Glass Shop AI Agent: Standardize Custom Cut Measurements First

An AI agent for a glass shop can check a custom cut before anyone approves it, but only when every order carries the same measurement fields, units, reference points and revision records. Standardizing those details before any software is connected gives the agent a fixed standard to compare each order against. Without that baseline, the agent can only repeat the guesses already built into the shop's paperwork.

Key takeaways

• Define one reference point and one unit system for every custom cut before adding an agent.
• Store glass type, thickness, edge work and tolerance as separate fields, never as free text in a note.
• Record where each measurement came from, who took it and when, so a remeasure can be traced.
• Let the agent flag mismatches and hold orders, while a named person keeps the release decision.
• Start with one product line, such as shower enclosures or mirrors, and widen the scope only after the checks hold on real orders.

Why custom glass orders go wrong before the saw

Glass leaves little room for correction. A panel cut a few millimeters off cannot be trimmed back into shape. It has to be remade, and the remake waits on material and a second production slot. Most errors begin earlier than the cutting table. A width is taken from a rough opening, a dimension lands on a sticky note, a sketch is not updated after a door swing changes, or a contractor's phone order uses a different starting edge than the office expects.

Experienced staff usually catch these problems by habit. That habit is valuable, but it does not transfer well to a busy week, a new estimator or a second location. An agent does not remove the risk on its own either. It can only compare what it is given. If the order says 1200 and the site note says 1195 measured from another edge, the agent needs a rule that says which reading governs and what to do when the two disagree. That is a records question before it is an AI question, and it is the right place to start.

Which measurement fields to standardize first

Every custom cut should carry the same set of measurement fields, filled in the same way, whoever takes the reading. The list below is a practical starting point for a shop that makes glass to order for openings, frames or fixtures.

• The reference point for each dimension, such as the finished face of the opening or the inside edge of the frame, written as one defined term used everywhere.
• The unit and the rounding rule, including whether dimensions are kept in millimeters or inches and to what precision.
• Width and height readings at more than one position, with the smallest and largest values both stored.
• Diagonal readings for any opening that is not a true rectangle, so the agent can see racking before the part is cut.
• Glass type, thickness, coating and edge finish as their own fields rather than inferred from a product name.
• Cut-outs, drilled holes, notches and corner radii, each with its position stated from the reference point.
• The tolerance the shop allows for that product, written as a value the agent can compare against.

The test for each field is simple: could someone who has never seen the job check it without asking a question? A field called "size" fails that test. Two fields called "opening width, top" and "opening width, bottom" pass it, and so does a tolerance written in the product rules rather than remembered by a supervisor.

Which order records an agent needs to see

A measurement set is only useful when it is attached to the order that uses it. Each custom cut should point to one measurement record, one sketch or drawing revision, one customer-approved revision and one production status. When those links exist, the agent can ask the questions a careful estimator would ask: has the drawing changed since the measurements were taken, and did the customer sign off on the version being cut?

The order log should also keep a short change history. Each entry should show what changed, who changed it and the reason in plain words, such as "customer moved hinge side after site visit." An agent reading that history can flag an order where the sketch was revised but the measurement set was not, without inferring it from a message thread. Keep customer notes separate from production notes, because a mix of both is hard to check and hard to audit later.

Checks an agent can run before a cut is approved

The checks below are deliberately ordinary. They are the comparisons a careful reviewer would make, written in a sequence an agent can repeat on every order. Each one either passes quietly or produces a flag with the mismatched fields shown side by side.

• Confirm the measurement set is complete, with every required field present and its unit stated.
• Compare the ordered dimensions with the measured dimensions and flag any difference beyond the tolerance for that product.
• Check that the sketch revision matches the latest revision the customer approved in the order log.
• Confirm that every cut-out sits inside the panel after the edge clearance defined in the product rules.
• Compare the glass specification with the product rules, so a panel that requires a specific glass type is not released with a different one.
• Hold the order and open a review task for a named person when any check fails, listing the conflicting values together.
• Write the result to the order log with the check version and the time it ran.

Those steps are only as trustworthy as the testing behind them. Anthropic's guidance on evaluating AI agents recommends building test cases from real work, which for a glass shop means past orders, including the ones that were remade, and confirming that the agent flags them. Run the checks against a sample of closed orders first. If the agent misses a known error, fix the rule or the record before it touches live work.

What the agent should never approve on its own

An agent in a glass shop should hold, flag, draft and prepare, but not release. Cut approval, changes to a tolerance, remeasure exceptions and any commitment to a customer about remake costs belong to a named person. The NIST AI Risk Management Framework treats human oversight as something defined in advance for each risk, and that is the right way to think about these boundaries. Permissions should follow the same logic: read access to orders and measurements, write access only to flags and status notes, and no path for the agent to change a dimension in the production record.

Illustrative scenario: A shop takes a phone order for six shower panels. The site note gives the opening width at the top of the frame, but the sketch was drawn from the bottom reading. The agent compares both readings, finds a difference beyond the product tolerance, and holds all six orders. It opens a review task for the estimator with both readings, the photo reference and the affected order numbers. The estimator revisits the site, records the corrected measurement with a new timestamp, and releases the orders. The agent never edited a dimension.

The scenario is invented for illustration. The point is the division of work: the agent catches the mismatch quickly, and a person decides what the correct dimension is.

How Be Human handles this in a managed setup

Be Human starts with one useful role for the glass shop rather than a general assistant. The first engagement defines the measurement and order fields that role reads, the tools it can use, the approval boundaries and one observable success criterion, such as every out-of-tolerance order being held before a cut is released. Managed operations then keep that setup working: monitoring the checks, running recurring evaluations against new orders, reviewing permissions when staff or tools change, and making controlled improvements to the rules when the shop's products change. The same pattern applies to adjacent custom work, such as Superintelligence Prep for Custom Cabinet Shops: Quotes and Shop Drawings, where field measurements and drawings must agree before fabrication starts. Custom orders with tight tolerances also appear in Superintelligent Agents for Jewelry Stores: Repairs and Custom Orders, where the record of each specification matters as much as the work itself.

If you are not sure which custom cut task should come first, the free Be Human AI diagnostic will help you find the first task worth handing to an agent.

Frequently asked questions

Can an AI agent approve a custom glass cut by itself? No, and it should not be set up that way. The agent can compare measurements, check revisions and hold an order when something does not match, but release of a cut stays with a named person. Keeping that boundary in writing makes the process easier to audit when something goes wrong.

Do we need to standardize our measurements before an agent can help? Yes, at least for the fields the agent will check. An agent can only compare values that exist in a consistent form, so a width written as a rough note in one place and a precise figure in another gives it nothing reliable to work with. Start with one product line and define its fields before widening the scope.

What if our installers and our office measure differently? Record both readings with their source, the person who took them and the time, then let the tolerance rule decide whether they conflict. The agent should not silently pick one. When readings differ beyond tolerance, the order should be held and a person should confirm the correct dimension on site or in the drawing.

Do we need new software before an agent can check orders? Not necessarily. Many shops already hold most of the needed information in an order system, a spreadsheet or a shared folder, and the first task is to see which fields exist and who can read them. What matters more is that the agent has defined access, a clear rule set and a place to write flags, so its work can be reviewed.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.