Terrakotta

Remove the model layer and what remains is a commodity dialer.

Voice cloning is not a feature attached to this product, it is the reason the product exists. A studio inside the platform builds a voice model from a representative's own recordings, and every subsequent voicemail is synthesised in that voice, written for the individual prospect and split tested against alternatives on callback rate. The company describes itself as adding voice as a channel, and its accelerator listing states the proposition directly.

The data half is equally model driven. Skip tracing is delivered as a prompt rather than a form, described by the vendor as being as easy as asking a language model, and a conversational research tool answers property questions against the database. Call summaries are generated after each conversation.

What remains deterministic is the telephony: the dialer, disconnected line detection, number verification and the live connection when someone answers. Those are necessary and none of them is distinctive.

The founders' own launch material asks readers for feedback on how accurate they find the voice cloning, which places the model at the centre of what they think they are selling.

Ask what the voice model is built from and how many samples it requires.

The division of labour is clean: software handles dialling and the no answer path, a person takes every real conversation.

That matters more than it sounds. The vendor states that when a prospect answers, the representative is connected immediately in live conversation with no lag, as if the call came from their own handset. So the synthetic voice never holds a dialogue, never fields a question and never represents the company to someone who is actually listening. It operates only where nobody picked up. Several records built this week let generated voices conduct entire conversations; this one deliberately does not.

Before dialling, two filters operate: automatic detection of disconnected lines, and bulk verification and grading of numbers, so effort is not spent on dead records.

Split testing runs across voicemail variants, and calls are recorded and summarised.

What is missing is control over the generated artefact. Two passes located no approval step before a voicemail is synthesised and left, no review queue, no audit trail and no stated limit on daily volume. A representative approves the voice once at cloning and thereafter messages go out in it.

Ask whether individual voicemail scripts can be reviewed before they are left.

The most sensitive question this product raises is answered directly in the questions section, which is why this sits above the usual grade.

Asked whether voice clone information is stored, the vendor answers yes, states that voice data is held on encrypted servers, that a user can delete it at any time through account settings, and that it is never shared with third parties without explicit consent. Voice data is biometric material, and user controlled deletion is the part almost every vendor omits. Publishing that answer where a prospective buyer will read it, rather than in a policy document, is a deliberate choice.

Against that, nothing else is disclosed. Two passes located no provider, no model name, no version and no accuracy figure for the voice cloning, and nothing at all about what generates the personalised voicemail text, the skip tracing inference, the property research tool or the call summaries.

The accuracy question is not academic here. The founders' own launch material asks readers how accurate they find the cloning, which suggests fidelity is the thing being iterated on, and no measure of it is published.

Ask how faithful the voice clone is and what generates the voicemail script.

Accelerator backing and published job listings are the checkable evidence; the customer proof is thin.

The verifiable material is unusually concrete for a company this small. A named accelerator batch in early 2024, three founders with traceable histories covering two named universities, a data platform, a consultancy and a ride hailing company, a stated team size of five, and open roles published across three cities with salary bands and equity ranges attached. A prospective buyer can establish that this is a real operating company in a few minutes.

On customer outcomes the record is much lighter. Four testimonials appear, one carrying a specific figure of three hundred calls in a single day and a statement that the user would not be where they are without the software. None of those retrieved carried a name or a company. A customers page exists and was not opened.

Coverage in a commercial property sector newsletter describes the product accurately.

What two passes could not locate: any named customer, any case study, any review count or rating on an independent platform, and any measured callback rate improvement, which is the one metric this product exists to move.

Ask what callback rate lift the split tested voicemails actually produce.

No do not call screening appears anywhere, on a power dialer calling numbers obtained by skip tracing.

Two passes across the pricing page, the tier feature lists, the questions section, the home page and independent coverage located no reference to do not call registers, no screening, no scrubbing and no suppression capability of any kind. Every tier lists unlimited calls and voicemails.

That combination is the concern. The platform locates telephone numbers for people through public record searches, then dials them at volume, and calling a registered number in the United States is unlawful rather than merely unwelcome. A vendor built earlier in this same sweep includes register screening on its cheapest tier at sixty dollars, so this is not a question of company size.

Number verification and grading do exist, and they answer a different question: whether a line is live and reachable, not whether it may lawfully be called.

Also absent: any calling hours restriction, any consent basis for skip traced numbers, and any notification to a recipient that a call is being recorded, in a product sold across states with differing consent rules.

One favourable commitment: uploaded contact data is never shared with any third party.

Ask what do not call screening exists before a number is dialled.

Two commitments published in plain language, and both answer the question a buyer would actually ask.

On voice: data is stored on encrypted servers, the user can delete it at any time through account settings, and it is never shared with third parties without explicit consent. Voice is biometric material subject to specific statutes in several United States jurisdictions, and self service deletion is the control most vendors holding such data do not offer.

On contact data: a stated zero sharing policy, with the vendor undertaking never to share, use or distribute uploaded contact data to a third party under any circumstances. The word used is doing real work in that sentence, because it forecloses training on or monetising customer lists as well as selling them, and very few vendors in this index make the second commitment at all.

