Coldreach

The removal test leaves nothing. Every load bearing part of this product is the model doing something a filter cannot. Signals are described in plain English and matched against more than 113 million companies, and the vendor makes the architectural point explicitly, that there are no preset categories or rigid filter boxes, which means the matching is comprehension rather than field comparison.

The research step reads a company's website, its regulatory filing, its hiring pages, recent news and social activity and extracts a pain point, a status quo and an urgency, which is interpretation rather than retrieval. The message is then generated from that finding. Strip the models out and what remains is a subscription to some public data feeds with nothing able to read them.

The vendor's own definition of the category confirms it, describing an AI sales development representative as software that can independently decide who to reach out to and what to say without a human writing each email.

The vendor defines the product category by the absence of a human in the loop, describing an AI sales development representative as software that independently decides who to reach out to and what to say without a person writing each email, and it markets an auto pilot mode in which the agent works around the clock reaching out automatically.

What the operator controls is upstream configuration rather than downstream review: custom prompts defining what counts as a buying signal, and custom instructions shaping tone. Both of those govern what the system looks for and how it sounds, not whether any individual message should go. Nothing published describes an approval step, a review queue, a confidence threshold on the signal match, a guardrail that withholds a message, an escalation path or an audit trail of what the agent did. One artefact does reach the operator, a weekly export of the targeted account list, but a list delivered on a schedule is visibility after the fact rather than a gate before the send.

Bottom of the band, earned on mechanism rather than on model identity. The published description of how the system works is unusually specific for this category: it names the document classes read for each lead, covering the company website, the regulatory filing, the hiring page, recent news and social activity, and it names what is extracted from them, namely the pain point, the status quo and the urgency.

It also explains the matching architecture, that a plain English description is evaluated against five or more data sources rather than resolved into filters. That is real mechanism disclosure. What is entirely absent is the model: no provider, no family, no version, no inference location, no retention position for the material read or the content generated. And one gap matters more than any of those, because it goes to the product's central claim.

The marketing states that every message opens with a specific, verified reason for the outreach, and no accuracy or error rate is published for the signal detection, so verified is asserted without a standard behind it. A wrong signal produces a confident email about something that did not happen.

Above the corpus norm on specificity and undercut by one comparison. The good part is genuinely good: two testimonials carry a full name, a job title and a named employer, a customer outcome is published with real figures covering 1,768 leads reached, 103 replies and 34 interested leads in a first month, and the vendor publishes its own internal numbers with a denominator, stating that its own agent books around 70 percent of its inbound pipeline and produces ten to fifteen qualified meetings a week from 700 to 800 outreach attempts.

Publishing the attempts alongside the meetings is rare and creditable. A reply rate of 3.8 percent is stated across more than 500,000 messages, which is a real sample size. The problem is the framing around that figure. It is presented as roughly ten times an industry baseline of 0.3 to 0.5 percent, and that baseline is unsourced and sits far below widely published benchmarks for the same channel, several of which put the average above three percent. The multiple therefore depends almost entirely on which baseline was chosen. No analyst evaluation or independent study exists, and the customer logo strip is unnamed.

The regulatory surface is unaddressed on every page read. This product sends unsolicited email and professional network messages at scale to people who have not consented, and separately profiles individuals by reading their social activity and job changes, and no regulation is named anywhere: no commercial email rules, no European direct marketing position, no Canadian consent regime, no unsubscribe or suppression mechanism, and no statement of where responsibility for the lawfulness of the outreach sits between the vendor and its customer.

Only two legal documents are published on the site, terms of service and a privacy policy, and neither was read this pass. Recorded as observed rather than concluded: this is a failure to locate a position rather than a finding that none exists, and either document could carry one that the marketing does not. Re verify there before this row is quoted.

The published legal surface is two documents, terms of service and a privacy policy, neither read this pass, with no processing addendum, no jurisdiction specific page, no data protection contact, no European representative and no sub processor list located.

