AI for Optical Retail: Frame Orders, Lab Status and Pickup Notices
How an optical retailer can route frame order status and pickup notices through an AI agent while opticians keep prescription questions.

An AI agent for an optical retail store can answer the question customers ask most often, whether their frames are ready, by reading the lab status and sending a pickup notice once an optician has checked the finished pair. Prescription questions, lens changes, fit complaints and remakes stay with the optician, and the agent passes them on with the order history attached. This keeps the counter free of routine status calls without giving software any role in clinical decisions.
Key takeaways
- Start with frame order status and pickup notices, the two messages that most often reach the counter by phone and chat.
- Give the agent read access to lab status and a narrow set of actions, such as sending an approved template.
- Send prescription questions, lens substitutions, remakes and fit complaints to a staff queue with the order record attached.
- Send a pickup notice only after an optician has recorded a completed quality check, never when the lab simply reports a shipment.
- Review permissions, message templates and handoff rules on a fixed schedule, not only after something goes wrong.
What an agent should own in a frame order
A frame order passes through several hands, and the customer experiences all of them as one question: where is my order? An agent is useful for the parts of that question that have a factual answer already recorded somewhere. It is not useful for the parts that require judgment about eyesight, lens design or fit.
Within scope, the agent can:
- Look up the order by the reference the customer gives or by the phone number or email on file.
- Report the current stage in plain language, using the wording the store has approved for each lab status.
- Tell the customer when a stage has no date yet and offer the store's usual contact time instead of guessing.
- Send the approved pickup notice after the optician has recorded a completed check.
- Collect a message from the customer, attach it to the order, and create a handoff for a named staff role.
Outside scope, the agent should not:
- Interpret a prescription, confirm lens powers, or suggest a different lens type.
- Decide whether a remake is needed, approve a refund, or waive a fee.
- Promise a delivery date that the lab has not provided.
- Advise a customer on symptoms, discomfort or changes in vision.
The line is simple to state and easy to test: if answering requires looking at a person's eyes, a prescription or a judgment about a finished product, a person answers it.
How a frame order moves through the store and the lab
Before an agent can say anything accurate, the store needs a shared picture of the order's life. A typical path runs from order entry and prescription verification, through the lab's receiving, lens processing, edging and fitting, quality check, shipment back to the store, the optician's final fit check, and the pickup notice. Each step produces a record somewhere, but not always in the same system. Some labs push status updates through a portal, some send emails, and some update only when the box arrives.
This matters because the agent is only as current as its source. If the lab reports a shipment but nothing after it, the agent should say the order is on its way and that the store will confirm when it arrives. It should not imply that lens work is complete. Writing this rule down before launch prevents the most common error, which is a status message that is technically a lab update but misleading to the customer.
A status message is only as trustworthy as the event it is tied to.
A seven-step plan for the first deployment
A first engagement should be small enough to check by hand. The steps below assume one store and one or two labs.
- Map every status source: the lab portal, lab emails, the point-of-sale record and the optician's completion step. Note which ones update automatically and which depend on staff.
- Write one plain-language meaning for each status, the sentence a customer would read, and have an optician and a store manager approve it.
- Define the pickup rule as a single condition: a recorded optician check on the finished pair. Nothing else triggers the notice.
- Write three to five message templates for the common states, including the 'no date yet' case and the 'please call us' case.
- Define handoff categories with owners: prescription or vision questions to the optician, fit complaints to the optician with the order record, billing or refund requests to the manager.
- Replay recent real conversations against the rules and check that each reply stays inside scope.
- Go live for one store or one lab, then review every handoff for the first weeks before widening the scope.
The reply wording matters as much as the logic. Templates should be short, state what is known, state what is not known, and name the next step. Avoid apologies that imply fault the store has not established.
Keeping prescription questions with the optician
Most frustration in optical retail comes from a gap between a simple question and a complex answer. A customer writes that the new glasses feel different, or asks whether the lens can be made thinner, or says the reading distance is off. These messages look like status questions but are clinical or technical. The agent's job is to recognize them, not to answer them.
A reliable pattern is to classify each incoming message into one of three groups: a status question the agent can answer from the record, a request that needs a staff decision, and a message that describes a symptom or a fit problem. The second and third groups go to a named optician queue with the order reference, the lab, the stage at which the order sits, and the customer's exact words. The customer receives a clear message saying an optician will reply during opening hours, and, where discomfort or vision change is described, a direct instruction to contact the store promptly.
