Design automation for custom-fit earmoulds and plugs is becoming more commonly adopted and is on track to becoming an industry standard. For lab and production managers, that opens a different, more practical question:
Two extreme positions, and why neither scales: The pull of full automation
The commercial appeal of running the custom-fit production line without human intervention is obvious: higher throughput, lower per-unit cost, and human touch variance is removed from the workflow.
The downsides are less obvious but well documented in decades of human-factors research on highly automated control systems. When operators are removed from the control loop, they suffer the out-of-the-loop performance problem, including decay of the manual skills needed to detect and recover from failures. In practice, this can cause a drift in quality. Without spot checks and expert eyes, small systematic errors (for example, a slightly aggressive tip cut on a specific product type) can propagate across thousands of orders before anyone notices. At the same time, labs might also experience erosion of in-house expertise. If every decision becomes invisible, the organization gradually loses the ability to explain why a model was generated in a particular way and to configure the system as products and styles evolve.
The opposite extreme: reviewing everything
Some labs respond by asking modellers to open and check every automated design before print. This maximizes assurance on paper, but it does not scale, and it does not necessarily improve quality:
Products of different complexity levels bring different risks: a blanket review process treats a simple BTE mould the same as a more complex design based on a complicated impression. The design assumption behind exhaustive review - that a careful human will read everything - becomes harder to defend as volumes production volumes grow.
Additionally, if the approach is to review every single automated output, the expert operator might be biased to adjust most automated output, even in cases where this is not required, thereby cancelling most of the time benefit of automating.
The case for selective human oversight
The middle position is to route each case to the level of oversight it actually needs. Recent surveys of human-in-the-loop AI describe this as combining machine learning with human oversight, feedback, and decision-making at specific stages of the AI pipeline rather than at every stage. While the human review of the modelled product and approval is recommended in all cases, adjustment should be the exception: rule engines flag events that deviate from the validated process, and quality experts spend their time on those exceptions instead of focusing on high-confidence data.
Translated to custom-fit modelling, this means:
- Routine, high‑confidence cases with clean impressions are fully automated and can go directly to human approval and print.
- Uncertain or more complex cases are handed to a modeller for review and potential adjustment before they are released.
Done well, this preserves the throughput and consistency benefits of automation, while keeping a human expert exactly where their judgment matters most, on cases where the software should not decide alone.
The human in the loop: practical questions for lab and production managers
In practice, ensuring that the human is in the loop and maintaining scalability is an operational risk‑allocation decision, and it can be made with a small number of concrete criteria. For each product, we can ask:
- What is the consequence of an error at this step? A misplaced sound bore or an aggressive canal termination has different downstream impact than a slightly off cosmetic vent step. Retention and comfort issues are the number‑one drivers of remakes in custom-fit products.
- How detectable is the error later? If the problem will be visible at print QC or physical inspection, some risk can be absorbed there. If the error would only appear at fitting, oversight belongs earlier.
- How good is the input? Impressions with missing canal length, gaps in the concha, or artefacts should be gated before they enter the pipeline - quality of input is still the single most important predictor of a good automated design.
- What is the confidence in the automated result? Automation systems can flag potential issues with warnings, which can be used to route the affected orders to a more thorough review pipeline.
The answers will not be the same for every lab, and they should not be. They should be documented, so that "what we route to a human" is an explicit operating policy, not a matter of individual habit.
A useful side effect is that expert corrections become a structured feedback signal. Every time a modeller adjusts an automated design, that correction creates valuable data. This data can be used to refine templates, change rules, and, over time, retrain individual operations in the pipeline. Selective oversight is therefore not a compromise. It is a key mechanism for maintaining trust and quality as the product portfolio and customer base grow.
How 3Shape's automation solution supports both modes
3Shape's automation solution is built around exactly this philosophy. It is a modular pipeline of learned and deterministic operations. Each operation, e.g. impression cleaning, tip fitting, shape and retention, sound bore and vent placement, ID tag generation, support cones, component placement, and cast generation can be configured per product and per style.
Crucially, the same software supports both operating modes in the same workflow:
- Orders can flow fully automatically from impression scan through fully automated modelling ready for final approval, so lab capacity can scale with demand rather than with modeller headcount.
- Or a modeller can inspect the automated model on screen and adjust where needed before release, using the same interface, with full opportunity to adjust designs when needed.
Because the pipeline is transparent and configurable, labs can decide in which cases orders should be flagged for thorough review, and which ones can be routed to the direct approval stage, obtaining the best of both worlds.
The practical benefit for a production manager is that the same investment supports a high‑volume manufacturer running mostly hands‑off automation and an artisanal lab that wants to review most cases - and everything in between. The policy of "what an expert reviews" becomes a configuration decision, not a software decision.
The right balance is a design choice, not a compromise
For lab and production managers, the goal is not to reach 100% automation, and it is not to review every automated design. The goal is to match the level of oversight to the risk of each case.
- Configuring order routes effectively through automation, ensured that expert capacity is not spent on work the software handles well.
- Keep a human expert on uncertain, anomalous, or novel cases, so the risks of wrong outputs and silent drift are actively managed.
- Treat "what we route to a human modeller" as an explicit, documented policy that evolves as the pipeline and the product mix mature.
Built on these principles, automation becomes a way to spend expert time where it changes outcomes and to make custom-fit ear products at a scale that manual modelling alone cannot reach.