Eight point three basis points. That is the entire economic distance between the two numbers in Sela's funding announcement: one billion dollars of mortgage loans originated monthly, and ten million dollars in annual recurring revenue. Divide the second by the first, annualize the ratio, and a company that claims to touch a trillion dollars of loan flow every year is capturing less than one-tenth of one percent of it. The headline is a billion a month. The footnote is a ten-million-dollar business. Nobody put those two figures next to each other in the release, and that adjacency is the most interesting thing in it.
Sela raised $21 million — the announcement describes it as a combined seed and Series A — led by Costanoa Ventures with participation from Emergence Capital. Its product is a voice AI agent that speaks with mortgage borrowers, walks them through the loan process, and escalates to a human only when the conversation requires one. Seventeen employees. Six of the "top ten independent mortgage bankers" integrated into production workflows, per the company. A nine percent conversion lift. A forty percent improvement in profit per loan.
Trust is a protocol, not a promise — and the first question any protocol must answer is who counts, and what gets counted. I have spent sixteen years asking that question inside crypto, mostly about numbers nobody could audit. Reading this announcement, my instinct was not to doubt Sela. It was to notice that the mortgage industry has no block explorer, and that this absence is doing an enormous amount of work in the story.
The disclosure map, before the analysis
Start with what is stated. Sela's agents conduct borrower conversations end to end, handle complex interactions, and hand off to a licensed human when the situation exceeds their design envelope. Internal A/B testing covered more than ten thousand borrowers. A separate comparison against a competing voice AI vendor ran across roughly seven thousand leads. The results were framed as conversion and cost-per-loan improvements, which is the correct frame: the funding thesis, in the investors' own words, is that every lender's income statement reduces to those two variables.
Now list what is absent, because the shape of the silence is more informative than the shape of the claims. No base model. No disclosure of whether the model is proprietary, fine-tuned, or a rented API. No latency figures. No word error rate. No accent or dialect coverage. No pricing model. No gross margin. No churn or net revenue retention. No valuation. No legal characterization of the agent — whether it is a tool sold to a licensed lender or something closer to an unlicensed originator.
That last omission matters more than the others. Mortgage origination in the United States sits under the SAFE Act, state licensing through NMLS, ECOA and Regulation B, UDAAP, TRID disclosure rules, and the TCPA — and in February 2024 the FCC confirmed that AI-generated human voices fall under the TCPA's "artificial or prerecorded voice" category. An AI cannot hold a mortgage loan originator license. There is no mechanism for it. So whatever Sela's agent does, the legal personality must belong to a natural person at a licensed institution. We govern the gray areas between blocks — and the grayest area in agentic finance is not the model architecture. It is the moment a synthetic voice says a number that becomes a legal representation.
I spent 2025 as a governance architect on an Africa-focused Layer 2, negotiating real-world asset tokenization with institutions that arrive at every meeting carrying a compliance officer. The pattern is consistent: the technology is rarely the objection. The objection is always liability allocation. When I ran the governance token distribution for a 500-person artist collective in Lagos in 2021, the hardest work was not the smart contract. It was making the voting rights defensible to people who had been excluded from every prior structure. The mechanism was trivial. The legitimacy was not.
The arithmetic of the gap
Here is where a crypto reader has a specific, hard-won literacy that most fintech coverage lacks.
We have spent a decade arguing about what volume means. Total value locked that counts the same dollar six times as it recurses through leveraged loops. DEX volume inflated by wash trading between wallets under common control. Aggregators reporting routed flow as their own. Bridge TVL that counts a deposit on both sides of the crossing. Every one of these is a definitional choice, and every definitional choice is a commercial one. The reason we ended up building on-chain attestations, proof-of-reserves dashboards, and independent data providers is that we learned the hard way: when attribution is unauditable, the party being measured controls the denominator.
Sela's numbers live entirely inside that fog. "One billion a month" could mean one billion in loans that closed after a Sela agent touched the borrower, or one billion routed through a lender who happens to use Sela in a single product line, or something in between. The eight-point-three basis point ratio is not proof of inflation. It is entirely consistent with a seat-based or per-conversation pricing model, where revenue is decoupled from flow. But it does mean the trillion-dollar annualized figure was selected for narrative rather than for accounting — because if Sela captured even a quarter of the hundred to two hundred basis points that origination economics actually represent, its revenue would not be ten million dollars.
The absence of a shared ledger is not a neutral fact. It is a pricing decision. Attribution opacity is what allows a vendor to sell the story of volume and the reality of subscriptions in the same paragraph.
I have watched the same mechanism inside lending protocol governance. The base rate and the slope on the largest decentralized lending pools are set by governance vote — adjusted in discrete steps, anchored by a kink in the utilization curve that somebody chose. It resembles price discovery the way a thermostat resembles weather. That is not a criticism of parameterization; every system needs parameters. It is a warning about the stories we attach to them. A nine percent conversion lift, absent a disclosed baseline and a disclosed methodology, is a parameter too. Someone chose which cohort counted as the control.
