
Superintelligence for Dental Prosthetics Labs: Orders and Deadlines
Learn how dental labs can log orders and per-piece deadlines so an AI agent reports status without promising dates, with clear rules and human handoff.
Want to apply this in your company?
Book a call with Juan →An AI agent in a dental prosthetics lab should find the order, read the stage each piece is in, and reply only with what the team has recorded: the current stage, the internal deadline in force, and any pending items blocking the work. It should not estimate dates the bench has not confirmed. The result depends less on how sophisticated the model is and more on an order and deadline record kept per piece, with history, owners, and clear escalation rules.
Key takeaways
- Every order needs its own identifier, and every piece needs a recorded stage, owner, and deadline.
- The agent reports the confirmed status and never creates a date the bench has not set.
- Deadlines change because of insufficient impressions, shade adjustments, and rework, so the record must keep the history of each change.
- Urgent requests, complaints, and clinical questions leave the agent and go to a person on the team.
- Before connecting the agent to WhatsApp, test it with simulated orders and an agreed success criterion.
Why a prosthesis status is hard to answer
A prosthesis goes through several stages: receiving the impression or digital scan, planning, fabrication, try-in with the dentist, adjustments, finishing, quality control, and dispatch. Each stage depends on the one before it. If the impression arrives with a defect, the work goes back. If the clinic takes a long time to approve the shade or the try-in, the piece stops. When a dentist asks "is it ready?", the honest answer is almost always "it is at this stage, and this is still missing."
That dependency explains why a single date given over WhatsApp often causes problems. The date may have been a guess by whoever answered, it may ignore the clinic's try-in, and it may then be repeated as a commitment. The agent needs to treat a deadline as recorded information, with a source date and an owner, not as a default answer.
What to record for each order and each piece
Before any agent is involved, the lab needs a record that the team keeps up to date every day. A minimal model usually includes:
- Order identification with an internal code, the responsible dentist, and the clinic, without exposing the patient's full name in the messaging channel.
- For each piece: type of work, material, chosen shade, number of units, and planning notes.
- Milestones with real recorded dates: impression received, try-in scheduled, try-in completed, dispatched.
- The current stage of the piece, with the internal person handling it.
- Pending items that block the deadline, such as an insufficient impression, a shade awaiting approval, or a reply expected from the clinic.
- The internal deadline in force, with the date of the last change and the reason for it.
- Who can change a deadline in the system, and a log of who made each change.
This record also matters for data handling. Health information is sensitive personal data under Brazil's data protection law (LGPD), and patient data from dentists usually reaches the lab in that category. The agent should therefore access only the minimum it needs to answer, and the record should not store clinical details beyond what production requires.
How the agent answers without promising a date
The conduct rules can be written as a simple decision guide that the team reviews before the agent goes live:
- If the piece has a current stage and an up-to-date internal deadline on record, the agent reports the stage and the deadline in force, making clear that it is the lab's current forecast.
- If the piece is stopped by a pending item from the clinic, the agent reports the pending item, explains what is missing, and hands the case to the team.
- If there is no recent record, the agent does not estimate. It says it will confirm with the team and opens a handoff with the context already summarized.
- If the order requires a fixed date, such as delivery for a scheduled appointment, the agent does not confirm. It logs the request and passes it to a person.
- If rework happened, the agent says the piece was adjusted and the team will update the deadline, without detailing internal failures.
- If the conversation turns into a complaint or a cancellation request, the agent stops automated replies and transfers the conversation.
When a conversation happens outside the 24 hour window of the WhatsApp Business API, contacting the customer depends on approved templates and prior permission. We cover this in our article on the WhatsApp 24 hour customer service window, and it is worth settling before any automatic "piece ready" notice is sent.
Illustrative scenario: the dentist asks about a crown
Illustrative scenario, not based on a real client. A dentist writes to ask about a crown from an order in progress. The record shows the piece is at the try-in stage, the try-in is scheduled for next week, and the clinic has not yet confirmed the shade. A suitable reply would be:
The crown is at the try-in stage, with the try-in scheduled for next week. We still need the clinic to confirm the shade. I will ask the team to confirm the timing with you.
The same information used badly would lead to a line like "it will be ready Friday." The agent has no way of knowing whether the try-in will be approved the first time. The difference between the two replies lies in the record and the rules, not in the ability to hold a conversation.
The same kind of stage-based triage appears in other sectors that depend on sequential steps, such as the one we describe in our article on order triage and deadline tracking for notary offices.
How to test the agent before it talks to dentists
Testing should use simulated data and real questions, drawn from the messages the front desk and the bench already receive. A simple plan:
- Collect a list of frequent questions and difficult ones, including requests for fixed dates and pressure to "confirm it quickly."
- Create test orders at every possible stage, including orders with pending items, rework, and changed deadlines.
- Check that no reply cites a date without a matching recorded milestone.
- Deliberately try to get the agent to promise a deadline, and confirm that it declines and hands off.
- Measure accuracy against the criterion agreed with the team, and review every incorrect handoff case.
Our guide on mock tools and sandboxes for AI agents explains how to build this environment without using real orders. For the evaluation itself, Anthropic has published engineering notes on evaluating AI agents, which are useful for thinking through test cases and approval criteria.
How we apply the method in this kind of operation
In our way of working, the path follows four steps: map the current flow of orders and updates, limit what the agent can read and do, pilot one type of piece with complete records and agreed metrics, and then transfer the operation and decide whether to expand, adjust, or stop. In a lab, this means starting with a small group of pieces, with human handoff from the first day. The decision to scale depends on what the logs show, not on an initial expectation.
The recommendation to keep deterministic flows where the rule is clear and to use the model only where interpretation is needed also appears in Anthropic's engineering writing on building effective agents. For the lab, this means the status lookup is a controlled flow, while natural language conversation stays in the layer that understands and responds.
Frequently asked questions
Can the agent give the dentist the delivery date for a prosthesis? It can share the internal deadline in force that is recorded in the system, identified as the lab's current forecast. It should not confirm delivery dates for fixed appointments or promise deadlines that depend on a stage not yet completed. When in doubt, the case goes to a person on the team.
Which patient data should stay out of the agent's channel? As a rule, the agent should work with an order code rather than the patient's full name, because the messaging channel does not need that information to report status. Detailed clinical information does not need to appear in replies either. Since health data is sensitive under the LGPD, limit access to the minimum necessary and record who accesses what.
How long does it take to put an agent like this into operation? It depends on how well organized the order record already is and how many stages the lab uses. Usually the work starts with mapping and defining the record, then moves to a pilot with one type of piece and tests using simulated orders. There is no universal timeline, and promises of a fixed timeline for this kind of rollout deserve skepticism.
What should we do if the agent makes a mistake and promises a date? The first step is to keep logs of every conversation so you can identify the case and the rule that failed. Then the team corrects the record or the agent's instructions, adds the case to the tests, and checks that the failure does not repeat at other stages. If the error reached a dentist, the team should contact them directly with the correct deadline and the reason for the change.