HubSpot Sales Hub

Remove every model and a complete, saleable and very large product remains: a system of record with deal pipelines, forecasting, email sequences, calling with recording, meeting scheduling, document tracking, quotes, task queues, reporting, forms and an application marketplace. That is what this company sold for close to two decades and it is still the substance of what a sales team buys.

The agent layer sits above it and is packaged as an addition in the most literal sense, since agents are metered per action on top of a seat price rather than being included in it, and the vendor describes charging when an agent resolves a conversation, drafts outreach or answers a question. A capability the vendor itself bills as a consumable alongside the subscription is being positioned by its own price list as an enhancement to the product rather than as the product. This is the established platform pattern applied consistently, and it separates this vendor from the model native startups in the index by vintage rather than by marketing volume.

Governance here is administrative and it is real. Agents are individually activatable rather than on by default, with a console showing which are running, how each is performing and which have not been turned on. What data any artificial intelligence feature may reach is separately configurable, covering records, conversations and files independently.

Changes to those settings require the highest permission level and are written to the audit log, so a buyer can reconstruct who enabled what and when. A builder lets customers scope custom agents to defined tasks rather than open ended instruction. Per action metering supplies an incidental but genuine control, since every autonomous act is billed and therefore visible on an invoice.

Held off the top band because the benchmark on this axis remains a competitor that publishes engagement rules governing when an agent may work a lead, guardrails that withhold a violating response rather than display it, configurable human approval escalated by risk, and an action level audit trail of agent outputs. Activation control and data scoping are the outer boundary of what an agent can touch; they are not a published account of what happens when one is about to do something it should not.

The artefact that earns this is one almost nobody in this corpus produces: published model cards, per feature, describing how each capability works, what data it uses and what safeguards apply, offered explicitly so a buyer's security team can review and verify rather than take the marketing on trust.

Around them sits a register of processors that names the model providers engaged and, for each one, the location where data is processed against the customer's own hosting region, so the question of what runs where has a published answer rather than an inference. A dedicated trust page for artificial intelligence sets out the design and deployment guidelines followed, states that every feature is tested before launch and monitored after, and points to where the detail lives.

The vendor is also specific about which of its own features are trained across accounts rather than describing training in the abstract. Held at the top band on the strength of per feature documentation, named providers and published processing locations together; what remains unverified from the pages read is model version detail, and the cards themselves were not opened on this pass and should be re verified.

The evidence base is deep and most of it sits outside the vendor's control. The company is publicly listed, so customer counts, revenue and retention are filed and independently auditable rather than asserted, and its scale claims of more than 299,000 customers across more than 135 countries are checkable against those filings. One review platform carries more than thirteen thousand reviews at a rating above four, which is among the largest independent samples in this corpus.

A substantial case study library is published, and named customers give specific figures, including one quantifying roughly three hundred and fifty hours recovered from a single automation with the basis for the estimate stated.

Recorded as defects rather than disqualifiers: headline percentages such as the share of customers reporting increased revenue carry no method, period or sample; and independent pricing guides disagree with each other about what a professional seat costs and what it includes, which is a sign the packaging is complex enough that even specialists reading the same public pages reach different answers.

The compliance machinery shipped to customers is among the strongest in the index and it is mostly aimed at the electronic mail channel. Lawful basis to communicate tracking lets a customer record, per contact, why that person may be messaged, which is a considerably more demanding artefact than a bare unsubscribe link and is the single best consent instrument graded here. Subscription types and preference management give recipients granular control rather than a binary opt out.

Permanent record deletion supports erasure requests. Cookie consent banners are configurable by region. Held off the top band on two gaps. No enforcement mechanism was located of the kind that earned another vendor in this category the top band, meaning no published complaint or bounce threshold, no account review on abnormal unsubscribe rates and no stated sending conduct policy.

And the calling side is unaddressed: the product records calls and dials from the record, and no telephone marketing regulation, do not call obligation, calling window or recording notification position was located anywhere on the pages read.

Two halves and the second is what lifts it. On the vendor's own obligations: a data processing agreement covering the European regulation, commitments on the other processors engaged, breach notification timelines, and international transfer mechanisms named as both standard contractual clauses and the current adequacy framework rather than an invalidated predecessor. A register of processors is published and broken out by hosting region.

An independent enterprise privacy certification is held. Sensitive data can be stored in dedicated properties visible only to the highest permission level and protected by per tenant encryption. On the customer's obligations, and this is rarer: the platform ships the machinery a controller needs to discharge its own duties, including a permanent record deletion function built for erasure requests, lawful basis to communicate tracking so a customer can record why each contact may be messaged, subscription and preference management, and cookie consent banners configurable by region. Most vendors in this index publish their own compliance. This one also builds their customer's.