What raises the stakes is the product's own example signals, which the vendor publishes to demonstrate capability and which describe individual level profiling in plain terms: one worked example is a company that onboarded a data engineer in the last three months who mentioned a particular technology on their professional profile. That is a named living person being read, categorised and used as a targeting trigger, at a scale of more than 113 million monitored companies.

No notification route, self service lookup, objection path or removal mechanism exists for those individuals anywhere on the site. Recorded as observed rather than concluded. Re verify by reading the published privacy policy directly.

The source types are named with real specificity and not one supplier is. The vendor lists what it reads across five or more data sources, covering job postings, company news, website content, social posts, employee changes and securities filings, which is more informative than the single word disclosures common in this category. The useful distinction to draw is that those sources are not alike.

Regulatory filings, job postings and a company's own website are public record, and reading them at scale is defensible and needs little justification. Social profile activity and employee changes are not the same thing, and the published example of detecting that a named individual joined a company and mentioned a technology on their profile describes systematic collection from a platform, with no source, method, licence or permitted use position stated anywhere. Contact data is the third gap: deep lead enrichment is listed as included on every tier and nothing describes where the email addresses come from, nor is any notice given to the people concerned.

Bottom of the band on a single clause the vendor wrote itself. Answering its own question about outreach channels, the site states that professional network outreach is paced like a human to avoid account flags. That is evasion framing rather than a conformance position: the pacing exists to defeat the platform's detection of automation, and the vendor says so.

Alongside it, the research layer reads social posts and profile changes at a scale of more than 113 million monitored companies, which is systematic collection from the same platform with no stated method or basis, and the platform's own sales product is listed as an integration.

Held above the lowest grade, and the distance is real: there is no account rotation, no multiple account support, no proxy infrastructure, no credential custody described and no marketing programme built around undetectability, which is what has taken three other vendors in this index below this band. The rest of the integration surface is entirely official, covering established customer relationship, engagement and revenue intelligence platforms.

The training and tenancy questions are unaddressed on every surface read. Nothing published states whether customer data trains or improves models, whether any learning is scoped to a single tenant, how long prompts, research output and generated messages are retained, or which sub processors see the material.

The question has weight here because of what flows through the system: customer records synced from a system of record, replies from prospects, and the research corpus assembled on more than 113 million companies including individual profile activity. Two legal documents are published, terms of service and a privacy policy, and neither was read this pass, so this is a failure to locate a position rather than a finding that none exists. Recorded as observed. Re verify in both documents, which are the conventional home for a processing and retention statement.

The clearest case of manufactured effort in the index, and it is the strongest version of that shape because the effort is not simulated at all. Elsewhere on this axis a generated icebreaker implies research that never happened. Here the research genuinely happened: the system read the company's regulatory filing, its hiring page, its news and its social activity, found something specific and true, and opened the email with it. What did not happen is a person doing any of it.

The recipient receives a message demonstrating exactly the individual attention that would once have taken a representative an hour, produced in seconds against a monitored universe of more than 113 million companies, and there is no signal available to them that would distinguish the two. The research is real and the researcher is not.

Nothing here impersonates a named human, uses a synthetic voice or invents an identity, and the sender is the customer's own representative, which is why this sits at the middle of the band. Article 50 of the EU AI Act is not addressed, and the vendor's own definition of its product, a system that independently decides what to say without a person writing it, places it squarely in scope.

Bottom of the band on a small but well chosen and specifically named set. The integrations cover the three categories a research and outreach layer actually needs to sit inside an existing stack: systems of record with two major platforms named, engagement platforms with three named, revenue intelligence with one, plus team notification and the professional network's own sales product. A programmatic interface is listed alongside them for custom work.

For a product that positions itself as a layer rather than a replacement, that set is the right shape rather than a long count. Two things hold it here. Synchronisation with a system of record is gated above the entry plan and included only from the second tier upward, which means the cheapest customers generate researched leads with no documented path into the place their pipeline lives. And no connector count, marketplace, partner directory, public developer documentation or agent facing endpoint was located on any page read.

