Airspeed

AI native by construction and by origin. Transcription, speaker separation, signal extraction against named qualification methodologies, risk detection and the agents that act on all of it are the product; there is no non AI remainder to fall back on. The company was founded by two former Google DeepMind research scientists and its published positioning is explicitly that the previous generation of tools surfaced insight while this one executes work. Remove the models and nothing sellable remains.

The gap sits exactly where the product is boldest. The platform is sold on agents that update systems, send follow ups and flag risks without a human trigger, and customers are reported to have built thousands of custom agents on it, yet nothing published describes the oversight model around any of that.

No statement of what an agent may send unreviewed versus what requires approval, no description of what a misconfigured custom agent can reach, and nothing on what stops a run once it is live against real customer records and real inboxes. Autonomy is the pitch and containment is unaddressed.

No model or provider is named anywhere public, no version or update policy is stated, and nothing explains how the platform derives a close probability or a risk flag from a conversation, although the company publishes editorial arguing that such scores need reasoning behind them. Transcription accuracy is disclosed with its scope, which is unusually honest, but that measures output quality rather than explaining the system. For a product whose extracted signals feed forecasting, the derivation is the transparency question and it is unanswered.

Better than most of this cohort, and short of the top band on basis. A named customer figure is published with a timeframe attached: an enterprise software customer reporting more than 193,000 dollars saved within 90 days and roughly six hours reclaimed per sales representative per week.

Product accuracy is quoted with its scope stated rather than as a bare number, at over 95 percent transcription accuracy on standard sales calls in English, with most calls fully processed within five minutes. Usage figures are specific: revenue up fourfold year on year, monthly run volume close to tripling between January and April 2026. What is missing is the method behind the headline saving, and the wider claims arrive through funding coverage rather than published measurement.

Recording consent is handled well and is graded on the recipient axis. What is not addressed is the sending side, which the product now touches directly: agents draft and send follow up email on a representative's behalf, and nothing published covers opt out handling, suppression, or how the obligations attaching to automated commercial email are met when an agent rather than a person composes and sends it. No public enforcement action located.

A real posture, published in product documentation rather than buried in a policy: GDPR compliance claimed, encryption in transit and at rest, single sign on and SAML, data residency options offered, retention configurable from 30 days to seven years, and an explicit statement that customers retain full ownership of their data and may request deletion at any time. Configurable retention on recorded conversations is a meaningful control that many peers do not offer. Short of the top band because no data processing agreement, subprocessor list or data subject rights process was located, and lawful basis for recording third parties rests with the customer.

Unambiguous by construction. The inputs are the customer's own calls, emails, support tickets and CRM records, and no third party contact or intent data is sold or blended in. The company states that customers retain ownership of that material. Short of the top band because the flow of conversation data through processing and any sub processors is not documented, which for recorded third party speech is the licensing question that would matter in a dispute.

Low exposure and the method is disclosed, which is what separates this from the silent end of the category. The platform joins meetings as a named participant rather than capturing audio covertly, and reaches CRM, email and ticketing systems through their own interfaces. Recording bots do sit within the terms of the meeting platforms they join, and the company states no position on those terms, so the grade rests on a disclosed method rather than a stated conformance commitment.

This is the question the category turns on and it is unanswered. Nothing published states whether recorded conversations, extracted signals or the custom agents customers build inform models or capabilities that serve other customers, and no opt out or contractual exclusion is documented. Ownership and deletion are addressed, retention is configurable, and neither of those answers the cross client boundary. A buyer whose competitors are also customers cannot tell from public material whether their pipeline conversations improve the instance those competitors use.

One of the strongest positions in the index and it is published as a direct answer to a direct question. The company states that it does not record without consent, that its recorder joins calls as a named participant and announces its presence, that consent handling follows the customer's own policy with participants able to opt out, and that consent first recording workflows are supported for compliance sensitive industries.

Disclosure by default to the people being recorded is exactly what this axis asks for. It falls short of the top band because the obligations in force since 2 August 2026 under Article 50 of the EU AI Act are not addressed by name, and nothing states whether agent drafted follow up email reaching a prospect identifies itself as machine generated.

Bidirectional by design and central to the value proposition: the platform positions itself as the layer between the conversation and the system of record, pushing and pulling across CRM, email, calendar and support tickets, with structured deal data written back rather than transcripts attached. Named CRM coverage spans the two major platforms. Depth thins in public material, with no API reference, object mapping or sync behaviour documented before purchase.

Multi tenant cloud with residency options stated to be available, which for a company at this stage is more than most offer and is the right answer for a product holding recorded conversations from customers across twenty countries. Short of the top band because the regions themselves are not enumerated, so a European buyer knows an option exists without knowing what it is until they ask.

SOC 2 Type II certification is claimed and stated consistently across product documentation, alongside encryption in transit and at rest and single sign on with SAML. That is a credible posture for a Series A company holding sensitive recorded material. Short of the top band on the usual two points: no trust centre or document portal was located, and no audit date, report recency or penetration test attestation is visible to an outsider.

No published pricing was located. There is no pricing page, no tier structure, no unit of pricing and no indication of whether the platform is charged per seat, per recorded hour or per agent action, and the platform pages route pricing questions to a sales conversation or a demo booking. A buyer cannot form an order of magnitude estimate, which is a notable omission for a product aimed at mid market teams who are the segment most likely to self qualify on price.

Two of the pieces are present and published: customers are stated to retain full ownership of their data and can request deletion at any time, and retention is configurable from 30 days to seven years, which gives a customer real control over what accumulates. The rest is missing.

No documented export path for recordings, transcripts, extracted signals or the custom agents a customer has built, no post termination data rights, no deletion timeline, and no public commercial terms a prospect can read before signing.

The product now sends on a representative's behalf, with agents generating and dispatching follow ups automatically, and nothing published addresses the sending craft that comes with that: which mailbox is used, what volume governance applies when agents fire across a whole book of accounts, or what happens to sender reputation if agent generated follow ups draw complaints. Deliverability is not claimed and not addressed.

The segment is stated plainly as mid market business to business sales teams and the evidence supports it: named customers including a document experience platform, a vector database company and a pricing software vendor, roughly 180 to 200 customers across 20 countries, and operations split between London and New York. The gap is language. Transcription accuracy is published for English and nothing states what coverage looks like in the other markets the customer count implies, which matters more for a speech product than for most.