A critical assessment of the $21M raise behind AI agents that call and qualify mortgage borrowers — a genuinely hot vertical with strong claimed traction, sitting on the single most litigated fact pattern in fintech: an artificial voice contacting consumers about credit. The team is credible; the compliance story is thinner than the marketing.
The logos and growth would be excellent — if verifiable. But $1B/month, $10M ARR, “6 of 10 banks,” and both A/B lift studies are single-source and unaudited, with internal inconsistencies ($1B/mo vs. “$6B+ facilitated,” “tens of millions of calls” vs. “15M+ monthly”).
Maximal. Since the FCC’s 2024 ruling, AI-generated voices fall under the TCPA: written consent, $500–$1,500 per call, no cap. In Feb 2026, Lamb v. Mortgage One became a class action over exactly this — an AI voice cold-calling mortgage prospects.
Thin. Even Costanoa’s own thesis concedes voice generation is commoditizing. Sela discloses no LLM, TTS/STT, or telephony vendor, and is flanked by Kastle (servicing side, $24M) and compliance-first Sei — while horizontal players (Retell, Bland) undercut the infra.
Key Finding: Sela sits in a real, valuable vertical with a credible founding team. But its public compliance posture and its actual legal documents do not match: a product claiming 15M+ borrower calls a month ships a generic, GDPR-flavored privacy policy that names only Google Analytics and is silent on call recording, voice data, consent capture, GLBA, and subprocessors. The unaddressed structural question — where automated “qualifying” ends and licensable loan origination begins — is the exact line regulators and plaintiffs’ attorneys are now probing.
Sela’s agents don’t just answer FAQs. Its own marketing says they “qualify,” “educate borrowers,” and do “objection handling” before transferring to a licensed loan officer. Under the SAFE Act, some of that activity is licensable — and it is a human license, not a model’s.
An inbound or purchased lead (LendingTree, Own Up) enters the CRM. “Speed to lead under 1 second.”
An artificial voice places the call. Under the FCC’s 2024 ruling this triggers the full TCPA regime.
The agent gathers borrower information and screens fit — the point where “pre-qual” can shade into taking an application.
The agent explains products and rebuts hesitation — conduct that in some states edges toward negotiating loan terms.
Only now does a licensed human loan officer join — the line Sela draws, but never publicly defines.
The liability sits with the lender, but a vendor whose tool caused the violation is an obvious co-target. Sela sells the exact capability at the center of active litigation, to customers who bear the exposure — and its own disclosures show none of the consent machinery the law now requires.
Sela’s site advertises “TCPA Compliant Dialing” and “Audit-Ready Call Records.” Its actual privacy policy (effective April 1, 2024) is generic website boilerplate: it says the data collected is “email, name, cookies and usage data,” names only Google Analytics, and invokes EU GDPR — irrelevant to a US-only mortgage dialer. For a product claiming millions of recorded borrower calls, the policy is silent on call recording, voice/biometric data, consent capture, retention, GLBA, CCPA, and subprocessors. The marketing and the legal document describe two different companies.
Telephone Consumer Protection Act — since the FCC’s Feb 2024 ruling, AI voices need prior written consent; $500–$1,500/call, no cap.
State mortgage-loan-originator licensing — generally requires a licensed human to take an application or negotiate terms.
Feb 2026 class action (E.D. Mich.) over an AI voice cold-calling mortgage prospects without consent or disclosure — Sela’s exact fact pattern.
Two Sela-run A/B tests (9% lead-to-lock lift; 41% vs. another AI) with no methodology, significance, or third-party audit.
No foundation model, voice API, or telephony vendor named — margin, latency, and quality all ride third-party rails Sela doesn’t price-control.
Revenue leans on a few large lenders in a rate-sensitive industry; a rate spike or one churn dents a $10M-ARR base.
Every headline figure that makes Sela impressive originates with Sela. That is not disqualifying — early companies rarely have audited metrics — but a buyer should treat the traction as a claim to verify, not a fact, and note that the moat sits in a commoditizing layer.
Larger than compliance-first Sei, comparable to Kastle’s fresh $24M (which attacks the servicing/collections half Sela doesn’t touch), and far below the horizontal voice-infra players (Retell, Bland, PolyAI) whose APIs Sela most likely depends on.
A vertical voice company almost certainly rides third-party foundation models and voice APIs — yet none is named. That makes gross margin, the “<1 second” latency, and the “outperforms humans” behavior all hostage to suppliers Sela neither discloses nor controls on price.
Kastle builds the deeper servicing/collections stack; Sei layers 100% call QA and FDCPA/TRID monitoring Sela lacks by design. Sela’s defensibility rests on mortgage workflow and data — real, but narrow, in a category the infra players commoditize from below.
What would change the verdict: an audited ARR or origination figure, a disclosed consent-capture and call-recording architecture, and a clear public statement of where its agents stop and licensed humans begin. Until then, the strongest, most verifiable facts about Sela are the gaps in its own disclosures.
Seven structural risks the $21M raise does not resolve.
AI voice contacting consumers about credit is the precise fact pattern in Lamb v. Mortgage One (class action, Feb 2026). $500–$1,500 per call, no cap; the lender holds the bag, but the vendor whose tool caused it is an obvious co-defendant.
“Qualify / educate / handle objections” before human transfer may cross into licensable origination activity in some states — a line requiring a licensed human. Sela has never publicly drawn it.
A boilerplate GDPR-flavored policy omitting call recording, voice data, GLBA, CCPA, and subprocessors — a concrete, verifiable weakness that directly contradicts the site’s “audit-ready / TCPA compliant” badges.
No LLM/TTS/STT/telephony vendor named; margin, latency, and quality sit on rails Sela doesn’t control. A model or pricing change could reset unit economics or degrade the “outperforms humans” edge overnight.
“$1B/month,” “$10M ARR,” “6 of 10 banks,” and all A/B lift figures are single-source, with internal inconsistencies between cumulative and monthly numbers.
Costanoa’s own thesis concedes voice generation is commoditizing. Sei and Kastle build deeper compliance/servicing; horizontal players undercut on infra.
Revenue leans on a handful of very large lenders in a rate-sensitive industry; a rate spike or a single enterprise churn materially dents a $10M-ARR base, and reliance on aggregator lead buyers adds channel risk.
Sela is a credible team attacking a real, valuable vertical — wrapped in a compliance story its own documents don’t support. The claimed traction, if audited, would justify the round; the regulatory exposure, if left unaddressed, could unwind it. The diligence question is not whether AI can sell mortgages — it is whether a 17-person startup selling artificial-voice calls about consumer credit has built the consent, licensing, and audit machinery the law now demands. On the public evidence, it hasn’t shown it.
Based entirely on publicly available information, anchored to the September 22, 2026 funding announcement. Every performance figure is company-sourced and unaudited; the privacy-policy, vendor-disclosure, licensing, and TCPA findings are the most solid, evidence-backed parts of this analysis.
