Aveni

Conduct testing, advice file review and suitability drafting are all model work, built on the FinLLM family and more than 150 specialist risk models. Nothing operable would remain once they were removed.

Agents handle structure and detail while the adviser applies judgment and reviews the result before anything reaches a client. Detect reviews every conversation and File Check every advice file, replacing sampled quality assurance, while Approve and Assure certify consumer facing AI before launch and oversee it live.

FinLLM's training on a 100 billion token UK financial services corpus labeled by compliance experts gives provenance, which is not evidence of performance. Accuracy and error rates for risk and vulnerability detection, the evaluation method and model documentation are unpublished, though the outputs feed quality assurance records a supervisor may test.

Prosser Knowles reports 60 percent less admin, Key Group six times faster quality assurance across more than 55,000 customers, and Octopus Money moved from 15 to 100 percent review coverage. The story summaries give no method or baseline period.

Customer and call data never trains the firm's models or reaches a third party provider for training, enforced as a technical control, and each customer sits in an isolated AWS production project. Retention terms for what external model providers receive are not stated.

Vulnerability detection infers ill health, cognitive decline and bereavement from recorded conversations, which engages special category handling. Its privacy and security page reaches the product, purging deleted data after 30 days and honoring manual deletion at once under UK and EU GDPR, but special category handling and a subprocessor list are not described.

Organization wide ISO 27001 certification is stated, with AES 256 at rest and regularly rotated root keys. Certificate, certifying body and audit period are not published.

The products are built against the Financial Conduct Authority's Consumer Duty, vulnerable customer expectations and suitability rules, along with the lines of defense supervisors expect. No formal admission backs it.

Speech based coaching on clarity and pace carries accent and dialect variance, and vulnerability inference errs both ways, missing a customer who needed support or labeling one who did not. Demographic accuracy and false classification rates are unpublished.

An adviser reviews generated documentation before it reaches a client, so the accountable person catches a suitability error. Neither an adviser scored unfavorably nor a customer wrongly assessed as vulnerable has a published route, and no accuracy guarantee binds the vendor.

Its own FinLLM model family runs alongside OpenAI, Anthropic and Google models used selectively. Which functions call which external model, and on what terms, is not said.

Iress lists Aveni Assist among its Xplan integrations, and Intelliflo is an announced strategic partner, so suitability reports and fact finds land in the adviser back office. That back office is the system of record for UK advice firms.

AWS hosts customer data with server side encryption, but no region, residency commitment or transfer mechanism is stated. Recorded client conversations and suitability files carry retention duties, so their location matters to a supervisor.

Engagement runs through a demo request, and nothing is priced for buyers from small practices to large banks. Pricing by adviser, interaction or file is equally possible.

Banks, wealth managers and advice networks are covered with role specific material for advisers, quality assurance assessors and compliance teams, and Lloyds Banking Group, Nationwide and Quilter are named users. One regulator's rules shape the product, with nothing for other jurisdictions.