This is the conspicuous gap in an otherwise exemplary disclosure record, and its shape is worth stating precisely. On the customer's own data the vendor is unusually clear, publishing processors, hosting regions, training positions and controls. On the data it supplies about other people it is silent.

Record enrichment was previously provided free and has been consolidated into a metered credit product, which makes the supply of third party company and contact information a commercial line rather than a courtesy. No provenance disclosure for that inventory was located: no sources named, no lawful basis stated for holding information about individuals who never contracted with anyone here, no accuracy or match rate published, and no notification or removal route for the people whose records are sold onward as enrichment.

Held above the bottom band because the enrichment is a disclosed and separately metered product rather than a one, and because the processor register and processing agreement provide a partial route into how supplier relationships are governed. The comparison inside this index is direct: the largest data vendor graded here publishes a source taxonomy, a transparency page and a per subject notice stating where information was obtained.

The buyer carries very little exposure here and the reason is structural: this vendor is itself the platform others build on, so its own terms govern the ecosystem rather than the other way round. Where it does depend on third parties, the connections are sanctioned. Mail client integration runs through delegated authorisation with the major providers and is distributed as published extensions and add ins through their own stores.

Professional network integration operates through the network's own sanctioned sales product rather than through scraping or browser automation. There is no credential rotation, no multiple account handling and no automation of a platform that forbids it.

Held off the top band because no conformance position is published for the bulk sender requirements the major mail providers introduced, which matter to any customer sending at volume from a connected mailbox, and because the dependency the buyer should think hardest about is the reverse one: an ecosystem of tools they may also be buying exists at this platform's discretion, and nothing addresses what happens to those tools if the marketplace terms change.

This is the most carefully drawn position on the training question in the index, and it earns the top band by disclosing more rather than by claiming less. Three separate things are stated and kept distinct. First, third party model providers are contractually prohibited from training on customer data, with zero retention enforced wherever possible, and the register of processors names each provider alongside the location where it processes data for each customer region.

Second, and unusually, the vendor states plainly that its own models do train on customer data, and then explains the mechanics rather than leaving them to inference: some models are trained only on a single account's data to tailor features to that customer, while others are trained across many accounts to improve accuracy platform wide, with email segmentation, spam detection and deal close prediction named as examples.

It states that the resulting models do not contain or reproduce individual customer data and that proprietary content is never shared between customers. Third, there is a real control. Training is an account level toggle requiring the highest administrative permission, changes are written to the audit log, opting out degrades no feature because it is a separate setting from data access, and accounts holding sensitive data are opted out by default and cannot opt in.

Set this against the other large vendors here, where a blanket assurance of no training sits beside marketing about insight drawn from aggregate customer patterns and the boundary between the two is never drawn. This vendor draws it.

What is published is the machinery a customer would use to treat recipients well, and it is genuinely good machinery: lawful basis tracking recording why each person may be contacted, subscription types and preference management giving recipients granular control, permanent deletion supporting erasure requests, and regionally configurable consent banners. What is absent is any statement of the vendor's own behaviour where its systems act on a recipient.

A prospecting agent researches accounts and drafts outreach, and nothing published takes a position on whether the person receiving that message is told a machine composed it, or addresses the European transparency obligations that took effect in August 2026, under which the marking duty rests principally on the provider of the system rather than on the deploying customer. Calls are recorded from the record with no notification position stated. The pattern matches the other large vendors here: extensive configurability for the customer's compliance, and silence on the duty the regulation places on the vendor itself.

This vendor is not merely well integrated, it is the platform other vendors integrate to, and the index itself is the evidence. Five separately graded vendors here name it: a digital sales room lists it as an integration, a conversation intelligence vendor names it as a customer, a data platform ships a connector for it, a video vendor offers both native and extension based versions of it, and a speed to lead vendor exists only inside it and cannot be bought without it.

Underneath that sits a public application marketplace running to well over a thousand listings, documented public interfaces, a developer platform, workflow actions third parties can extend, and a certified partner programme. The platform architecture is the point: the shared record layer means an integration written once reaches the marketing, service, content and data products as well as this one. Held at the top band on breadth, depth and demonstrated gravitational pull together.

Not established on this pass and worth re verifying: whether support exists for the emerging protocol that lets external agents query a vendor directly, which several smaller vendors in this index already ship.

The strongest residency record in the index and the buyer decides it. Product infrastructure is hosted in regional data centres across four named territories covering the United States, the European Union, Canada and Australia, and a paid account can be migrated to a preferred region rather than being assigned one by signup domain or address.

