Mili

The models are the whole of it. Live comprehension of a spoken client conversation, structured note generation against a firm's own rubric, extraction of actions and financial detail, and an agentic layer that then carries work across other systems are all inference, and there is no workflow spine, data asset or calculation engine underneath any of it. Remove the models and nothing remains that a firm would pay for.

The architectural choice not to record is worth noting here precisely because it sharpens the point rather than softening it: without a stored recording there is not even a transcript archive to fall back on, so the model's understanding of the conversation in the moment is the entire product.

Autonomy expanded substantially in February 2026 and no oversight design was published alongside it. The meeting product was low risk by construction, producing notes a person reads. The agentic platform launched since is described as handling complete workflows end to end across a firm's systems with bidirectional writes into the client record, which is action rather than analysis.

Searched for an approval step before a write reaches a system of record, an administrator control over what the agent may do, a record distinguishing agent authored entries from staff authored ones, or any audit trail, and located none in published material. The assurance that human judgement stays in the loop appears in a recorded interview with the chief executive rather than in any product documentation, and an interview is not a control a compliance officer can point to.

Searched for a transcription accuracy figure, any measure of how completely a generated note captures a conversation, a benchmark, a validation method or a revalidation cadence, and located none. The published figures are all efficiency rather than accuracy, being a 90 percent reduction in administrative work and more than six hours recovered a week, which measure how much faster the work is done and nothing about whether it is done right.

The risk is heightened by the design rather than reduced by it: with no recording retained, a firm that later doubts a note has no source to check it against, so the model's comprehension in the moment is unverifiable after the fact.

Both routes to this grade are satisfied independently, which is unusual for a company of roughly thirty five people. The independent route: the 2026 edition of the largest adviser software survey placed this product top of its category on user rating at 8.69, measured across more than fourteen tracked competitors, which is a number produced by a third party polling the profession and not selectable by the vendor.

The named customer route: an employee owned registered investment adviser managing twelve and a half billion dollars is identified as having evaluated several competing tools before selecting this one, and the wealth arm of a four and a half billion dollar community bank is identified as having deployed it across its adviser team, both with their size stated. A wealth advisory firm also took part in the funding round, which is a customer putting capital behind the judgement.

The most sensitive artefact is never created, and the training question is untouched. Not recording the conversation removes the single largest stewardship exposure a meeting product has, and that is a real and unusual choice which is credited on the privacy row. It does not answer this axis.

Searched for any statement on whether generated notes, extracted financial detail, firm specific rubrics or the records written into client systems are retained, used to improve the service, or used to train, tune or evaluate any model, and located none in either direction. Whether one advisory firm's documented conversations can shape output another firm sees is precisely the kind of thing a company positioning itself on trust would benefit from stating.

An architectural commitment sits where most vendors offer a retention policy, and it is the stronger of the two. The product makes no recording of the client conversation and puts no participant into the call, so the audio artefact that a firm would otherwise have to store, secure, supervise and eventually produce is never created at all.

A commitment that something does not exist cannot be undermined by a retention setting, a misconfiguration or a subpoena, which is why this reads above the norm. Held below the top grade because it covers only the audio: notes, extracted financial detail and the records written into client systems are all created and stored, and searched for a retention schedule, a deletion route or published processing terms covering those, and located none.

Recorded as an unverified absence, with one construction worth flagging. Security is claimed repeatedly and always by adjective, the company describing itself as combining bank grade security with industry expertise and as the most secure solution in its category, and an adjective is standing exactly where an enumerated credential belongs. This is the fourth instance of that pattern found across recent builds.

Searched for a named attestation, an information security standard, an audit report or a dedicated trust page and located none, and no security page was reached. A vendor selling to a bank owned wealth business has passed that institution's review, so credentials likely exist. Queued check: a trust or security page.

A software supplier holds no licence and requires none, and no penalty attaches. The vendor engages with its customers' obligations more directly than most, describing itself as removing compliance risk at the architectural level and its product as a compliant digital worker, and both claims are about the customer's position rather than any supervised status of its own.

The unaddressed question follows from the agentic expansion: records written automatically into a firm's client system are examinable material, and nothing published describes how a firm demonstrates to an examiner which entries a person made and which the agent made.

Searched for any evaluation of comprehension across speaker groups, any fairness testing, or any account of how the agentic layer decides which workflows to run and in what order, and located none. Two exposures sit unaddressed. Live comprehension of speech varies with accent, speech pattern and background conditions, and a client understood less accurately gets a thinner advice record with no recording available to correct it later.

And a system that assembles and prioritises follow up work across a book is allocating a firm's service capacity, so a consistent tendency in what it raises directs attention toward some clients and away from others.

Searched for a warranty, an accuracy commitment, a correction obligation or any allocation of responsibility between vendor and advice firm, and located none. The exposure grew in February 2026 without any published term growing with it, because an agent that writes into a firm's client system and runs workflows end to end performs acts the firm owns entirely as far as its clients and its regulator are concerned.

The client whose conversation was interpreted has no described route to see the record made of them, and here that route could not exist even in principle, since there is no recording against which anyone could check what was actually said.

Searched the product pages, the funding announcement, the platform launch release and third party coverage for a model provider, a model family, a version, a speech recognition supplier or an inference host, and located none. The gap is more pointed for this vendor than for most because of what it claims: a company arguing that its architecture removes compliance risk is making a claim about the path a client's spoken words take, and a firm cannot evaluate that path without knowing which third parties sit on it. Not recording a conversation and not sending it anywhere are different commitments, and only the first is published.

Ten systems named individually across four categories, with the connections stated as bidirectional, which is specific where most vendors this size are vague. Six client relationship systems are named, covering both the general enterprise platform and the adviser specific packages an independent firm actually runs, alongside a financial planning platform, both dominant calendar services and a custodian.

The custodian connection is the one that signals seriousness, because it means the assistant can see account context rather than only conversation context. Held below the top grade because searched for a programmatic interface, any account of write depth such as custom fields or objects, and any statement of how contact matching or schema mapping works, and located none, so the breadth is established and the depth is not.

Recorded as an unverified absence. The product is delivered as a hosted service processing audio in real time from a device or platform, and nothing about hosting, region, tenancy or residency was located in the material opened. Searched for a storage location or residency commitment and located none. One inference a reader might draw is worth blocking: processing audio without recording it says nothing about where that processing happens or where the resulting notes are stored.

Searched the product material, the launch announcements and third party listings for a rate, a tier, a per seat figure or any account of whether the agentic platform is priced separately from the meeting product, and located none. That last question is the live one, because the company shipped a substantially larger product in February 2026 and a buyer cannot establish from anything published whether existing customers received it or are being sold it. Queued check: a pricing page.

Two professional buyers and a genuine spread of firm size within them. The product is sold to wealth advice firms and to tax planners, and the named customer base runs from ordinary advisory teams up to an employee owned registered investment adviser managing twelve and a half billion dollars and the wealth arm of a community bank managing four and a half billion, which are different organisations with different supervision requirements. Reported reach is hundreds of advisory teams.

Held below the top grade because the market is the United States, with the qualifier beyond it left undefined, and because no bank, insurer or asset manager is addressed as a buyer in its own right, the community bank appearing as the owner of an advice business rather than as a banking customer.