6sense

The predictive model is the product rather than a layer on top of one. The company was founded in 2013 as a predictive intelligence engine and its own trust page still describes it that way, the buying stage prediction it sells is patented, and the homepage claims models built per customer against a signal graph the vendor dates back thirteen years.

Remove the model and what remains is a raw intent feed and an advertising activation tool, which is a materially smaller product rather than the same product with a feature switched off. That is what separates this from the incumbents whose assistants sit on top of a complete system of record or content library.

Held below the top band because the newest layer is separable in exactly the way the core is not: the conversational assistant, the sales copilot and the email agents are recent additions that could be stripped while the scoring engine kept working, and the marketing volume around them outpaces what is evidenced. Ask which AI branded components are separately licensed, and whether account scoring is available without them.

The agents send and reply without a human approving each message, and the controls published are configuration rather than runtime. The email agent page states that agents create and send outreach, send follow ups, read responses, send context aware replies, and decide when to loop in a sales representative.

The oversight described is all set before launch: preview the sequence, adjust reply tone, add custom instructions for handling objections and meeting requests, and preview agent responses. Once running, no published mechanism lets an operator halt a batch in flight, cap volume, or require review of a reply before it reaches a buyer. Two things pull the other way and both are real.

The trust page names guardrails as input validation, output filtering and continuous output monitoring, and describes penetration testing of language model features against a recognised industry risk list. And the privacy policy states, in a section disclaiming automated decision rights, that processing is not solely automated and human intervention is actively part of the process, which is a commitment made where it carries legal weight. That statement covers scoring and profiling rather than send authority. Ask what stops an agent mid sequence, what the daily send ceiling is, and whether a reply can be held for human review.

More architectural disclosure than most, and nothing on the models themselves or on whether the predictions work. On the disclosure side the vendor states that models are built per customer rather than pooled, describes output as cited intelligence answering why this account and why now, and publishes a service organisation control report scoped specifically to the AI email product, which is an unusually granular thing to break out.

Against that, no model provider, family or version is named on any page read, no model card exists, and for a company whose central claim is predicting buying stage, no accuracy, precision, validation or backtest figure for the predictive engine appears anywhere. The comparative claim of ten times the intent coverage of other providers carries no measurement basis. The gap is specific: a buyer can learn how the system is governed and cannot learn how well it predicts. Ask for the predictive model's measured accuracy against realised outcomes for a stated customer cohort, and which third party models process customer content.

Quantified customer outcomes in volume, and a contradiction sitting inside them. Named accounts carry figures across the site: a close rate improvement at an enablement vendor, deal size and reply rate at a field service software company, business development hours saved and pipeline generated at a data management vendor, and open rate and hours saved at a cloud services firm. None states a population, a baseline or a measurement period.

More seriously, one page gives two different results for the same customer, describing an increase in sales productivity of twenty six percent in one place and an increase in average opportunity value of two hundred and sixty percent in another, for the same named account. Those are different metrics at a tenfold difference in magnitude and both cannot be the headline finding.

Third party analyst placement is genuine and independent, with leader positions in evaluations of intent data providers, revenue marketing platforms and conversation automation, though those measure vendor standing rather than customer outcome. The vendor also claims ten times the intent coverage of other providers with no methodology attached. Ask for reply rate and pipeline conversion measured across a stated customer population and period, and ask which of the two figures for the same account is correct.

Suppression is pushed into the customer's workflow rather than left as the customer's problem. The vendor operates an opt out list inside the product: when a data subject asks to be removed from the database, that person is added to a customer specific list, primary administrators receive a weekly email of new entries, and the documented instruction is that administrators review the list monthly and remove those contacts from their own internal systems unless they hold an independent lawful basis.

That is a vendor propagating a suppression obligation downstream to the people who exported the record, which almost nothing in this index does. Alongside it the privacy policy states that telephone do not call and do not mail lists are maintained as required by law, that Global Privacy Control and Do Not Track browser signals may be used to opt out of collection and sharing, and that the vendor follows the standards of the advertising industry trade body.

Held below the top band because nothing published states what the sending product enforces at send time: whether the agent checks the opt out list before a message goes out, or whether that check depends on an administrator having done the monthly cleanup. Ask whether opt out suppression is enforced automatically at send time by the email agents, or only through the manual list review.

The strongest privacy documentation recorded in this index, and almost all of it is verifiable without asking anyone. The policy names its lawful bases individually and states that individualized notice is given to data subjects whose records are added to the business database, which is the specific obligation that applies when data is collected from somewhere other than the person, and very few vendors say they do it.

Transfers run on the current framework, with certification covering the European arrangement, its United Kingdom extension and the Swiss arrangement, an independent dispute resolution route through a named non profit, stated regulator jurisdiction and residual binding arbitration, plus standard contractual clauses for customer data.

