Dialpad

Speech recognition is in the transport path rather than layered over it, and the commercial choice proves the architectural one. Transcription, live sentiment analysis, call summaries and voicemail transcription run on every call across every plan including the fifteen dollar entry tier, where competitors commonly meter the same capabilities as paid add ons. A vendor giving away at the floor what rivals sell as an upgrade is telling you the capability is not a separable module.

The autonomous voice agent product is fully model dependent by definition, since removing the model leaves nothing to answer the call. Against that, the underlying business is a cloud phone system carrying voice, messaging and meetings, and strip every model and a functioning telephony platform remains, which is what most of the customer base buys. So this is a genuine model dependency in the layer the vendor differentiates on, sitting on infrastructure that predates it. Ask which capabilities disappear entirely without the speech layer, and whether the agent product is licensed separately from the phone system.

An autonomous product is sold with the human explicitly removed, and the governance around it is named without being published. The agent product is described as handling voice and text interactions end to end without a human in the loop, which is the highest autonomy level of any vendor built in this project so far: it speaks to a customer, decides what to say, and completes or escalates on its own judgement.

The trust centre does carry a dedicated artificial intelligence section listing an overview, a governance entry, an evaluation and red teaming entry, and a separate policy governing AI use and development, so the questions are anticipated and answered somewhere. None of that substance is readable without an access request.

What a buyer cannot establish before a sales conversation is what the agent may commit to on a call, what triggers escalation to a person, what caps exist, and what a recording shows afterward. Ask what the agent can commit to unsupervised, what escalates, and request the evaluation and red teaming documentation.

A model provider is identifiable from the vendor's own subprocessor list, which is a more reliable disclosure than a marketing claim. The trust centre publishes a subprocessor list naming, among others, a large language model provider and a data annotation company, so a buyer can see both who processes inference and that human labelling is part of the pipeline.

Naming an annotation vendor is unusual and quietly informative, because it indicates audio or transcript material is reviewed by people somewhere in the workflow. Alongside it the centre carries an artificial intelligence overview, a governance entry and an evaluation and red teaming entry, and a standing policy on AI use and development, which is a fuller governance shelf than most vendors maintain.

Held below the top band because the substance sits behind an access request, no model version or evaluation result is published openly, and no accuracy figure exists for transcription or sentiment analysis despite those being the product's core claims. Ask which models handle transcription against which handle generation, what the annotation vendor reviews, and for published evaluation results.

Recognisable customers at scale and no measured outcome behind them. The trust centre names large media, telecommunications, staffing, retail and healthcare organisations as customers, which establishes the platform operates at enterprise volume, and independent review coverage is extensive across the business phone category.

What is absent is any published figure tying the product to a result: no transcription accuracy rate, no connect rate improvement, no measured effect on handle time, resolution or conversion, and no study across a stated customer population and period. For a vendor whose central claim is that speech AI improves conversations, transcription accuracy is the foundational measurement and it appears nowhere.

Independent review coverage also raises a recurring operational complaint about billing clarity and support responsiveness, which is a cost of ownership signal from outside the vendor rather than a product one. Ask for measured transcription accuracy by accent and call condition, and for outcome measurement across a stated population.

A carrier grade platform with the regulatory apparatus present in the ecosystem and absent from the vendor's own published surface. Everything located on the specific obligations that govern this category came from third parties rather than from the vendor: a connectivity partner describes providing caller authentication signing and attestation logging, application to person messaging registration with opt in management, emergency location compliance and short duration call filtering for organisations running this platform, which indicates those controls are commonly sourced alongside rather than assumed within it.

On the vendor's own pages no statement was located on consent capture, preference register screening, calling window enforcement, abandoned call rate limits or where responsibility transfers to the customer. Aloware, graded in this same lane, engages the regulator by name, cites the abandoned call limit and the narrowed dialer definition, and states plainly which obligations remain the customer's. Ask what preference register screening and calling window enforcement are built into the outbound products, and whether caller authentication attestation is native or requires a connectivity partner.

A broad and correctly constructed privacy shelf, with the documents themselves gated. Transatlantic transfer certification is listed alongside named regime coverage spanning the European regulation, the California statutes in both their original and amended form, the Brazilian law and the Canadian federal law, which is genuine multi jurisdiction breadth rather than a single badge. A processing addendum is published as a featured document.

The subprocessor list is public and enumerates actual companies rather than categories, which is the single most useful privacy artifact for a platform that routes call audio to third parties. Data retention, data management, breach notification and cookie entries each have their own place in the centre, and employee privacy training is maintained.

Held below the top band because the substantive documents require an approved access request, no retention period is stated openly, and no data protection officer or European representative was located, where the vendors at the top of this axis publish request metrics, registration numbers and named representatives. Ask for the retention schedule covering call recordings and transcripts, and the processing addendum.

