AI Assistant in Manufacturing: What It Does and Why Decisions Remain with Humans
Where in the manufacturing vertical does the AI assistant operate, what it does on its own, and what it only suggests, how human action confirmation is structured, how much it costs in credits, and what happens with data.
In the manufacturing vertical, an AI-powered assistant can be found in many places: recognizing a product from a photo, reading drawings, building a kitchen from a description, advising on how to reduce batch costs. This is convenient, and at the same time raises a valid question: what if it makes a mistake?
This article explains exactly where the assistant operates, what it does on its own, what it only suggests, and how the system ensures that decisions are made by a human.
Main Principle
The assistant prepares and suggests. The human reviews and confirms.
This principle is repeated everywhere the assistant is present. A recognized drawing remains a draft until a human confirms it. A suggestion to replace a defect awaits the foreman. Advice to change the time standard is applied only after the owner presses a button.
The reason is simple. The assistant is fast and doesn't get tired, but it doesn't know your factory as well as you do. And the cost of an error in manufacturing is spoiled metal and a missed deadline.
Where the Assistant Operates
Quick Estimate from Photo
You upload a photo of the product, and the assistant suggests a design: what parts the product consists of, their size, and material.
What a human does: reviews and corrects the design before building the model. Dimensions are also entered by a human, as the scale is not visible from a photo.
Learn more about Quick Estimate
Drawing Recognition
You upload a photo of a dimensioned drawing, and the assistant recognizes the profile, length, holes, and slots.
What a human does: checks every number and confirms. Until confirmed, the result has a draft status and does not go into production.
Learn more about Drawing Recognition
Design Chats
You describe a kitchen in words, and the assistant arranges modules from your catalog and shows the result in 3D.
What a human does: reviews, requests changes, coordinates with the client. The assistant does not consider the project complete if the review finds an error.
Cost Advisor
The Manufacturing Advisor analyzes the calculation and suggests how to make it cheaper: a different route, calibration of time models.
What a human does: decides whether to apply the advice. Calibration of time models is explicitly marked as requiring human approval.
Calculator from Your File
The assistant reads your price calculation table and suggests a calculator that reproduces its logic.
What a human does: compares the results on real products, saves, and enables it. Until then, prices do not change.
Learn more about the Calculator
Shop Floor Actions
The assistant can suggest an action to the foreman: replace a defect, reallocate work, change the batch route.
What a human does: confirms each action. The foreman's panel explicitly states this: the assistant can suggest these actions, but each is confirmed by a human.
Automatic Work Distribution
The automatic distribution button suggests who does what.
What a human does: unchecks what is not suitable and clicks "Distribute".
Connecting an External Assistant
In addition to built-in functions, an external AI assistant can be connected to the workspace, for example, one you already use for work. It gains access to workspace data via a special connection and can read orders, specifications, shop floor status, and prepare actions.
The owner configures the connection. They also determine what the assistant is allowed to do.
How Confirmation is Structured
Actions that the assistant can perform are divided by risk.
- Reading. View an order, find a part, calculate. Performed immediately: nothing changes from this.
- Safe changes. Create a draft, add a note. Can be performed immediately if the owner has allowed it.
- Changes affecting production or money. Start an order, write off a defect, publish a proposal. Go into the approval queue.
Approval Queue
An action that requires confirmation is not performed but waits. A person with the appropriate rights sees exactly what the assistant wants to do, with what data, and why. They approve or reject.
The number of pending actions is visible on the main workspace dashboard in the "AI Actions for Approval" block. An empty block means nothing is pending.
Delegation
The owner can pre-authorize the assistant for certain types of actions without separate confirmation: for a limited time and within limited boundaries. For example, allow it to independently design kitchens within a single project.
Delegation can be revoked at any time.
There is no way for the assistant to grant itself more rights. Rights are only given by a human.
What the Assistant Never Does
- Does not start a robot. It can prepare a welding program, but the cycle on the cell is started by the operator.
- Does not start welding parts without an assembly. This rule applies to everyone, including the assistant.
- Does not invent data. If the catalog does not have the required module or there is no price for the material, it states this, rather than substituting something similar.
- Does not declare work complete if a check found an error.
- Does not receive personal identification data of employees, such as document numbers.
AI Credits
The assistant's work is paid for with workspace AI credits. The remaining balance is visible in the top line of the screen, next to the lightning bolt icon.
What consumes credits:
- Photo and drawing recognition;
- Conversation in design chat;
- Translation of help articles;
- Responses from the built-in assistant.
What does not consume credits:
- Cost calculation based on your data;
- Building a 3D model from a confirmed design;
- Planning, cutting, work distribution.
That is, credits are only used for tasks where a language model is involved. Regular system calculations are free.
If credits run out, functions with the assistant stop working, and the system prompts you to top up the balance. The rest of the system works as usual.
What Happens to Data
- Photos for recognition are sent to the language model. A working copy remains in the browser tab, and a description of the design is stored on the server.
- Access keys to models are stored on the server. Your browser does not communicate directly with language models.
- The client in the proposal does not see the cost, margin, and suppliers, even if accessing via the API.
- Every action of the assistant is recorded in the log: what was done, when, and who confirmed it.
How to Use the Assistant Smartly
- Check numbers. Especially dimensions and quantities in recognized drawings. One wrong digit means a batch of defects.
- Provide accurate input data. Dimensions, material, requirements. The more precise the request, the fewer corrections.
- Do not enable delegation unnecessarily. Confirmation takes seconds, an error takes hours.
- Start small. First recognition and advice, then suggested actions, and only later delegation.
- Monitor credits. Recognizing a complex photo costs more than a simple one.
Frequently Asked Questions
Can the assistant mess something up?
Actions that affect production or money are not performed without human confirmation. The biggest risk is careless confirmation. Therefore, read what is being proposed.
The assistant suggested an action, but it wasn't performed.
It's waiting in the approval queue. Check the "AI Actions for Approval" block on the main dashboard.
Why does the assistant say something is missing instead of suggesting a replacement?
That's by design. Silent replacement is more dangerous than an honest refusal.
Credits ran out. Will the factory stop?
No. Only functions with the assistant will stop working. Orders, shop floor, planning, and calculations work without credits.
Can I work without the assistant?
Yes. Nothing in the core work depends on it: everything it suggests can be done manually.
Who sees what the assistant did?
The owner and administrators, in the action log.