Recordsure

The Hadrius argument applies exactly and it is the cleanest test in this lane. Strip the models and what remains is a recording application and a document repository, which is an archive, and an archive does not discharge a supervision obligation because the rules require review rather than retention.

Everything that makes this a compliance product is model work: transcription of face to face and remote conversations captured in live environments, classification of those conversations into structured review themes aligned to quality assurance checklists, and analysis of adviser files for suitability and ongoing service checks. The company describes proprietary artificial intelligence built from 2012 and models trained by specialist teams across millions of data points and thousands of cases.

This is the strongest autonomy position recorded in the index and it is stated as a property of the technology rather than as a policy. The company publishes that its platform limits the ability for any ultimate customer outcome or advice suitability decision to be made by artificial intelligence, directly or through user intervention, and the second half of that sentence is what makes it exceptional: it closes the loophole every other constraint in this index leaves open, where a user can simply accept the machine's recommendation and call it a human decision.

The positioning throughout is consistent, with the machine doing precision work and human judgement making the call, and the product designed to identify which files deserve attention rather than to conclude on them. Where Federato names what is automated and audits it, Recordsure names what may never be automated and builds the prohibition in.

Three things earn this and none of them is an accuracy figure. The models are described as thirteen years mature and purpose built for the domain rather than adapted from general purpose tooling. The deliberate use of predictive over generative architecture is itself a model risk argument, because a deterministic classifier can be validated and reproduced in a way a generative system cannot, and the company makes that argument explicitly rather than leaving it implied.

And the output is stated to have been examined by a regulator and in litigation, which is external scrutiny of model output under conditions designed to find fault. What is absent is the conventional package: no classification accuracy, no false negative rate on flagged issues, no validation documentation, and for a supervision product the unmeasured direction is the conversation that should have been flagged and was not.

Per deployment outcomes are quantified in the units a compliance function actually manages. At a major United Kingdom retail bank running roughly 5,000 mortgage advice conversations a month, quality assurance coverage went from under 5 percent of sales to 100 percent, review times fell by more than half, and the bank reports improved advice standards and fewer complaints, with roughly 60,000 conversations now analysed annually.

Platform scale is stated separately at more than 90 million documents processed, over 200,000 ingested per day, average handling time cut by more than 85 percent, and 4 billion pounds of service fees checked for whether the service was actually delivered. Customers are described as tier one financial institutions and central government.

One claim stands apart and is unmatched in this index: the technology is stated to be proven in practice with the regulator and in United Kingdom courts, which is adversarial testing of a kind no other vendor here asserts. Institutions are described rather than named.

The architectural choice is deliberate, reasoned and published, which is rare. The company uses predictive rather than generative models for compliance critical work and argues the distinction explicitly, that generative artificial intelligence is accessible but is not always the right tool where a decision must be defensible, so the platform is designed around determinism rather than fluency.

That forecloses the hallucination surface by construction rather than guarding against it, and it is a harder commercial choice to make than adopting the prevailing technology. Models are described as trained by the company's own specialist teams on its accumulated case corpus. Held at B because no cross customer boundary is stated, and nothing says whether conversations captured at one institution inform models serving another.

Two named commitments carry this and the payload makes them necessary. The company states its practices and technology are fully compliant with the European data protection regime and certified to the international information security standard, and describes secure automatic upload from the capture application to its portal.

That matters because the material is audio of clients discussing their finances, health, family circumstances and retirement plans in advice meetings, which is among the most intimate payloads any vendor in this index handles, and it is captured in person rather than over an already monitored channel.

Held at B because no retention schedule, subprocessor list or deletion commitment was located, and nothing states how long a recording of a client conversation persists or what the client is told about it.

The international information security standard is named as held rather than gestured at, alongside a stated commitment to the European data protection regime and to building and operating to the highest security standards. One named framework verifiable by a buyer puts this ahead of most of the index, including both incumbent financial crime vendors, neither of which publishes an attestation.

What holds it below the top grade is that only one framework appears, no trust centre or compliance portal exists, and no detail is published on encryption, key management or how recordings are segregated between client firms.

