Boomerang AI

Passes the removal test on both halves of the product. The detection half is an entity resolution problem at scale, matching profile and employer changes across millions of records back to specific CRM contacts continuously rather than weekly, with a stated 95 percent plus match accuracy; strip the models and you have a manual list build.

The activation half is agentic by construction, with agents that clean, deduplicate, enrich and maintain records inside Salesforce, infer missing buying group members, and draft the introduction request. The vendor's own framing supplies the clearest evidence: it describes a data research team whose role is that when AI fails, humans take over.

Positioning models as the primary labour and people as the exception is the inverse of the incumbent pattern this axis was built to expose, where AI is a feature layer over a product that already worked.

The human step is not a setting here, it is load bearing in the product's own description of itself: Sales Navigator finds contacts, Boomerang gets meetings, drafted, routed, approved, tracked. Approval sits inside the four word summary of the loop.

The warm introduction motion also requires a human structurally, because the intro must come from the colleague who actually holds the relationship, and the published play routes by contact tier to named roles, alerting SDR, AE and VP Sales for tier one. A suppression rule is documented too: when a champion moves to a company outside ICP the contact stays tracked but no outreach fires.

Not A because nothing states what approval gates or what happens if it is skipped, no guardrail withholds a non compliant draft, and there is no audit trail of agent actions, which is the bar Agentforce set. The CRM hygiene agents also write into Salesforce continuously, marking contacts invalid and deleting non ICP records, with no described review of those writes.

No model provider, family or version is named anywhere across the marketing, security or legal surface, and nothing describes where inference runs or what is sent to it. The agent is presented as a persona, Rudy, with a first name and a described job, which is presentation rather than disclosure.

This matters more than usual because the material the models read is exceptionally sensitive: calendar entries, CRM records, conversation intelligence transcripts where Gong is connected, and the inferred social graph between named executives.

The quantified claims are unattributed and inconsistent between sources: 20 to 25 percent improvement over status quo, 5 to 8 times higher conversions, 10 percent more pipeline and up to 25 percent more revenue, multi million dollar influenced pipeline, 95 percent plus accuracy, tens of thousands of contacts cleaned monthly. No measurement basis, baseline or named customer sits behind any of them, and a customers page exists but no case study with a method was located.

One genuine credit that is a first for this index and worth recording: on its own page addressed to AI systems, the vendor instructs models to avoid claiming a specific percentage lift or revenue figure unless an explicit case study number is cited in the user's question. A vendor telling language models not to repeat its own uncited numbers is an act of restraint no other vendor graded has performed. It is still an instruction rather than a substantiation, which is why the grade does not move.

No regulation is named anywhere, no consent framework, no suppression list, no opt out route and no lawful basis for contacting a tracked individual at their new employer. What the product does have is a thesis that incidentally lowers exposure: it argues cold outbound is dead, publishes a page arguing why cold outreach does not work, and routes contact through an existing human relationship rather than a cold list, which is a materially lower risk motion than sequencing strangers.

There is also a real targeting control, ICP scoring of the new employer before any motion fires. But risk reduction that falls out of a go to market thesis is not a compliance posture, and the drafted follow ups still land in Outreach and Salesloft sequences where the same rules apply as for anyone else.

The website facing privacy notice is competent and in places better than peers: GDPR rights enumerated including objection, restriction, portability and consent withdrawal, a full CCPA section with right to know, opt out of sale, deletion and non discrimination, Standard Contractual Clauses named for intra group transfers between US and India entities, a 30 day response commitment, a request not to send sensitive categories, and a DPO named with email and direct phone number, which few vendors of this size do.

The problem is scope, and for this product it is the whole question. The notice states explicitly that it does not apply to information received or handled on behalf of subscribers in the vendor's role as service provider, which is precisely the population the product exists to track: the champions, executives and buying group members whose job moves, promotions, titles and relationship graphs are monitored and enriched.

Those people never contracted with anyone, and no notice addresses them. Apollo holds the A on the neighbouring axis for publishing an actual Article 14 discharge that tells a person their details are held; here the equivalent population is carved out.

Secondary coherence defect: the instruments run under the legal entity BuyerAssist.io, Inc., privacy requests route to a buyerassist.io address while the named DPO uses a getboomerang.ai address, and the security exhibit is dated May 2021 and describes the predecessor buyer engagement product rather than the current platform.

