Why I Kept AI Out of the Most Important Decision in My Own Tool
By Dilum SenevirathneI’ve worn hearing aids since 2021. So when I decided to build a tool to help people navigate the OTC hearing aid market, I wasn’t approaching it as an interesting technical problem. I was approaching it as someone who has sat in that confusion himself.
The OTC hearing aid market opened up in 2022 and it’s genuinely useful for people with mild to moderate hearing loss who don’t need or can’t afford a full audiologist evaluation. But the buying experience is rough. Marketing language is thick, specs are hard to compare, and the number of options is overwhelming for someone who just wants to hear their grandchildren more clearly.
I wanted to build something that could cut through that. What I ended up learning in the process was less about hearing aids and more about where AI belongs in a tool that people are trusting with a decision that matters.
The design decision that shaped everything
The obvious approach would have been to let AI handle the whole thing. Describe your hearing situation, get a recommendation. Clean, simple, impressive in a demo.
I didn’t do that.
The AI in HearingShortlist does two specific jobs: it understands what a person describes in their own words, and it explains the results in plain language. That’s it. The actual selection, the ranking of devices against a person’s situation, is handled by plain, auditable code running against a fixed dataset of verified device specifications.
Same input, same output, every time. No inference, no interpretation, no variance.
The reason is simple. When someone is making a decision about their hearing, they need to be able to trust that the tool isn’t guessing. AI is extraordinarily good at understanding natural language and generating clear explanations. It is not the right tool for a decision that needs to be consistent and defensible. Those are different jobs, and conflating them is how you end up with a tool that feels smart but behaves unpredictably.
What the flow actually looks like
Here’s a simplified version of what happens when someone uses the tool:
User input (natural language)
|
v
AI interprets: hearing profile, lifestyle, budget range
|
v
Red-flag check: symptoms that need a professional?
YES --> Stop. Refer to audiologist.
NO --> Continue
|
v
Scoring engine (plain code, verified data, no AI)
|
v
Top matches ranked and returned
|
v
AI explains results using only facts from device dataThe red-flag check runs twice. Once by the AI, once by a simple set of rules in the code. If either flags something, the tool stops and points toward a professional. This isn’t a fallback. It’s a feature. The referral path is part of the design, not an edge case.
A before-and-after
This is roughly what a real interaction looks like.
Input: “I have trouble following conversations when there’s background noise, like at restaurants or family gatherings. TV is fine at home but I miss words when people aren’t facing me. My budget is around $200.”
What the AI extracts: Mild to moderate loss, difficulty with speech in noise, directional hearing affected, budget-constrained.
What the scoring engine returns: Three devices ranked by fit, with scores based on noise reduction rating, directionality features, and price. No device without a strong noise reduction spec makes the list regardless of anything else.
What the AI explains: Why each device fits the described situation, in plain language, using only the specifications in the dataset. It cannot say a device is “highly rated” or “popular” unless that data exists. It cannot introduce claims that aren’t in the source.
That last constraint matters more than it might seem. It means the explanations are auditable. If someone asks why a particular device was recommended, there’s a real answer rooted in real data.
The honesty built into the scoring
One thing I was deliberate about: the scoring engine never sees commission rates or affiliate relationships. A device without an affiliate program can score higher than one with a generous commission if the specs are better. The ranking is based on fit, not on what’s financially convenient for the tool.
This sounds obvious. It’s apparently not standard practice in comparison tools generally. I wanted it to be unambiguous here.
What this means for businesses putting AI in front of customers
The lesson I’d take from building this isn’t specific to hearing aids. It applies to any business that’s thinking about deploying AI in a customer-facing context.
Use AI where language and nuance matter. Understanding what a customer is actually asking, in their own words, is something AI does exceptionally well. Generating a clear, human explanation of a complex result is something AI does exceptionally well. Those are the right jobs for it.
Keep AI out of decisions that need to be consistent or defensible. If a customer asked why they got a particular recommendation and the honest answer is “the model thought so,” that’s a problem. Decisions with real consequences need to be traceable.
Always design the handoff path. For HearingShortlist, that’s the referral to an audiologist when symptoms suggest something beyond OTC. For a business chatbot, it’s the moment the bot says “let me connect you with someone.” That path should be intentional, not an afterthought.
The tool is live at HearingShortlist.com. It covers mild to moderate hearing loss, which is an honest limitation worth naming. More severe loss, or anything involving ear pain, sudden changes, or one-sided symptoms, goes straight to the referral path.
This is the same thinking I bring to client work at AInfuse. AI that’s deployed carelessly erodes trust. AI that’s deployed with clear boundaries, honest constraints, and a well-designed human handoff builds it.