What two passes could not locate: a processing agreement, a subprocessor list, a retention schedule for call recordings and transcripts, a deletion commitment for anything other than voice, or a residency statement.

The holdings extend beyond what those two commitments cover. Call recordings, transcripts, generated summaries and skip traced records about individuals all accumulate.

Ask how long call recordings and skip traced records are retained.

The method is described in general terms, which is more than most manage, and no source is named.

Asked how skip tracing works, the vendor states that the system pulls from a variety of publicly available resources and cross references multiple data sources to validate matches. That describes both the character of the inputs and a validation approach, and it is a fuller answer than the silence recorded on most data records in this index.

It still names nothing. Two passes located no supplier, no record count for the property database, no refresh cadence, no accuracy figure and no consent basis.

The activity deserves stating plainly for what it is. Skip tracing is the practice of locating a person's current contact details by assembling public and commercial records. The people located have not asked to be found, have no relationship with the broker who will call them, and in most cases will not know how their number was obtained. That the underlying records are public does not make the assembled result something the individual chose.

The property database adds ownership, tenant and market context to the same records.

Ask which public sources the skip trace draws on and how matches are validated.

Telephony and integrations both run through sanctioned paths, with one surface worth noting.

Calls are placed through the vendor's own infrastructure, and customers may bring an existing number and register it in the platform, so nothing depends on operating someone else's telephony without permission. Record system connections are named across a major platform and three of its commercial property overlays, plus four further systems, delivered through an interface and prebuilt connectors, with custom work available at the top tier. Those are partner arrangements rather than workarounds.

Two passes located no social network automation, no account renting, no scraping claim and no evasion language anywhere.

The surface worth noting is the browser extension, which displays owner names and telephone numbers on any web page. That overlays the vendor's own data onto third party sites rather than extracting from them, which is a materially different act from the extension based harvesting recorded elsewhere in this sweep, and it does mean the tool operates visually inside properties the vendor does not own.

A messaging application assistant delivers research through a consumer messaging channel.

Ask how the browser extension obtains the data it displays on third party pages.

The stewardship commitments are good and the one safety control this product most needs is undescribed.

Stewardship first, because it is genuinely above the norm. Voice data on encrypted servers, deletable by the user at will, never shared without explicit consent. Contact data never shared, used or distributed to any third party. Both stated plainly rather than buried.

The missing control is verification of voice ownership. The product clones a voice from submitted recordings, and two passes located nothing describing how the platform establishes that the voice belongs to the person submitting it. No consent capture from the voice owner, no identity check, no restriction preventing a user uploading samples of somebody else's speech. Every safeguard published concerns protecting the voice model after it exists; none concerns whether it should have been created.

For a technology whose misuse case is impersonation, that is the control a buyer should ask about first, and it is also the one a brokerage should want before letting staff clone voices.

Also absent: any model disclosure, any fidelity measure, any incident process, any named security contact and any vulnerability disclosure route.

Ask what verifies that a submitted voice belongs to the person submitting it.

A synthetic copy of a named person's voice says words that person never said, to strangers, at scale. This is the sharpest instance of the problem recorded anywhere in this index.

The prospect plays a voicemail. It is a specific broker, addressing them by name, referring to their property. The voice is unmistakably that individual's, because it was cloned from their own recordings. The words were generated, personalised for this recipient, and split tested against alternatives to find the version producing the most callbacks. The broker never spoke them.

Every verifiable signal is authentic. The number is real, the broker exists, the voice matches. The artefact is manufactured, and two passes located no disclosure to the recipient anywhere: no statement in the message, no notification, no configuration for adding one.

Compare the two account renting records built this week, where a real person's professional identity was used by someone else. This goes a step further, using a real person's voice to utter sentences they never uttered, and voice carries a stronger presumption of authenticity than a name on a profile.

Two genuine mitigations, neither of which is disclosure. The broker consented to their own cloning. And if the prospect calls back or answers, they reach that actual person live.

Ask whether the voicemail can state that it was generated.

The record system coverage is chosen for one industry and it is the right list.

Alongside two general platforms sit three commercial property overlays built on top of a major record system, plus a property sector native system and a spreadsheet import path. A brokerage does not run a generic configuration, it runs one of those overlays, and supporting them specifically rather than offering a generic connector is the difference between working and not working on day one.

An interface is published with prebuilt connectors, and custom integrations are available at the top tier. Bulk number verification runs through that interface at up to a thousand records per request with sub two second per record latency, which is a concrete performance commitment.

Two further surfaces are unusual. A browser extension surfaces owner names and numbers on any page. And an assistant delivers comparables and owner research inside a consumer phone messaging application, which puts research where a broker already is rather than requiring them to open a platform.

What two passes could not locate: published interface documentation, an authentication scheme, webhooks, a protocol server or a status page.

Ask whether the interface is documented publicly and whether webhooks exist.

Two concrete technical statements and nothing about where anything runs.

What is stated: voice data is held on encrypted servers, and the verification interface processes up to a thousand numbers per request at under two seconds each. The second is a real performance commitment rather than a claim, and a buyer can test it.