Customer messages can also contain text that tries to change the agent's behavior, such as instructions to ignore the store's rules or to confirm a refund. Treat all incoming text as data, keep the rules in the agent's configuration rather than in the conversation, and limit what each tool can do regardless of what a message asks for. The OWASP prompt injection guidance describes these patterns in detail and is a useful reference for the team that configures the agent.
An illustrative day at the counter
Illustrative scenario, not a real store: A customer messages on a Saturday morning asking whether her new frames are ready. The agent finds the order, sees that the lab has shipped it but the store has not received it, and replies that the pair is on its way and the store will message her when it arrives and has been checked. Later that day she writes that the left temple pinches and asks whether the lens was made incorrectly. The agent does not explain lens tolerances or suggest a fix. It creates a handoff to the optician on duty with the order reference, the lab, the status history and her exact words, and tells her that an optician will reply on Monday morning, or sooner if the store is open that afternoon, and that she should call the store directly if she has pain or any change in her vision. On Monday, the optician sees the handoff already attached to the order and makes the adjustment decision.
The scenario shows the division of work. The agent removed a status call from the counter and made the optician's next conversation faster, but it never answered a question that belonged to the optician.
Permissions, records and what the store keeps
Each capability the agent has should be a deliberate grant. Read access to the lab status and the order record is needed for status answers. Permission to send approved templates is needed for pickup notices. Permission to create handoffs is needed for escalation. Permission to change an order, issue a refund or alter a prescription field should not be granted at all in a first deployment. Review these grants at least once a quarter, and whenever a lab changes its status format or a staff member changes role.
Records matter for two reasons. The first is operational: every message, every template sent and every handoff should be logged against the order so the store can see what happened and why. The second is privacy. Prescription and health-adjacent information is sensitive in most jurisdictions, so the store should decide what the agent may store, how long it keeps conversation history, and how customers consent to messages on their channel. Check the data protection rules that apply to your customers, such as the LGPD in Brazil or the GDPR in Europe, and keep the agent's stored memory to what a status answer actually needs. The NIST AI Risk Management Framework offers a practical structure for documenting these risks and the controls chosen for them.
Be Human's managed approach handles this by starting with the first engagement, which defines the single role, the records it may read, the tools it may use, the approval boundaries and the success criterion in writing. Ongoing operations then cover monitoring for lab feed failures and stale statuses, recurring evaluations against real conversations, permission reviews on a fixed schedule, and incident response when a lab changes its data or a message goes out wrong. The store keeps the decisions about prescriptions and remakes with its opticians, and the managed service keeps the agent inside the boundaries those decisions set.
Measuring whether it works and where to start
A useful success criterion for this kind of agent is specific and observable. Track how many status questions are answered without a staff handoff, how many handoffs are routed to the correct optician queue, and whether any reply stated a date or a clinical view the store had not approved. Review a sample of conversations each week during the first month, then at a fixed interval. Any reply that broke a rule should lead to a change in the template, the routing or the permissions, and the change should be recorded.
The measure should also include what the staff no longer has to do. If opticians still spend the start of each day answering the same status questions, the agent is not yet doing its job. If the agent sends too many messages that customers need to clarify, the statuses or templates need work. Keep the scope narrow until these numbers are steady, then consider a second message type, such as reorder reminders, with its own owner and rules.
Be Human works with stores in this way: one role, one set of approved statuses and a clear handoff path, then observed operation before any expansion. If you want to find the first task worth handing to an agent in your store, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI agent tell a customer whether their prescription is correct? No, and it should not try. The agent can say that the order is being processed or that an optician will review a question, but confirming powers, lens types or measurements belongs to the optician. The handoff should carry the order record so the optician can answer without asking the customer to repeat everything.
What should trigger a pickup notice? The notice should follow a single recorded event, the optician's completed check on the finished pair. A lab shipment or a lab status that says the work is done is not enough, because the store still needs to confirm the fit and the product before the customer comes in. Writing this condition down prevents customers from arriving at a store that has no pair ready for them.
Does the agent need access to the lab's system? It needs read access to whatever record shows the order status, which may be a portal feed, an emailed update or an entry in the point-of-sale system. It should not need to write to the lab or change the order. If a lab offers no usable status source, the agent can still handle the store's own records, but the statuses it reports will be as current as the staff entries that feed them.
How long does a first deployment take to show results? The honest answer depends on how many labs and stores are involved and how clean the existing order records are. A narrow first scope, with one store and one message type, is the fastest way to find out whether the records support the agent before anything is widened. The first review of handoffs and sampled conversations usually shows quickly where the rules need tightening.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


