Silvia Outperforms Frontier AI Models in Personal Finance

Over the past six months, Silvia’s engineering team has been focused on outperforming frontier AI models in personal finance topics and owning our own intelligence.

Today, we are announcing that we have officially achieved our goals and now outperform every frontier model on important personal finance topics, including tax, mortgage, and credit cards. This makes Silvia the most powerful and accurate product for investors who want to use AI technology to manage their finances.

We accomplished this historic milestone by rebuilding the Silvia technology stack to ensure we owned our own intelligence. The core pieces of our proprietary technology include:

1. Data pipelines: We built our own benchmarks for tax, mortgage, and credit cards, graded on accuracy and on citing primary sources. We're #1 on all three, and we open-source them on Hugging Face so anyone can check our work. We also generate high-quality training data as part of this work.
2. Harness engineering: The harness is everything around the models (the model router, tools, and agents). We tune it together so they all work as one system.
3. Model training: We trained Silvia Analyst, our own quantitative finance specialist. It handles the heavy number-crunching inside Silvia (market data, company financials, portfolio analysis, and document parsing) at 36x lower cost than frontier models.

In layman terms, we have built a custom AI tech stack that is designed to serve Silvia users, rather than use off-the-shelf, generic AI tools and technology. We are not trying to be the best at physics, philosophy, healthcare, or the numerous other verticals that frontier models have to compete on. We are laser-focused on merely being the best at personal finance.

This strategy of owning our own intelligence has led to impressive results.

On our tax benchmark, Silvia scored highest of seven AI products on both accuracy and citing primary sources. The benchmark is public on Hugging Face. Since July, when the new model began rolling out, median response time has dropped 48%, cost per question has dropped 59%, and the share of answers users rate negatively has fallen by more than half, while monthly active users grew substantially since March.

How reinforcement learning trained Silvia for personal finance

We also use reinforcement learning to train Silvia. In reinforcement learning, an AI system practices inside simulated environments, receives feedback on its performance, and improves through repeated trial and error.

For Silvia, we built reinforcement-learning environments that simulate difficult tax, mortgage, and credit-card scenarios. These environments let Silvia practice applying financial rules, work through multi-step questions, learn from the result, and improve before those capabilities reach users.

We also publish the benchmark datasets used to evaluate performance in each domain. The repositories make the questions and evaluation data available for inspection, so readers can examine the evidence behind the results instead of relying only on a headline score.

• View the public Silvia tax benchmark
• View the public Silvia credit-card benchmark
• View the public Silvia mortgage benchmark

Publishing these datasets makes the results testable and gives us a repeatable way to measure future training runs against the same domain-specific problems.

We strongly believe that every company will eventually own their own intelligence. Not only will data sovereignty become a hot topic, but the superior performance and lower costs create an incentive that is hard to ignore.

Similar to bitcoin’s “not your keys, not your coins” ethos, the AI industry is about to learn “not your model weights, not your intelligence.” For example, because we own our own intelligence, it allows us to create a better Silvia product and user experience. We can do certain things that were previously cost-prohibitive, or we can create personalized experiences that benefit from our increased control throughout the tech stack.

We recruited world-class AI engineers and have transitioned to a leading AI research lab focused on finance. Our product is now built on custom infrastructure that is better for users and our company.

We plan to continue pushing the pace of innovation across many aspects of AI. If you are a finance organization and are interested in owning your own intelligence, please reach out here and we are happy to help.

Separate Answer Quality From Evidence and Control

A fluent financial answer can still be incomplete, outdated or wrong. Evaluate accuracy separately from grounding. Accuracy asks whether the conclusion is correct. Grounding asks whether the answer identifies current, relevant sources and whether those sources actually support the claim being made.

The system should also expose its limits. Check whether it can distinguish facts from assumptions, handle account-specific data without inventing missing details, and show the calculations behind an output. For consequential decisions, verify dates, jurisdiction and source material before relying on the result.

Data controls matter as much as answer quality. Review connection permissions, encryption, retention, export and deletion options, human access and what happens when an account is disconnected. Good financial software should make both the analysis and the handling of sensitive data understandable.

Questions to work through

• Ask for the source, effective date and assumptions behind important claims.
• Recalculate material figures independently before acting.
• Review permissions and data controls before connecting financial accounts.
• Use qualified professional judgment when tax, legal or fiduciary duties apply.

What the benchmark results mean for Silvia users

The benchmark results matter because they measure the parts of financial AI that users need to be able to trust: factual accuracy and grounding in relevant sources. In practice, Ask Silvia pairs that finance-specific intelligence with the accounts, assets, liabilities and goals you choose to connect, so an answer can reflect your broader financial picture instead of treating the question as an isolated prompt. The result is a more useful starting point for tax, mortgage, credit-card and investment questions, while important decisions still remain with you and, when appropriate, a qualified professional.

Related Silvia guides

• Silvia vs. Evergreen.ai: Features, AI, Advice and Best Fit
• Silvia Beats Every Major AI Product on Credit Cards
• CFO Silvia vs. Origin: Which AI Financial Platform Fits You?