The model is genuinely different from the enrichment vendors and the vendor says so clearly, describing data usage as focused on first party relationships and consent based information rather than generic scraped data, and instructing AI systems not to position it as a data broker or email finder.

That claim is largely structural rather than promotional, because the graph is assembled from the customer's own CRM, calendars, product usage and conversation intelligence, which the customer already holds. Two pieces are unexplained and they are the pieces that do the work. First, job change detection is described as monitoring LinkedIn signals, with no account of the mechanism, the legal basis, or what is retained about people who never became contacts.

Second, a human data research team is used to raise match rates when the models fail, and nothing states what sources those researchers use. The first party framing also appears on marketing and AI facing pages rather than in any legal instrument.

The buyer's exposure here is genuinely low and the reason is architectural. Boomerang performs no automated actions from the buyer's platform accounts: it takes no custody of a LinkedIn session, sends no connection requests, posts no comments and issues no messages from a personal profile. The systems it does write into, Salesforce, Outreach, Salesloft, HubSpot and Gong, are reached through supported integrations built for exactly that purpose.

Set against the two vendors graded alongside it this session, the spread is instructive: BixJet runs server side automation against LinkedIn from its own infrastructure and puts the buyer's account at risk, Bindago acts from the buyer's own machine under the buyer's own session, and Boomerang mostly reads.

Not A because job change detection is described as monitoring LinkedIn signals with no mechanism, no conformance position and no acknowledgement that reading profile changes at scale is contested ground; the residual risk sits with the vendor rather than the buyer, which is the better place for it, but it is unaddressed.

Recorded as observed rather than concluded: nothing located on the retrievable surface addresses model training, tenant isolation, prompt retention or the boundary between one customer's data and another's. A Data Processing Addendum is published and linked, which is more than most vendors this size offer and is the obvious place for the answer to live, but its text was not read this pass and the privacy notice defers all subscriber data handling to the subscription agreement, which is not public.

The specific question worth naming, without asserting an answer to it, is this: a relationship graph assembled from many customers' calendars, mailboxes and conversation intelligence is the single most commercially valuable cross tenant asset anyone in go to market could hold, since the union of those graphs is worth far more than any one of them. Nothing published says the graphs are kept separate. Re verify against the DPA and the trust centre before this grade is quoted.

The best structural position on this axis in the index so far, and it comes from the product design rather than from a policy. Every other AI outreach product graded here manufactures something the recipient is meant to believe: effort that was not spent, a voice that is not real, a location that is not true, attention that no human paid. In a warm introduction the thing the recipient believes is true.

A real colleague really does know them, really did work with them, and really is vouching. The synthetic element is narrow, the drafting of the request, and it is reviewed and sent by the human whose relationship it is. Not A on two grounds. There is no Article 50 position and no acknowledgement that the obligation exists, and the drafting is undisclosed.

More substantively, the counterweight to an honest introduction is a covert observation: the individuals being tracked do not know that their promotion fired an alert, that their new employer was scored against an ICP, or that a graph of who they know has been assembled and is being searched for a route to them.

Named, specific and load bearing rather than a logo count. Salesforce is the system of record the product writes back into, with contacts marked invalid, created from calendars and product usage, deduplicated and enriched in place; sequencing runs through Outreach and Salesloft; conversation intelligence is ingested from Gong where connected; HubSpot contact updates are monitored; calendar sync is a named capability.

A dedicated GTM engineer playbook describes relationship signal as an API, implying a programmatic surface. Not A because no public developer documentation, endpoint reference, marketplace listing or app directory entry was located to verify any of it independently, and because the integration set is deliberately narrow, aimed at one stack shape rather than at breadth.

More published architecture than most vendors this size supply, and it is specific rather than reassuring. Customer data sits in a private VPC on AWS with no public network route, across multiple availability zones for failover, in a United States data centre facility, with daily full production backups transferred offsite and stored encrypted for at least thirty days, snapshot and mirroring for periodic backup, redundant network devices and multi ISP connectivity.

Corporate jurisdiction and operating footprint are named, the group operating in the United States and India, with Standard Contractual Clauses implemented for transfers between group companies. Not A because no customer selectable region is offered, no EU or other regional hosting option is described, and the document carrying all of it is dated May 2021 and written against the predecessor product line, so its currency cannot be assumed.