No external corpus is licensed and the provenance question here is about voice. Nothing is purchased, scraped or resold; the material the platform holds is call audio, transcripts and messaging generated by the customer's own conversations.

The question that follows is whether that audio trains the speech models the vendor differentiates on, and it is sharpened by the vendor's own subprocessor list naming a data annotation company alongside a model provider, since annotation implies human review of material somewhere in the pipeline. Nothing published states what that vendor reviews, whether customer call audio is included, whether recordings or transcripts contribute to model improvement, or what consent basis covers it.

Speech models improve with real conversational audio, which makes this the commercially obvious question and the one the open surface does not answer. Ask whether customer call audio or transcripts are used to train or tune models, what the annotation vendor processes, and whether that behaviour can be disabled.

The vendor owns the layer that usually constitutes the exposure. Rather than reselling another provider's telephony, it operates its own carrier infrastructure with published interfaces for session initiation and trunk registration, so voice delivery is not dependent on a platform whose terms could be withdrawn. Software integration runs through official routes into named customer record platforms, productivity suites and messaging tools.

Nothing resembling scraping or unsanctioned automation appears anywhere. One structural qualification pulls the other way and is worth a buyer's attention: third party connectivity providers market services for organisations on this platform that need control over numbering, routing and caller identity, and regulated or high volume outbound operations are named as the typical case, which suggests the native carrier layer does not cover every compliance sensitive scenario.

The trust centre also records critical dependence as yes, the vendor's own assessment of how much a customer would rely on it. Ask what the native carrier layer covers for high volume outbound, and where a connectivity partner becomes necessary.

A purpose built governance shelf for artificial intelligence, held off the top band by access rather than by content. The trust centre carries a dedicated section with four distinct entries covering an overview, governance, and evaluation and red teaming, plus a standing policy on AI use and development held alongside the access and acceptable use policies.

Red teaming named as its own maintained artifact is rare in this index and is the entry that matters most for a product placing autonomous agents on live customer calls. The compliance list includes the European artificial intelligence regulation, and supply chain risk management and third party risk assessments are separately maintained, with the knowledge base showing buyers already asking for evidence of risk assessment on a critical model provider.

The subprocessor list names the model provider, so the supply chain is visible rather than described. What holds it below the top is that every substantive document requires an access request, no independently audited management standard for artificial intelligence is claimed, and no training position on customer call audio is stated openly. Ask for the governance and red teaming documentation and a written position on training with customer audio.

Autonomous agents speak to customers and no disclosure position is published. The agent product handles voice and text interactions end to end without a human in the loop, which means a person calling a business can hold a complete conversation with a synthetic voice, and nothing on the open surface states whether that person is told, whether disclosure is configurable by the operator, or what the agent says if asked directly whether it is a person.

The listed compliance coverage includes the European artificial intelligence regulation, whose transparency provisions bear directly on exactly this question, so the vendor has evidently considered it somewhere, and the detail is not readable without an access request. A second disclosure surface sits underneath: every call across every plan is transcribed and analysed, and consent to record varies by jurisdiction with several requiring all parties to agree.

Nothing published describes what announcement the platform makes or whether it is automatic. Ask whether the agent discloses that it is not human, whether that is configurable, and what recording consent announcement is applied by jurisdiction.

Integration across the categories a communications platform must reach, with a published interface layer at the telephony level as well as the application level. Named connections cover customer record platforms, two major productivity suites, messaging and support tools, and integration behaviour has its own dedicated security entry in the trust centre rather than being treated as an afterthought.

Below the application layer the vendor publishes session initiation interfaces and trunk registration, which lets connectivity partners and enterprises with their own carrier arrangements interoperate directly, and that is a deeper form of openness than a connector list. Status monitoring is maintained and a vulnerability reporting route is published. Two limits hold it below the top band.

Customer record integrations are gated above the entry tier, so the published integration list overstates what a buyer at fifteen dollars receives. And no developer documentation was reached on the routes taken this pass, so interface depth and rate limits are unverified. Ask which integrations are gated to which tier, and for the interface documentation.

The hosting provider is named and the residency answer is not published. The trust centre names the underlying cloud provider, maintains entries for status monitoring, denial of service protection, backup protection, a disaster recovery plan and tabletop exercises, and states a risk profile with data access restricted and impact substantial, which is a self assessment more candid than most vendors publish about themselves.

What is absent is the customer facing residency layer: no region named, no European or United Kingdom hosting election described, no tenancy model stated, and no recovery time or recovery point figures published. That gap is sharper for a voice platform than for most software, because call audio is subject to jurisdictional recording law and telecommunications regulation as well as data protection, and a buyer in a regulated European market needs to know where recordings are stored before they can lawfully deploy. Ask which regions call audio and transcripts are stored in, whether European recordings can be kept in region, and what the recovery objectives are.