Setup is described as taking five minutes, and customers may bring their own telephone number and register it in the platform, so the telephony layer does not require provisioning.

What two passes could not locate: a hosting provider, a region, a data centre, a residency commitment, a tenancy or isolation model, a backup position, a continuity plan, an uptime commitment or a status page.

The holdings sharpen the residency question beyond the usual. Voice biometric templates for every user attract specific statutory treatment in several United States jurisdictions, with requirements covering storage, retention and destruction that depend on where the data sits. Call recordings and skip traced records about individuals accumulate alongside.

A buyer in a state with biometric privacy legislation would need the processing location before enrolling staff voices.

Ask where voice models and call recordings are stored.

Two legal documents, and custody of material that warrants considerably more.

The footer carries terms of service and a privacy policy. Two passes located no certification, no audit report, no examination, no penetration test, no trust portal, no dedicated security page, no access control description, no incident response process, no breach notification commitment, no named security contact, no vulnerability disclosure route, no subprocessor list and no status page.

Encryption is asserted once, in a questions answer about voice storage, without detail.

What sits behind that is the concern. This platform holds voice biometric templates capable of reproducing a named individual's speech convincingly, alongside recordings of their calls, transcripts, and contact records for people located through public record searches. A compromise of the voice models specifically would hand an attacker the ability to impersonate identifiable professionals in their own voices, which is a category of harm distinct from a data leak.

The vendor's own commitments show awareness of the sensitivity, which makes the absence of any security material to support them the gap.

Ask what protects stored voice models and where a researcher reports a vulnerability.

Three tiers are published, none carries a price, and all three list identical features.

The structure is genuinely unusual. A solo plan at one user, a team plan at two or more, an enterprise plan at ten or more. Each lists unlimited calls and voicemails, call recording with generated summaries, skip tracer access and telephone number verification. Those four items appear verbatim on all three.

The only stated difference anywhere in the table is support response time: four hours, two hours and one hour respectively.

So the tier comparison tells a buyer nothing about what they get and nothing about what it costs. Two tiers offer a subscribe action with no figure attached and the third routes to sales. Third party coverage confirms pricing is custom and varies by team size.

One consequence is worth drawing out. Because capability does not vary by tier, the sole thing being sold at higher prices is faster support, which is either an admirably simple model or a table that exists to look like a pricing page. A buyer cannot tell which without a call.

A free trial is available and custom integrations are stated as typically requiring the top tier, which is the one capability difference mentioned anywhere.

Ask what each tier costs and what else differs beyond support time.

One deletion right is explicit and user controlled, which is rare, and everything else is undescribed.

The explicit part: voice data can be deleted at any time through account settings, by the user, without contacting support. For biometric material that is the right design, and it is the only self service deletion control recorded in this index.

The zero sharing commitment on contact data works in the customer's favour on exit too, since uploaded lists are not retained for other purposes or passed onward.

An interface with prebuilt connectors and synchronisation to named record systems means call activity can flow into a system the customer owns as the work happens.

What two passes could not locate: any export mechanism for call recordings, transcripts, generated summaries or skip traced records, any format, any retention position after cancellation, any deletion commitment covering anything other than voice, and any account closure process.

Call recordings are the material with the clearest ongoing value and the least described route out.

Ask what can be exported besides voice deletion, and what happens to recordings on cancellation.

Number quality is handled well and caller reputation is not addressed at all.

What exists is real. Bulk verification and grading run through an interface at up to a thousand numbers per request in under two seconds each, and the dialer automatically detects disconnected lines and filters them. Together those stop effort being spent on dead records, which is the first discipline any high volume calling operation needs, and grading implies a quality signal beyond a simple live or dead answer.

Customers may bring their own number rather than dialling from a pool.

What is absent is the other half of this axis. Two passes located no caller identity reputation monitoring, no spam flag detection, no local presence matching, no number rotation and no per number volume ceiling. A vendor built earlier in this same sweep places live test calls to handsets on three carrier networks to observe how its numbers actually display, and replaces flagged ones daily.

That gap matters at the volumes this product enables. A testimonial cites three hundred calls in a single day from one user, and sustained volume from a single bring your own number is how a number gets flagged.

Ask what monitors whether the calling number is being flagged as spam.

One vertical, served properly, with a broader claim nothing supports.

The commercial property focus is deep and coherent. Named audiences cover brokers, acquisitions teams and commercial services professionals. Expired listings has its own page as a use case. The record system list includes three overlays built specifically for the property sector plus a sector native platform. The database covers ownership, tenants, comparables and market context, and the research assistant answers property questions. Every surface examined addresses that industry.

That depth is a genuine strength rather than a limitation, and it is why the integrations work where a generic dialer would not.

The tension is with the company's own broader positioning, which describes an artificial intelligence platform for phone first sellers generally. Two passes located no material addressing any non property vertical, no case study outside it and no feature suggesting adaptation to another industry. The voice cloning and dialer would transfer; the database, research tool and integrations would not.

What is absent: any geography, any company size guidance beyond user count thresholds, and any statement of which property markets the database covers.

Ask which geographic markets the property database covers.