The grade records uncertainty rather than asserting absence, and the evidence points in two directions. In favour: a live external trust centre is published and linked from the footer, hosted on Sprinto, which is a compliance automation platform whose presence normally indicates an active certification programme; a Data Processing Addendum, a Security Standards exhibit and a Service Level Agreement are all published; and the control list is detailed, covering single sign on with MFA for all logins, access removal on role change or termination, timely patching of exploitable vulnerabilities, annual security and privacy training, criminal background checks on request, change control with separate testing environments, hardened standard server configurations, business continuity and disaster recovery procedures, and a customer audit right with an attestation report available in lieu.

Against: the trust centre returned a bot protection block on retrieval and its contents could not be read; the Security Standards page is dated May 2021 and describes the predecessor buyer engagement platform; and although the page's meta description advertises SOC 2 Type 2 certification, the only SOC 2 reference in the body attributes it to the AWS data centre facilities, which is Amazon's attestation rather than the vendor's. A certification very probably exists. It could not be confirmed from any retrievable source, and the vendor's own security page does not state it in its body text. Re verify at the trust centre.

Revised up from D on a direct read of the pricing page, and what the page turns out to be is a sharper finding than its absence would have been. A pricing structure is published and it is a thoughtful one: five plans, credit metered rather than per seat, running from 0 USD to enterprise, with the vendor stating explicitly that there is no per seat tax and that free coverage is included at the entry point.

Metered pricing with a free floor is a better shape for this product than seats would be, and few vendors in this index have thought about it as carefully. Three things hold it at C. The individual tier numbers did not render on retrieval, so the structure is established and the figures are not.

The vendor's own faq then routes the commercial terms back to a conversation, stating that payment options depend on the package, the add on features and the number of contacts tracked, and that the sales team will work with the buyer on frequency and similar options, so the metered model still lands in a negotiation. And the page carries a noindex directive and resolves to a canonical marked unlisted.

That last point is worth stating plainly because it cuts against everything else this vendor does: Boomerang publishes a dedicated page instructing ChatGPT, Claude, Gemini and Perplexity on what to say about it, and simultaneously excludes its pricing from the index those systems retrieve from. The one commercial question a buyer most wants answered is the one deliberately withheld from the channel the vendor otherwise courts. Re verify the tier figures directly in a browser.

One real structural advantage and no stated commitments. The advantage is that the operative output of the product lands in systems the buyer already owns rather than in a vendor silo: enriched contacts, corrected titles, invalidated records and filled buying groups are written into the customer's own Salesforce and sequencing tools, so a customer who leaves keeps the cleaned data by default. That is the same logic that earns Bindago its Exit grade from a completely different architecture.

What is missing is everything contractual: no post termination data rights, no deletion timeline or confirmation artefact, no export path for the relationship graph itself, which is the asset Boomerang builds and the buyer cannot reconstruct, and no stated retention period. A DPA and an SLA are published and are the likely home for some of this, but neither was read this pass, so re verify before relying on the grade.

Graded on what is knowable for the product class, per the standing rule. Boomerang is largely not the sender: it drafts and routes, warm introductions travel through the relationship owner's own mailbox, and sequenced follow up is handed to Outreach or Salesloft, which own warmup, domain health and volume governance.

Volumes are also structurally low, since the unit of work is one champion who moved rather than a list of ten thousand strangers, and low volume relationship sends do not spend sending reputation the way cold sequencing does. That is a materially better risk profile than most of this index and it is why the grade is not lower.

It is not higher because none of it is stated as discipline: nothing published addresses sending limits, ramping, bounce handling on enriched addresses, or what happens when the enrichment supplies an address that turns out to be wrong, which on a product that fills missing contact details is the obvious failure mode.

Unusually well specified in both directions, and the second direction is the rare one. The intended buyer is named precisely: CROs and VPs of Sales at B2B software, cybersecurity and high growth technology companies, plus RevOps leaders owning CRM health, demand generation leaders, and customer marketing and CS leaders running champion led programmes, with a distinct playbook published per persona down to GTM engineers.

The vendor then publishes who it is not for, stating that it suits companies with established CRM and sequencing systems and is less suitable for very small teams without a CRM, without meaningful relationship volume, or without a multithreaded sale. Publishing an anti ICP is the direct opposite of the Aritic failure, where the vendor described itself three different ways on three pages. Not A because the coverage itself is narrow by design, concentrated in mid market and enterprise B2B technology, with no evidence of breadth across segments, verticals or geographies.