The United Kingdom office publishes its regulator registration number, and an external data protection officer and a separate European representative are named with full postal addresses and a telephone number. The California request metrics are published as a table with real counts for the 2025 calendar year, including four hundred and four deletion requests, all complied with, at a median of two days. That table is the kind of artifact a marketing team cannot write. Two small defects: the footer dates the policy one day later than the document header, and the address on the California data broker registration does not match the registered address on the site.

The sourcing model is described in categories and the individual sources are not named. What is disclosed is substantial. Business profile records come from third party providers, partners and integrators drawing on data cooperatives and publicly available sources. Web and intent identifiers come from networks, exchanges, aggregators and publishers.

The vendor states it does not reverse engineer those identifiers to a named person and instead resolves them to the company, which is an architectural limit on the intent side rather than a promise about behaviour. It is a registered data broker in California and declares itself one under Texas law, and its own disclosure lists the exact categories of business personal information sold in the preceding twelve months.

It also draws a boundary that matters commercially, stating that customer provided data is not added to the shared business database under current subscriptions. What is missing is the layer beneath: no publisher is named, no cooperative is named, no third party provider is named, and the assertion that intent vendors maintain opt in and consent policies with their publishers appears on an executive blog rather than in any document. Nothing addresses indemnification. Ask which cooperatives and publishers supply the proprietary intent, and whether the vendor indemnifies a customer against a provenance claim.

Activation runs through sanctioned channels and one industry rule set is named explicitly. The privacy policy states membership of the advertising industry trade body and adherence to its standards and guidelines, which is a named conformance position governing the advertising activation that forms a large part of the product.

Integration into customer systems runs through official connectors, with a partner library, a productized data interface, a model context protocol server for agent access, and named delivery into customer record platforms, sales engagement tools and data warehouses. Sending runs through the customer's own provider by election, with the vendor's own infrastructure, a major webmail service or a major enterprise mail service offered as choices, rather than through unsanctioned relay.

Two things hold it below the top band. A browser extension is shipped and nothing published states what it collects or under whose terms it operates. And the web deanonymization capability depends on identifier collection from third party networks whose own platform terms are described nowhere. Ask what the browser extension collects and under which platform terms, and what happens to the deanonymization capability as third party identifiers are deprecated.

Model governance is independently audited and the certificate is downloadable rather than described. The vendor holds the international standard for artificial intelligence management systems, assessed annually by an outside auditor, and publishes the certificate itself as an open document.

Around it sits a governance programme with substance rather than a policy statement: a cross functional council drawn from data science, engineering, security and legal, a secure development lifecycle specific to AI, documented model risk assessments, penetration testing of language model features guided by a recognised industry risk list for such systems, and named runtime guardrails covering input validation, output filtering and continuous output monitoring.

The vendor also publishes a separate service organisation control report scoped only to the AI email product, meaning the component that writes and sends to buyers has been examined on its own rather than folded into a platform wide attestation. It has signed a cloud industry trustworthy AI pledge. This is the second A recorded on this axis and it rests on a broader base than the first: the same audited standard, plus a product scoped attestation and a published testing methodology. The gap is output quality rather than governance, since no evaluation or error rate for agent generated email is published. Ask what evaluation covers agent written copy before it reaches a buyer.

The vendor markets undisclosed machine authorship as a selling point, in its own words. The email agent page describes personalization that makes a message sound like the customer, and then adds that it does so even when it is not the customer writing.

The same page states that agents send outreach, read responses and send context aware replies automatically, so a buyer can hold a multi turn exchange with a system while believing they are corresponding with a person, and the handoff to a human happens only when the agent decides to make it.

Nothing on that page, the trust page or the privacy policy states that AI authorship is disclosed to the recipient, offers an operator setting to disclose it, or describes any marking on generated messages. This is the axis where the product's design and the vendor's own marketing point the same direction, which is why it lands here rather than at a grade held down for silence.

Pulling slightly the other way, and the reason this is not the bottom of the scale, the vendor does give individualized notice to people added to its database and publishes a public opt out route, so a person is told their data is held even though they are not told who wrote the email. Ask whether AI authorship can be disclosed in the message, and what a recipient asking directly whether they are corresponding with a human will be told.

Distinct delivery surfaces rather than a connector count. Intelligence is published into customer systems four different ways: native integrations with customer record and sales engagement platforms, a productized data interface sold as its own line item, a model context protocol server aimed at agent consumption, and direct delivery into customer data warehouses.

A partner library is published with separate views for technology integrations and agency partners, and the intent layer accepts third party feeds from named providers rather than only its own, which is an openness most signal vendors avoid. A browser extension extends the sales product into the browser, and a public status page reports platform availability.

Held below the top band on retrieval rather than on architecture: no developer documentation was reached on the routes taken this pass, so the interface depth, authentication model and rate limits behind the data interface are unverified, and the integration inventory was not counted against the partner library. Ask for the data interface documentation and the current verified integration list.

The hosting is named and everything a buyer needs in order to place their data is not. The vendor states the platform runs on a major public cloud with redundant multi system architecture designed against defined recovery time and recovery point objectives, and that it is a United States company whose operations necessarily involve international transmission. Cross border movement of customer data is supported by standard contractual clauses where required.