A live trust centre of real depth, with a certification set that goes beyond the usual pair and every document behind a request. Published on the face are a service organisation control report, the information security standard in its 2022 revision, the cloud security and cloud privacy standards, payment card and health information frameworks, a national cybersecurity framework, a cloud application security assessment at its highest tier, a cloud alliance registry entry and a state government cloud authorisation at level two.

The last two are independently verifiable outside the vendor's own site, and the registry listing dates to 2019, which evidences duration rather than a recent push. Around them sit roughly twenty maintained categories covering application security, endpoint security, change management, asset disposal, physical security and training, plus a penetration test report, cyber insurance and a public vulnerability reporting route.

Held below the top band on the usual ceiling: no audit period, auditor or certificate is on the public face, and the reports require an approved request. One caution recorded rather than graded: an unclaimed third party directory presents this vendor's regime coverage as an eleven item certification list including data protection regulations, which are laws rather than certifications. The vendor's own centre separates them correctly. Ask for audit periods and the auditor.

Real published rates across several product lines, undone at the top by minimums and add ons rather than by silence. Published figures include a business phone system at fifteen and twenty five dollars per user monthly on annual billing, with monthly billing stated at twenty seven and thirty five, and a sales suite at thirty nine, ninety five and one hundred and fifty dollars per user monthly.

A one user minimum at the entry tier means a solo operator can buy without purchasing empty seats, which is a genuine accessibility choice. Core speech features are included at every tier rather than metered, so the headline rate covers the differentiating capability rather than excluding it. Three things keep it off the top band and they are the ones that decide the real bill.

The enterprise tier is quote only and carries a reported one hundred user minimum, which is a hard floor rather than a negotiation. Load bearing capabilities are gated upward, with customer record integrations and telephone support absent from the entry tier.

And a substantial add on menu sits outside the headline rate, with additional numbers, conference lines, larger meetings and international calling each separately priced, and an unlimited international package reported to carry a fifteen hundred minute cap per office. The autonomous agent product is quote only on a per conversation basis. Ask for the enterprise rate and its seat minimum, the per conversation agent price, and a quote including every add on your deployment needs.

Retention is maintained as a named topic and its terms are not readable. The trust centre carries dedicated entries for data retention, data management, data backup and secure asset disposal, so the questions are tracked as part of the security programme, but each requires an access request and no period, scope or deletion timeline appears on the open surface. Nothing states what a departing customer can export or in what format.

Two exposures are specific to this category and neither is addressed. Call recordings and transcripts accumulate as a business record that many customers are separately obliged to retain, and whether those leave in a usable form or remain in a platform the customer has stopped paying for is a compliance question rather than a convenience one.

And telephone numbers are the harder asset: porting numbers away is governed by carrier process rather than by the vendor's export tooling, and nothing published describes what that process involves or how long it takes. Ask what the retention schedule is, whether recordings and transcripts export in bulk, and what the number porting process and timeline look like on termination.

The vendor operates its own carrier infrastructure, which is the deliverability layer for a voice product, and publishes almost nothing about how it manages it. For telephony the equivalent of deliverability is call completion and caller reputation: whether calls connect, whether they arrive labelled as suspected spam, and what attestation the carrier applies. Those are determined by the platform rather than the customer, and no published statement addresses any of them.

Third party connectivity providers marketing to this platform's customers offer caller identity branding, spam label protection, short duration call filtering and traffic throttling, which indicates the questions are live for high volume users and that the native answer is not sufficient for all of them. On messaging, application to person registration governs whether text traffic is delivered at all, and no native registration process is described on the vendor surface. Ask what caller authentication attestation level the platform signs at, what spam label remediation is available, and whether messaging registration is handled natively.

Genuine range from a single user to large enterprise, with the product line structured to match rather than one platform stretched across both. A one user minimum at the entry tier means the smallest possible buyer is served, and the customer roster establishes the upper end with named organisations in media, telecommunications, staffing, retail and healthcare, several of them global.

The suite structure addresses distinct buyers with distinct products rather than tiers of one product: a phone system for general business use, a sales suite for outbound teams, a contact centre suite for support operations, and an agent product for automated handling. Regulated sector coverage is evidenced by instruments rather than assertion, with health information and payment card frameworks both listed and a state government cloud authorisation held.

Held below the top band by the gap between the range advertised and the range purchasable: the entry tier excludes customer record integration and telephone support, and the enterprise tier carries a reported hundred seat minimum, so the middle of the market has a narrower real choice than the four price points suggest. Ask what the smallest deployment is that includes integrations and support.