What changed
Apple shipped a first-party on-device streaming speech-recognition API (SpeechAnalyzer) that makes real-time transcription free, private, and local (FACT, per the benchmark source). In parallel, FTC data show imposter-scam losses hit $3.5B in 2025, ~3x since 2020 (FACT, per FTC press release).
Why now
On-device STT removes the per-minute cloud cost, latency, and privacy toxicity that made 'analyze phone audio' impractical. The scam-loss curve gives the pain a headline number.
Converging signals
A new abundant capability (on-device real-time STT) x a large, documented, growing pain (imposter scams). Two signals, one bridge (ai x complaint).
Customer pain
Real and well-evidenced: adult children fear their aging parents will be scammed by fake fraud-alert / 'verify your account' / gift-card-coercion calls; FTC quantifies $3.5B/yr in losses. This is HYPOTHESIS as to willingness-to-pay for THIS product, FACT as to the underlying loss.
Who pays
Hypothesis: adult children buying elder-protection subscriptions; possibly banks/credit unions white-labeling to cut fraud losses. No demand_evidence provided for either buyer paying for a live-screener specifically.
Solved today
Carrier spam-labeling (AT&T Call Protect, Verizon Call Filter), Truecaller/Hiya caller-ID databases, bank fraud-text alerts, and post-hoc education. These label the NUMBER, not the CONTENT of the conversation.
Why current solutions are bad
Number-based blocking misses spoofed or novel numbers and does nothing once the elder is already talking to a convincing 'bank fraud department.' Content-level, mid-call warning would be genuinely differentiated — IF it were buildable.
Proposed product
iOS app that runs SpeechAnalyzer on live in-call audio, matches FTC-taxonomy impersonation scripts on-device with a lightweight classifier, and pushes a local warning to the elder plus an opted-in family contact.
MVP version
NOT BUILDABLE AS SPECIFIED. iOS does not expose live cellular/VoIP call audio to third-party apps — CallKit provides call state/UI only, not an audio tap, and the OS blocks recording of active phone calls. SpeechAnalyzer can transcribe the mic or imported recordings, not the far-end voice of a live carrier call. The founder's own KILL TEST ('verify the API can access live in-call audio') fails on the platform, not the API.
30-day build
Only honest 30-day move: prototype against the ALLOWED surface — speakerphone mic capture (ask the elder to put scam-suspect calls on speaker) or post-call voicemail/recording analysis — and measure whether that degraded UX has any buyer pull. Do NOT market a 'mid-call live tap' you cannot ship.
60-day build
If the speakerphone/voicemail wedge shows pull, build a scam-script classifier on the FTC taxonomy and validate detection precision on a labeled corpus (still-needed resource, per input).
90-day revenue plan
Consumer App Store subscription for the degraded version, or pivot the classifier to a channel that DOES have text: SMS/voicemail-transcript scam scoring, where you legally get the content. Live-call revenue: none, blocked.
Distribution path
Consumer iOS App Store — ad-spend-heavy, network-effect-adjacent, the founder's explicitly-avoided quadrant. Bank white-label is enterprise procurement (also avoided).
Pricing hypothesis
Hypothetical $3-6/mo consumer subscription; unproven, no demand_evidence.
Technical difficulty
The classifier is moderate; the blocker is an absolute platform entitlement wall on live call audio, which no amount of engineering removes.
Legal / regulatory risk
Two-party call-recording consent laws in many states; App Store rejects apps that record calls. Compounds the technical block.
Platform dependency
Total. Apple controls both the API and the entitlement that makes the core feature impossible. If Apple/carriers bake screening into the OS/network (input's own TIME note), the wedge closes anyway.
Founder fit
Low. This is a consumer app with network-effect distribution, ad-spend dependence, and platform-gated core tech — the opposite of his government-portal / forced-filer / complaint-mining sweet spot. No public money, no forced buyer, no per-filing transaction.
Breakout potential
Would be high IF the live tap existed — content-level mid-call warning is a real gap. It does not exist, so realized potential is low.
Final recommendation
KILL as specified. The single most important claim — real-time access to live call audio on iOS — is false, and it is the load-bearing assumption. Do not build the live screener. If the scam-loss pain is worth chasing, pivot to a channel where the content is legally accessible on-device (SMS/voicemail-transcript scam scoring), and even then only after finding demand evidence of a paying buyer.
Next action
Spend one hour confirming the platform block in Apple's docs (CallKit + AVAudioSession restrictions on active calls, App Store Review Guideline on call recording); document the failed kill test and either pivot to voicemail/SMS scam-scoring or drop it.