The residency question is unaddressed on every surface read this pass. No hosting region is named, no residency option is offered, no cross border transfer position is stated and no sub processor list is published. The gap is larger than it looks for a product of this shape, because the platform holds three distinct bodies of data whose location a buyer would want to know: the research corpus assembled on more than 113 million companies, the customer's own records synchronised from their system of record, and the prospect replies flowing back through the outreach layer.

Only terms of service and a privacy policy are published and neither was read this pass. Recorded as observed rather than concluded: this is a failure to locate documentation rather than a finding that none exists. Re verify in both documents.

No certification, trust centre, security page, audit report, penetration test, vulnerability disclosure policy or enumerated control set was located on any page read, and the footer carries exactly two links, terms of service and a privacy policy. Recorded as observed rather than concluded, and the reading may well be a false negative given the named enterprise integrations, since synchronising with a major system of record usually forces a security review somewhere.

One observation belongs on the record because the vendor supplied it itself. Among the published example buying signals demonstrating what the product can detect about other companies is whether that company holds a particular security attestation. The platform therefore sells the ability to identify which prospects are certified while publishing nothing at all about its own posture.

Five real figures exist, which puts this above the vendors here that publish none, and almost everything around them works against a buyer trying to budget. The numbers are 899, 1,099, 1,499 and 1,999 dollars a month for self service tiers and 3,000 for a fully managed service. Four problems follow. There is no pricing page: the navigation offers use cases, integrations, a blog, sign in and a demo booking, and the figures appear only inside a collapsed question on the homepage.

No tier carries a volume: no lead count, message count, seat count or signal quota is attached to any of the five, so the only stated difference between the entry tier and the one above it is that the second synchronises with a system of record. Every call to action on the site leads to the same external demo form, so there is no self serve purchase path despite the plans being labelled self service.

And the vendor's own blog states that self service plans start at 749 dollars a month for a stated message volume, contradicting the 899 on its homepage. A buyer can read five prices here and still not know which one applies to them or what it buys.

The architecture moves the valuable output out of the platform continuously, which is a real position, and no contractual commitment stands behind it. Two mechanisms do the work. The targeted account list is exported weekly as a published feature rather than an on request favour, so a customer accumulates their own copy of the research output as the subscription runs.

And from the second tier upward, researched leads, sent activity and replies are pushed into the customer's own system of record, which means the pipeline record lives in a system the customer already controls. A departing customer on those tiers therefore keeps most of what they paid for.

What is absent is everything a contract would say: no post termination retention period, no deletion timeline, no deletion confirmation, no formal export of the platform's own record, and no statement covering the entry tier, which has neither synchronisation nor any documented route out. Terms of service are published and unread. Re verify there.

A named control set stated as included on every tier, which is the right shape for this axis. The vendor enumerates authentication by protocol, covering sender policy framework, domain keys and domain based message authentication, alongside mailbox warmup, send pacing and reputation monitoring, and presents the whole as a deliverability stack rather than as a feature to buy separately.

Naming the three authentication standards individually rather than gesturing at deliverability is what separates this from most of the field, and pacing and warmup being included at the entry price rather than gated upward matters more than it sounds.

What is missing keeps it below the top of the band: no bounce or complaint threshold is published, no blocklist monitoring, no inbox placement testing and no suppression mechanics, and the sending architecture itself is never described, so a buyer cannot tell whether messages leave from their own mailbox carrying their own reputation or from the vendor's infrastructure carrying somebody else's.

Top of the band, on a clear competitive frame and an absent customer picture. The vendor defines its market mostly by contrast, and does it well: contact databases find leads without researching or writing, one rival requires building custom automations against a credit system, two others are described as enterprise priced with limited customisation, and this product positions as the replacement for a three tool stack. That is a coherent statement of where it sits.

The named customers span staffing, financial services and recruitment, which suggests horizontal reach, and a fully managed service alongside the software indicates a buyer who may not have an outbound team at all. What is not published is any description of who the buyer is: no headcount band, no revenue range, no industry concentration, no geography and no customer distribution. The tier ladder cannot substitute for it either, because none of the five prices carries a volume, so a prospective buyer has no basis on which to place themselves.