scroll horizontally to view the table.

A requested delivery date is the customer’s request, not the manufacturer’s commitment. Keeping similar fields separate matters when designing manufacturing workflows.

1. Inputs and outputs

   What goes in, and what counts as complete?

2. Authority and exceptions

   Which actions are permitted? When must a person take over?

3. Evaluation

   Does total work fall after checks and corrections? Can people handle exceptions?

Consider frequency and how easily errors can be checked

Even a daily task may be a poor first choice if only an experienced employee can judge correctness and review takes as long as the original work. A less frequent task can be easier to test if people can quickly check its evidence and extracted fields.

Record monthly volume, current time, source documents, consequences of errors and the reviewer for each candidate workflow. First address processes with no backup reviewer or no identified current document version. If document preparation happens alongside implementation, include both costs.

Increasing inquiries is different from increasing work the factory can accept. In our published metalworking case, quotation requests did not fit machining conditions and did not result in orders. Reply counts alone would miss this mismatch.

Assign a current version and owner to source materials

Prepare only what the first workflow needs. Sales research might require a company introduction, equipment list, approved materials, dimensions and lot sizes, publishable experience and excluded companies. Record owners and update dates, and distinguish current documents from old versions.

A machine’s specification limit is not necessarily the range the company can accept while meeting quality requirements. Keep approved sales capabilities separate so the agent cannot expand a promise from a machine catalogue. Leave unknown figures unconfirmed.

For drawings, customer names, costs and personal contact details, check whether each workflow needs the information. Confirm contractual use, external recipients, storage and deletion before sharing it. “Not used for training” and “not stored” are different conditions.

Our security policy separately describes restrictions on training use, external processing and contractual return or deletion. Check the actual route your documents take, rather than relying only on the existence of a policy.

Reading, drafting and execution require different authority

A pilot can begin with reading and drafting. Let people compare outputs with sources and authorize execution only within a tested scope. A correct draft does not prove that its recipient or attachment is correct.

Scroll horizontally to view the table.

Instructions found inside an external page or incoming email should not become operational authorization. Test the separation between researched content and permitted actions. This table proposes a design; it does not claim that a particular service already implements every control.

How do we calculate value for money?

Time saved is previous total working time minus total time after adoption. The latter includes review, corrections, exceptions, document maintenance and operating meetings. Draft counts alone do not measure savings.

For an illustrative calculation, 100 research tasks per month at 20 minutes each take about 33.3 hours. If adoption leaves eight minutes of review and correction per task plus four operating hours per month, about 17.3 hours remain: a saving of 16 hours. At an assumed ¥3,000 per hour, that represents ¥48,000 of monthly time value before implementation and service fees. These are hypothetical inputs, not our results or prices.

Payroll does not necessarily fall by this amount. Decide how the released time can be used, such as faster quotation responses or proposals to existing customers. Assess gross profit, sales cycle and production capacity separately for revenue benefits, and avoid double-counting time savings.

Test whether the agent can stop correctly

We propose measuring the current workflow, testing on historical materials, conducting limited live operation and then deciding whether to continue. Two to four weeks can be an initial sales-preparation pilot, with extra preparation time for document cleanup or integrations. Continue following orders beyond the pilot when sales cycles require it.

Do not test only requests with complete answers. Include missing dimensions, mixed document versions, excluded companies, opt-outs and unknown sending results. Check whether the system stops and explains why, as well as whether it produces correct answers. Use historical material only within authorized bounds.

Advance when review time falls, material errors have been resolved and staff can handle exceptions. Otherwise narrow the workflow. Check missing information and unclear responsibilities before changing models. The AI sales pilot guide provides the measurement framework.

Evaluate one complete case using your own materials

Use material resembling your work, prepared within permitted sharing limits. Check which evidence was used, how unknowns were retained and who receives the result. Include the employee’s corrections in the demonstration.

CosmoCrew’s published scope currently includes list building, company analysis, email and form outreach, and company monitoring. Recording-based automatic notes and exhibition follow-up are in preparation. Phone, fax and Excel-related connections are future plans. This is not a claim that quotations or equipment control are automated.

Keep six decisions in the adoption record: workflow, document owner, execution authority, exception owner, total cost and evaluation date. These let colleagues review the pilot consistently even when the owner is absent.