How Aveni Assures Customer Outcomes Across the AI Lifecycle (The Demo Room #26)
Welcome to The Demo Room – your front-row seat to the future of RegTech, RiskTech, and AI innovation.
In this series, we document our research interviews with the most forward-thinking vendors tackling the industry's biggest challenges. Each blog is built around a comprehensive product demo, providing clear insights into how these innovations address industry challenges.
On this occasion, we feature Aveni, a provider of AI-powered oversight and assurance for customer interactions in regulated financial services.
Customer-facing AI is moving from answering simple questions to influencing decisions and completing tasks. This shift is changing the assurance problem.
Conventional model testing can tell a firm whether an AI system is accurate on a benchmark or resistant to a defined technical failure. It cannot, on its own, show whether the system treated a vulnerable customer appropriately, gave a suitable response in context, or contributed to a fair outcome across a complete interaction.
Our AI in Risk & Compliance 2026 research finds that 65.6% of reported AI activity has reached production or beyond, while 69.3% of institutions say deployment is ahead of the measured control environment. Firms now need evidence that customer-facing AI continues to behave appropriately before, during and after release.
Aveni is building to assure AI at the level that matters the most: customer outcomes.
The Problem for Firms
Most regulated firms already have model risk, quality assurance and three-lines-of-defence processes. The difficulty is that these processes were not designed for an AI agent that can hold a long, dynamic conversation with a customer.
Pre-deployment tests may miss unusual combinations of persona, language and circumstance. Once a system is live, human reviewers cannot watch every conversational turn without recreating the cost and delay that automation was intended to remove. Post-interaction quality assurance is still frequently sample based, which means rare but serious conduct, vulnerability or complaint risks may go unseen.
There is also a practical trade-off between depth, latency and cost. A frontier model can be powerful, but applying the same heavyweight analysis to every turn may be too slow and expensive for a natural conversation. Firms need a proportionate way to identify emerging risk quickly, reserve deeper analysis for the cases that warrant it, and preserve evidence for human review.
In our research interview, Aveni described this as an operating-model question. Continuous AI oversight has to fit established risk and compliance accountability.
A Solution: Aveni's AI Assurance Lifecycle
Aveni's compliance suite covers three connected stages: Approve before deployment, Assure during live interactions and Detect after the event.
1. Approve: continual behavioural testing before and after release
Approve connects to a customer-facing AI through an API. Firms select representative customer personas, the risks they want to test and the scale and depth of the simulation. The system generates dialogues and assesses the resulting interactions against criteria such as vulnerability, advice boundaries, harmful language, understanding and escalation.
The output is intended to give model and risk committees evidence for a release decision, including the cases that passed, those that did not and the risks involved. The tests cover customer situations and outcomes as well as the model's technical behaviour. This allow teams to make decisions on improvements and approval to production, both for first release and for each model change or adjustment.
2. Assure: monitor the live conversation
Assure evaluates customer-facing AI turn by turn through their Conduct Risk real-time guardrail. Aveni describes a tiered architecture in which fast, task-specific models carry out an initial check within the latency budget. Potentially higher-risk interactions receive deeper analysis, with configurable responses that can allow the conversation to continue, alert a human, redact content or stop the interaction.
The system keeps the control proportionate by using bounded classifiers for fast detection and reserving more intensive reasoning for escalated cases.
3. Detect: assess the complete outcome
Detect performs post-interaction quality assurance as well as outcome testing across voice, chat, email, documents and related records. It applies the firm's own QA and Consumer Duty framework to the complete case, rather than reviewing a small sample or an isolated transcript. Findings retain supporting evidence for review, reporting and remediation. This allows scaled coverage without adding headcount, and can support upselling or early indications of customer churn.
Aveni says one QA use case reduced assessment time from 90 minutes to 15 minutes. It also reports that a specialist model layer achieved a 61% faster time to first token, a +3% improvement in F1 accuracy and 92% lower running cost than Gemini Flash 2.5 Lite in the associated benchmarked use cases. These are vendor-supplied results and should be read in the context of the specific tests and deployments from which they were derived.
“The bank that solves the oversight problem will be the bank that wins the race to become the most agentic bank.” – Kent Mackenzie, Chief Commercial Officer, Aveni
Parker & Lawrence’s View
Today's AI models are already so complex that deep technical evaluation can feel several steps removed from the question that ultimately matters: what happened to the customer. Technical evaluations remain essential, and researchers and model specialists must continue to challenge how these systems work. But they do not replace evidence about suitability, vulnerability, fairness and the quality of the final outcome. Aveni is focusing on these outcomes.
And they are doing so with unusual technical depth. Aveni has trained its own small language models on financial-services data and deployed them within major financial institutions. Its architecture orchestrates more than 150 models, from low-latency classifiers to specialist FinLLMs and selectively used frontier models. Each task is routed according to the subject matter, depth of reasoning, speed and cost it requires.
This architecture makes full-population testing practical. Fast models can assess every interaction, with deeper analysis reserved for cases that warrant it, while the firm retains evidence at customer-outcome level. The approach uses specialist models for a clear operational purpose, rather than treating proprietary AI as an end in itself.
For financial institutions, this combination could remove a significant obstacle to adoption. Product teams can bring AI to market faster and with greater confidence when they can test the complete population before and after release, monitor live behaviour and show risk and model committees what happened in each case. Aveni's contribution is to keep advanced model engineering tied to the human outcomes that regulators, firms and customers care about. And we’re very bullish on their prospects.
