AI Transparency

Understanding how artificial intelligence works in Nova Net Worth

Last updated: May 16, 2026 • Version 2.1

Our Commitment to Transparency

At Nova Net Worth, we use artificial intelligence (AI) and data-driven analysis to help you make better financial decisions. We believe you have a right to understand how AI affects your experience, and we're committed to being transparent about our technology.

This page explains what AI we use, how it works, and how you can control it.

TL;DR: We use AI to help you make better financial decisions. Your data is never sold or used to train external models. You control all AI features.

How We Use AI

1. Spending Forecasting

What it does: Predicts your future spending in each category based on your historical patterns.

Technology: LSTM Neural Network (Deep Learning)

Accuracy: 87% accurate on average

How it helps: Helps you plan budgets and avoid overspending

Your control: You can override any prediction manually

2. Anomaly Detection

What it does: Identifies unusual transactions that might indicate fraud or errors.

Technology: Statistical analysis + behavioral profiling

Accuracy: 92% detection rate, 3% false positives

How it helps: Protects you from unauthorized transactions

Your control: You can mark false alarms, improving accuracy

3. Financial Recommendations

What it does: Suggests ways to save money, reduce debt, and reach your goals faster.

Technology: Rule-based AI + personalization engine

Effectiveness: Average $2,400/year in identified savings

How it helps: Provides actionable financial advice

Your control: All recommendations are optional—you decide

4. Charlie (AI Financial Coach)

What it does: Answers your financial questions using natural language.

Technology: Multi-provider LLM (OpenAI, Anthropic, Google) + your financial data, wrapped in a 14-layer guardrail pipeline

Privacy: Your data is never used to train external models

How it helps: Provides personalized financial education 24/7

Your control: You can disable Charlie anytime in settings

14-Layer Guardrail Pipeline

Every Charlie response passes through fourteen independent checks before it reaches you. No single model is trusted to be correct on its own — each layer can block, redact, or downgrade output that fails its specific test.

#	Layer	What it checks
1	PII redaction	Strips SSN, account numbers, raw tokens from prompts before they leave Nova.
2	Prompt-injection guard	Detects adversarial instructions hidden in user data (transactions, notes, account names).
3	DLP scanner	Flags data classes (credit cards, balances over a threshold) that shouldn’t be sent to the provider.
4	Pre-check scorer	LLM-side risk score on the proposed prompt before the main call runs.
5	Compliance rules engine	Hard rules: no investment recommendations, no tax advice, no “guaranteed” outcomes.
6	Semantic risk scorer	Embedding-distance check against known-bad output patterns.
7	Hallucination guard	Cross-checks every numeric claim against the user’s actual financial facts.
8	LLM judge	A second model independently scores the first model’s output for accuracy and tone.
9	Human review sampling	A bounded sample of responses is queued for human review and feeds the eval set.
10	Sandboxed executor	Tool calls (e.g. fetch transactions) run with row-level security; Charlie can’t see other users’ data.
11	AI audit log	Every prompt, response, layer verdict, and tool call is persisted for 3+ years.
12	Financial facts grounding	Charlie answers from a structured facts table derived from your data — not from “model memory.”
13	Policy registry	Versioned policy documents (privacy, retention, AB-2013) are pinned per response.
14	Model-independence assertion	The pipeline must produce equivalent verdicts on a held-out eval set across providers; if it can’t, we don’t ship the model.

Why fourteen? Each layer catches a different failure mode — a leak, a hallucinated number, an off-policy recommendation, a regression after a model swap. We don’t collapse them because the cost of one false-confident answer in someone’s financial life is higher than the cost of running a few more checks.

Charlie Eval Suite

How we know the pipeline actually works — and how we keep it working as the underlying models change.

100

Adversarial cases

A held-out test set covering prompt injection, hallucinated numbers, off-policy advice, jailbreaks, and edge-case account data.

14/14

Layer-by-layer verdicts

Every eval case logs the verdict of all fourteen layers, so we can see which guardrail caught which failure — not just the final pass/fail.

Pre-ship

Gating, not reporting

A model swap or pipeline change that regresses on any high-severity eval case blocks deploy. The eval is a release gate, not a dashboard.

What we publish: When we ship a new model or pipeline change, we record the eval delta in our release notes. We don’t hide regressions; we either fix them or we don’t ship.

Your Rights & Controls

Right to Explanation

You can see why AI made a specific prediction or recommendation.

Right to Override

You can manually adjust or ignore any AI prediction.

Right to Opt-Out

You can disable AI features individually or entirely.

Right to Feedback

You can report errors, improving the AI for everyone.

Right to Deletion

Deleting your account removes all your AI data.

Right to Human Review

You can escalate AI decisions to our team.

To exercise these rights: Visit Settings → AI Preferences

Privacy & Security

Your Data Stays Private

AI models are trained on YOUR data only—not shared with other users. We never sell your data.

No Sensitive Info in Models

We don't use names, SSNs, or account numbers in AI training—only transaction patterns.

External AI Partners

For some features (Charlie AI, receipt scanning), we use OpenAI's services:

Data sent to OpenAI is encrypted in transit
OpenAI does NOT use your data to train their public models
Data is retained for 30 days maximum (OpenAI policy)
We have a Data Processing Agreement with OpenAI (GDPR compliant)

Encrypted & Secure

All AI data is encrypted at rest and in transit. Access is logged and audited.

Limitations & Risks

We believe in honest communication about AI capabilities and limitations:

AI is not perfect: Our spending forecasts are 87% accurate—meaning 13% of predictions may be off.

Cannot predict life changes:AI doesn't know about upcoming job changes, pregnancies, or major purchases unless you tell it.

Requires data: AI needs 3-12 months of transaction history to work well. New users see statistical estimates instead.

Not financial advice: AI recommendations are educational, not professional financial advice. Consult a certified advisor for personalized guidance.

Regulatory Compliance

Our AI systems comply with international regulations:

EU AI Act: Designed to align with High-Risk AI System requirements including transparency and human oversight

GDPR: Right to explanation, deletion, and access fully implemented

UK FCA Consumer Duty: Designed with fair outcomes, transparency, and consumer support principles

CCPA: California privacy rights respected

For regulators: Full technical documentation available upon request at compliance@novanetworth.com

AI Training Data Disclosure

In accordance with California AB 2013 (Generative AI Training Data Transparency Act), we disclose the following about the AI models used in Nova Net Worth:

Models Used

Charlie is model-independent. Charlie's reasoning runs on a tier of frontier LLMs (currently OpenAI, Anthropic, and Google), routed through a provider-agnostic adapter so we can swap models without re-deriving the guardrail pipeline. We also use data-driven analysis for cash flow trends, category prediction, and anomaly detection — those run separately from Charlie.

User Data & Training

We do not use your personal financial data to train AI language models. Charlie runs on base frontier models from third-party providers, and Nova does not train custom LLMs on your data. Conversations sent via API are not used by those providers for model training under their enterprise data policies (e.g., OpenAI enterprise privacy, Anthropic API privacy).

Nova's ML Models

Our proprietary spending forecasting and category prediction models are trained on aggregated, anonymized transaction patterns. Individual users can opt out of contributing to model improvements in their privacy settings. These models do not generate text or provide financial advice — they power numerical predictions only (e.g., spending forecasts, category assignments).

Data Retention

AI conversation logs are retained for a minimum of 3 years for compliance and audit purposes. You can request deletion of your conversation history by contacting privacy@novanetworth.com, subject to applicable legal retention requirements.

Last updated: May 16, 2026. This disclosure is reviewed and updated quarterly.

Questions or Concerns?

We're here to help you understand our AI systems. Contact us: