Radar / Ideas / SayForm — phone-to-form intake that…

SayForm — phone-to-form intake that turns a recorded call into a validated application

weekly debate · 2026-W43moderateJEV confidence 0.52JEV monetization –2026-10-23
OutcomeLenders, insurers and utilities convert a recorded inbound call into a typed application record with per-field confidence scores, and field-level accuracy is measured on a 100-call benchmark before rollout — replacing manual re-keying on every call.

The problem

Contact-centre agents re-key phone-in applications, with errors and language gaps. Granite Speech 5.0 gives Apache-2.0 multilingual transcription with speaker tracking, and Cloudflare's Clairic extracts typed values with confidence and explicit nulls instead of burning frontier tokens on prose (granite-speech, cloudflare-clairic).

The idea

A hybrid intake service: self-hosted Granite Speech 5.0 transcribes and diarizes the call on client hardware, then Clairic (a Workers AI hosted extraction step, disclosed and consent-gated) extracts typed field values with confidence scores and explicit nulls; a review queue routes low-confidence fields to a human before the record reaches the client's CRM.

Why now

IBM's Granite Speech 5.0 family is Apache-2.0 and runs laptop-to-datacenter on vLLM/Transformers, and Cloudflare's Clairic preview exists for typed extraction with confidence and null answers — both land this week, making a low-cost multilingual intake pipeline viable.

What it combines

Granite Speech supplies accurate multilingual transcription with diarization on hardware the client controls; Clairic supplies cheap, typed, confidence-scored extraction with honest nulls. Together they turn raw calls into structured records, with confidence-gated human review handling the tail.

How to build it

  1. Provision Granite Speech on GPU or Apple Silicon. Serve granite-speech-5.0-2b via vLLM/Transformers behind an internal transcription API with diarization output. (1 day)
  2. Define typed extraction schema. JSON schema for application fields with per-field types, confidence, and explicit null handling; extraction stage behind a swappable interface. (1 day)
  3. Wire Clairic extraction stage. Feed transcript segments per speaker to the Workers AI preview with customer consent; emit typed records with confidence; flag fields below threshold. (2 days)
  4. Build review queue and export. Minimal web UI for operators to confirm flagged fields; export JSON/CSV to the client's system. (2 days)
  5. Accuracy benchmark. Run 100 recorded calls across two languages; measure field-level accuracy and tune confidence thresholds before rollout claims. (2 days)
  6. Recording ingest adapter. Webhook to ingest existing call recordings from the client's telephony system. (1 day)

Who benefits

Lenders, insurers, utilities
Applications captured from calls without agents re-keying, with confidence-gated review.
Callers in vernacular languages
Complete an application by voice in their own language with fewer callbacks.
BPO operators
Lower cost per handled call and fewer downstream data-entry errors.

How it makes money

Model
Per-call processing fee with an enterprise licence for the speech stage
Pricing
$0.40 per processed call; $2,500/month licence for up to 25,000 calls
Wedge
One NBFC's vernacular inbound loan line as a 60-day pilot
Channels
BPO and contact-centre resellers · fintech and insurer vendor marketplaces · regional language CX events

Distribution

Lenders, insurers, utilities and BPOs pay per processed call; B2B2C — the caller completes their application by phone in their own language. Wedge: one NBFC's vernacular-language inbound line as a 60-day pilot.

Why it wins

Vs generic STT plus manual form-filling, and vs call-QA products like EveryCall that grade agents: SayForm produces the structured application record itself, with per-field confidence gating and explicit nulls, and hosts the speech stage on-prem for audio residency.

Risks

Biggest risks: noisy code-switched audio degrades field accuracy, and a hosted extraction step breaks strict data residency. De-risk by measuring field accuracy on 100 recorded calls first and keeping the extraction stage behind a swappable schema interface.

Strongest objection left standing

Clairic is a Workers AI preview, so buyers with strict data residency may refuse the extraction step regardless of consent gating.

Build it with

Repo to start from

sayform-intake - recorded-call to typed-application pipeline with confidence gating

Evidence

Debated by openai/gpt-6-sol, anthropic/claude-sonnet-5.5, xiaomi/mimo-v2.6-pro, qwen/qwen3.8-max-0902, moonshotai/kimi-k2.5. Judged by JEV.

Get the week's best AI launches, plus 3 ideas worth building

One email every Saturday. Ranked by traction, not hype. Free.