What is actually being sold
Strip the marketing and the technical artifact becomes legible. A production voice agent for a regulated financial workflow requires streaming speech recognition, a conversational model, high-fidelity low-latency speech synthesis, barge-in handling so the borrower can interrupt, intent and sentiment classification, a state machine that tracks where a borrower sits in a multi-week process, and — decisively — deep API integration into a loan origination system, most likely Encompass. The orchestration layer is the product. The intelligence is rented.
The evidence for that conclusion is structural rather than stated. A seventeen-person team cannot simultaneously maintain model research, speech engineering, regulatory choreography, enterprise sales, and implementation. Through 2023 and 2024 the voice infrastructure layer commoditized aggressively; telephony, turn detection, speech synthesis, and low-latency inference are now purchasable by the minute. When those primitives are commodity and the team is small, the differentiator cannot be the primitive.
Silence in the chain speaks louder than noise. The absence of a single technical benchmark in an AI funding announcement is not an oversight. Benchmarks exist for models that have something to claim.
So where does the durability actually live? Not in the algorithm. In three accumulated assets: integration depth into a customer's origination system, which is expensive to build and expensive to abandon; the compliance choreography — a maintained script library, escalation rules, and the audit trail that lets a lender demonstrate it treated borrowers consistently; and whatever proprietary conversation data the company can lawfully reuse, which is probably less than outsiders assume, because the underlying data likely belongs to the lender and cannot cross customer boundaries without contractual scaffolding nobody has described.
This is the lesson I learned in Lagos in 2017, auditing a vesting schedule for a token issuance while colleagues chased announcement metrics. I found an integer overflow and refused to sign off until it was patched. I lost that job. Three similar projects were exploited weeks later, in one case because a rounding error compounded across a cliff. The value of a system never lives on the landing page. It lives in the parts nobody demos.
The cost side nobody discloses
Voice inference is not cheap. Streaming synthesis at conversational quality, continuous recognition, and a hosted language model running in parallel consume compute by the minute, not by the loan. Human escalation consumes labor by the incident. Every one of those line items attacks gross margin from below, and none of them were disclosed.
If Sela's agents resolve the full conversation on most leads, the unit economics can resemble software. If escalation is frequent, the company is a business process outsourcer with a better interface — and BPO margins do not support software multiples. Which of those two companies exists is the single most consequential unanswered question in the announcement, and it is answerable with one number: the escalation rate. That number is missing.
I have my own history with cost-side collapse. In 2022 my DAO's treasury lost sixty percent of its value and I withdrew from public writing for months. What I learned in that silence is that governance frameworks built during expansion are almost always revenue-side frameworks — they debate how to distribute what arrives. The frameworks that survive are cost-side, because costs are the part of a system that does not care about your narrative. I read every treasury proposal through that lens now, and I read this funding announcement the same way.
The contrarian test
The consensus reading of Sela is that it is the blueprint for vertical AI agents eating financial services from the outside: narrow domain, deep workflow integration, value-based pricing, high-friction incumbent. It is a clean story. Three things should hold it at arm's length.
Notice what the forty percent cannot distinguish. A voice agent that reaches borrowers within minutes instead of hours will naturally conduct more conversations with people who were already going to convert. A profit-per-loan improvement measured inside a lender's existing funnel cannot separate automation from prioritization unless the methodology controls for lead quality, and no methodology was published. To its credit, the announcement itself flagged that the data needs auditing. That flag deserves more weight than the numbers surrounding it.
The harder threat is architectural. The loan origination system is the choke point, and it is owned by a very small number of vendors. If the company that owns Encompass ships native voice handling, an external agent becomes a feature request with a sales team. We have watched this exact dynamic in our own stack: the application layer gets absorbed by whichever layer owns settlement. Aggregators fold into intent routing. Wallets fold into chains. Integration depth is a moat only until the thing you integrated with decides to compete.
Underneath both sits the structural limit. An AI cannot hold the license, so it cannot own the borrower relationship in any enforceable sense. Whoever owns the relationship owns the renewal. That caps Sela's ceiling at whatever lenders are willing to outsource of the front door — which for a regulated institution is a deliberate, limited, and revisitable decision.
A smaller note on the genre: twenty-one million dollars covering two rounds means neither round, individually, was large. The word "raises" performs the work a stage label would have performed.
The question that outlives the round
Here is what I would track, and it is not the conversion lift.
Mortgage pools are moving toward tokenization, and I have spent the past year inside that negotiation. If origination becomes programmable and the resulting note becomes an on-chain asset, the agent that conducted the conversation acquires a keypair and an attestation history. At that point the question stops being whether the AI performed well. It becomes whether the inference was verifiable — whether a signed receipt exists proving what the agent said, to whom, under which script version, at which model revision. Zero-knowledge machine learning is not ready for large language inference. Trusted execution environments are, imperfectly, today. Either way the demand will arrive from the compliance side first, because the compliance side is the side that gets sued.
Vision without verification is just hallucination. The billion-dollar figure is a vision. The eight-point-three basis points is what happens when nobody can check it. Which layer of this stack — the licensed institution, the origination system, or the agent itself — ends up capturing the hundred to two hundred basis points of origination economics once the conversation is machine-generated and the note is tokenized? That is the governance question of the next cycle, and it will be settled long before it is debated.