The regulatory grounding is unusual in kind rather than in specificity. The company states its technology meets the requirements of regulators in two jurisdictions and has been proven in practice with the regulator and in the courts, which is external testing of the output under adversarial conditions and is the strongest form of validation available to an evidence product.

It also operates alongside a compliance consultancy with more than two decades in the field, so regulatory interpretation sits inside the business rather than being bought in. What holds it below the top grade is that no individual instrument is named: the conduct rules governing suitability, the consumer outcomes regime and the record keeping obligations are all implicit in the product and none is identified, and the court claim carries no detail that a buyer could verify.

The speech recognition variance finding applies and this is its eighth recorded instance, after Glia, Rulebase, Aveni, Reality Defender, Theta Lake, interface-ai and Featurespace. Transcription and classification accuracy varies with accent, dialect, regional speech and speech impairment, and here the subject scored is the adviser, whose conversations feed quality assurance findings and performance management, so an adviser can be flagged more often for how they speak rather than for what they advised.

Face to face capture in live environments adds acoustic variation on top. The mitigation is real and structural rather than rhetorical, since the platform prevents the machine from concluding on suitability, so a misclassified conversation routes to a reviewer rather than producing an outcome. No per accent or per demographic accuracy is published, and no analysis of how classification performs across adviser or customer populations was located.

The autonomy constraint doubles as the accountability mechanism, which is an unusual and effective design. Because the platform prevents the machine from reaching a suitability or customer outcome decision, responsibility for every consequential judgement stays with an identifiable person at the regulated firm, and the audit trail the system produces is described as authoritative and has reportedly been relied on in litigation. That is a clearer allocation than most vendors here offer.

The affected party also benefits by design rather than incidentally, since the whole purpose is identifying customer harm before it escalates, and the fee checking work has surfaced billions in service charges where the service may not have been delivered. What is missing is a route for an adviser to contest a classification that shaped their performance record.

The chain is short by deliberate design and the design is explained. Models are proprietary and built in house from 2012, trained on the company's own case corpus by its own specialists, and the choice of predictive over generative architecture means no external foundation model provider sits in the decision path at all, which closes the question that dominates this axis elsewhere. Named practice management integrations disclose where the documents come from.

What is not published is the remainder: no hosting arrangement, no subprocessor list, and no statement of whether any third party service touches audio during transcription, which is the one point where an external dependency would most plausibly exist.

Integration is aimed precisely at where this buyer's files actually live. Documents and transcripts are ingested automatically from named adviser practice management systems and from general document storage, with one of those integrations described as off the shelf, which matters because the wealth advice market runs on a small number of these platforms and connecting to them is the difference between a pilot and a deployment.

The capture application runs on both major desktop and mobile platforms for in person and voice over internet conversations, uploading automatically to the portal. What is not evidenced is the wider estate: no core banking system, customer relationship platform or case management tool is named, and no developer documentation or interface reference was located.

No hosting provider, region selection, residency commitment or private deployment option was located. The certification and data protection commitments imply arrangements consistent with European requirements, and the customer base is concentrated in one jurisdiction, but an implication is not a residency statement.

The question has weight given the payload, since recordings of client advice meetings are among the most sensitive records a firm holds and a regulated institution will want the location written down rather than inferred, particularly with an Australian footprint stated alongside the domestic one.

No rate card is published, but a defined and scoped evaluation route is, which is more than most of this index offers. The company has run an open offer to analyse up to 100 ongoing advice review files for a limited number of wealth management firms, with the participant committing access to their practice management system, so a prospective buyer can see the product work on their own files before contracting. That is a real access mechanism of the kind that earns Akur8 its grade. Nothing indicates the basis of charge for a production deployment, whether per file, per adviser, per conversation or per platform.

Within its chosen domain the coverage is deep, spanning tier one retail and commercial banks, wealth managers and advice firms, and central government, across mortgage advice, suitability assessment, ongoing service reviews and remediation exercises, with both face to face and remote channels handled. Regulatory fit is claimed for two jurisdictions.

The limits are deliberate rather than accidental: this is retail conduct and advice oversight, so it does not reach financial crime, trading supervision or institutional business, and the evidenced footprint is the United Kingdom with Australia stated alongside it. A specialist rather than a platform.