Documentation covers how regional hosting actually works, and it goes further than any comparable vendor here by publishing the limited circumstances in which processing can occur outside the chosen hosting location, which is the exception most residency claims quietly omit. The register of processors is broken out by hosting region, so a customer can see for each supplier where their data is handled, including for the model providers behind the artificial intelligence features.

Underlying infrastructure runs on major public cloud providers whose own certifications are named and attributed to them rather than claimed as the vendor's own, which is the correct handling of a distinction several vendors in this index get wrong.

Depth and, more importantly, access. Two attestations are held in the vendor's own name with their types stated, and the summary level report is downloadable from the trust centre with no confidentiality agreement required, which is the distinction that separates the top band from the one below across this index. Penetration test summaries and a security overview document are served the same way.

A trust centre publishes compliance materials alongside live status and incident information, a formal vulnerability disclosure programme runs through a named third party platform, and the security programme is described with annual independent audits, continuous monitoring and a defined incident response process.

Because the company is publicly listed, its key technology controls are additionally audited under the financial reporting regime and that is disclosed in its regulatory filings, which is an assurance layer no private vendor here can offer.

Product side controls are enumerated rather than asserted: federated single sign on with named identity providers, enforced sign on with exemptions, automated user provisioning, comprehensive audit logging, inactive session timeout, application level per tenant encryption and role based access. Health information handling is supported with an agreement incorporated by reference into the sensitive data terms. Recorded for accuracy: the international information security standard is held by the infrastructure providers rather than by this vendor, and the vendor attributes it correctly.

Substantially more is published than at any other large vendor in this index, and the things that decide the real bill are still missing. Published: four tiers with per seat rates, a seat taxonomy distinguishing full sales seats from cheaper core seats from free and unlimited view only seats, one time onboarding fees stated at fifteen hundred and thirty five hundred dollars for the two upper tiers, contact allowance overage sold in priced blocks, and a product and services catalogue in the legal centre that defines seat access contractually and can be read before signing.

Free view only seats are a genuine and unusual commitment. Against that, the load bearing pieces are unpriced. The agent layer bills per action, with the vendor stating that a customer pays when an agent resolves a conversation, drafts outreach or answers a question about their data, and no rate for any of those was located.

The credit economy that now carries enrichment, following the consolidation of previously free record enrichment into metered credits, is likewise unrated on the public pages. All paid plans require annual commitment. And third party reporting consistently describes heavy negotiated discounting, so list price and paid price diverge by an amount a buyer cannot compute. This is the highest grade any large vendor here has taken on this axis, which says as much about the peer group as about the vendor.

The architectural position is strong. Documented public interfaces mean a customer's records, activity and history are retrievable programmatically rather than through a support request, and the free tier means an account can persist as a readable archive after paid features lapse rather than the data disappearing with the subscription.

A permanent record deletion function exists and is built for erasure obligations, so deletion is a customer operated capability rather than a vendor promise. The data processing agreement sets out breach notification and processor commitments in writing.

Held off the top band because the end of the relationship as a whole is not addressed on the pages read: no post termination retention period, no deletion timeline for a closed account, no deletion confirmation artefact and no statement of what happens to call recordings, agent outputs and the artificial intelligence derived fields once a subscription ends. Deleting one record is documented; closing one account is not.

The selling product's architecture puts reputation in the right place. Sequences send through the representative's own connected mail account rather than through shared platform infrastructure, so the buyer inherits and keeps their own sending reputation and no other customer's behaviour can damage it.

Third party reporting describes an enrolment ceiling of fifty contacts at a time on sequences, which if accurate is a real volume governor built into the product rather than a policy asking for restraint. Subscription management and preference handling reduce complaint pressure at source.

Held off the top band because no specific control was located on the selling surface: no sender authentication standards, no warm up guidance, no blocklist monitoring, no complaint or bounce rate threshold and no placement testing. The wider platform's marketing product has a longer history on this axis and was not examined here, so this row grades the selling product a sales team would actually buy and should be re verified against the marketing side if a buyer is purchasing both.

The tier ladder is the segmentation and it is unusually legible. A free tier serves individuals and very small teams with a genuine product rather than a trial, a starter tier is priced per seat for lean teams, and the two upper tiers add automation, forecasting and governance for larger organisations, with the seat taxonomy letting a buyer mix expensive full seats for sellers with cheap core seats for everyone else and free seats for observers.

That structure describes who each tier is for more precisely than most segment pages manage. Scale evidence corroborates the breadth, with a customer base spanning more than a hundred and thirty five countries. Held off the top band on the pattern that has caught every large vendor here: the same platform is presented as suitable for a solo founder and for an enterprise simultaneously, no headcount or revenue band is published for any tier, and no ceiling is stated. The counter example in this index remains a vendor that publishes the point at which a customer should leave for a larger product, and nothing comparable appears here.