What is absent is any residency election: no region is named, no European or United Kingdom hosting option is described, no tenancy model is stated, and the recovery objectives are referred to without publishing their values. The transfer language is broad, covering the United States and unspecified affiliate and service provider servers.

For a platform holding identified buying signal across a customer's entire target market, where that data physically sits is a first order procurement question and the public surface does not answer it. Ask which regions the platform can be pinned to, what the stated recovery time and recovery point values are, and whether European customer data can be kept in region.

Certifications held, and the evidence published rather than promised. The vendor undergoes an annual independent audit for the international information security standard and publishes the certificate as an open document, and holds a service organisation control type two report covering all five trust services criteria rather than security alone.

Critically it publishes the public form of that report as a downloadable file, so a buyer can read an auditor's opinion before any sales contact, which is the step almost every vendor in this index stops short of. The auditing firm is identifiable from the certification seals. Independent validation letters for both the European and California privacy regimes are published as open documents from a named assessor.

A cloud industry consensus assessment questionnaire is filed in the public registry, annual network, web application and interface penetration testing is performed by an accredited assessor, a vulnerability disclosure programme runs on a public platform, a status page is live, and the subprocessor list is public rather than gated. The security organisation is led by a named chief information security officer with the privacy function under a named lead.

A deeper document set sits behind a trust portal requiring an agreement, which is the normal ceiling, but unusually little of substance is behind it. Ask for the audit periods covering the current reports.

A pricing page with no price, and a navigation promise the page does not keep. The site menu invites a visitor to explore sales intelligence pricing and get started for free, and links to a page titled pricing. That page describes three packages in real detail, naming what each contains down to individual capabilities and explaining that data credits unlock contact export, and attaches no figure to any of them.

There is no rate, no band, no unit of pricing, no credit price, no seat minimum and no free entry path visible on the page the free offer links to. Every call to action is a demo booking. The detail makes the omission conspicuous rather than excusable, because the vendor has clearly done the work of defining its packaging and has chosen not to price it.

Third party sources place most deployments between sixty thousand and one hundred thousand dollars a year with full enterprise access frequently above two hundred thousand, a range too wide to budget against. This is the second consecutive enterprise vendor at this grade and the pattern holds that the split is sales led against self serve motion rather than company size. Ask for the entry rate for each named package, the price per data credit, and the seat minimum.

Individual rights are documented in detail and the customer's own exit is not addressed at all. A person can obtain, correct, delete or port their record through a published privacy centre, by email, or through a toll free number, and the vendor publishes how many such requests it received and how quickly it answered them. That is the data subject side and it is handled well. The subscriber side is the gap.

Nothing on the public surface states what a departing customer can export, in what format, whether account scores, model outputs, engagement history and audience definitions come out or stay behind, what happens to enriched contact records after termination, or what the deletion timeline is. Retention is described only in purpose based terms, meaning as long as needed to provide services, without a stated period.

The exposure is specific to this category: predictive scores are derived from a signal graph the customer never possessed, so a departing customer may find that what leaves with them is far thinner than what they were using. Ask whether model outputs and scores are exportable or terminate with the subscription, what happens to enriched records after termination, and what the deletion timeline is.

Real sending machinery, described at the level of a product that has had to think about it. The email agent product gives the operator a choice of sending path, either the vendor's own infrastructure or the customer's existing webmail or enterprise mail service, which keeps reputation on the customer's domain where the customer prefers that.

Dedicated agent inboxes are provisioned rather than sending pooled across tenants, contact validation runs before outreach begins so invalid addresses are caught ahead of send, and bounce rate monitoring is presented as an operator facing control with sender reputation named as the thing it protects. That is a fuller set than most sending products in this index publish.

Held below the top band because none of it is quantified or committed: no bounce or complaint rate threshold is stated, no warmup or ramp process is described for a new agent inbox, no policy is published on what happens when a customer's sending degrades, and nothing addresses whether one customer's poor performance can affect another on shared infrastructure. Ask what bounce and complaint thresholds trigger intervention, whether agent inboxes are warmed before first send, and what isolates one tenant's reputation from another.

The target is stated precisely and the limits of the data are not. Nine industry pages are published covering asset management, banking, business services, financial services, financial technology, insurance, manufacturing, technology, and transportation and logistics, alongside five named use cases spanning account based marketing, inbound automation, outbound automation, deal acceleration and customer expansion.

The international footprint is real rather than claimed, evidenced by a registered United Kingdom office and a named European representative entity with its own address. Intent coverage is claimed across more than forty languages. Two things hold it below the top band and they run together.

Third party review analysis reports that signal quality is strongest in technology, software and financial services where the publisher network is dense, and weaker for niche industries and regions outside North America, and the vendor's own pages carry no such qualification anywhere. And the site is published in English only, with no localised versions, which sits oddly beside a forty language coverage claim. Ask for intent coverage and match rates broken out by region and by industry, particularly outside North America.