The Aiera Lift: Measuring the Value of Financial Data Across Foundation Models

Large language models are increasingly becoming the reasoning layer for financial research. But even the strongest models face a fundamental limitation: they can only answer from what it already knows or can retrieve.

That creates a serious gap for institutional financial research.

Training data becomes outdated. Public search does not provide access to proprietary sell-side research, expert interviews, or many of the point-in-time details professional analysts rely on. A model may be highly capable, yet still lack the information required to answer the question.

The Aiera Lift study was designed to measure what happens when that changes. We tested the same analyst-style questions under two conditions:

Baseline: The model using its existing knowledge and available web search

Aiera MCP: The same model connected to Aiera's professional financial data through the Model Context Protocol

The difference in answer quality between those two conditions is what we call the Aiera Lift.

The Headline Result: 2.4× Better Answers

Across the models that successfully used Aiera's tools, the average share of required key facts captured increased from:

13% without Aiera → 32% with Aiera

That represents approximately 2.4 times as many key facts captured.

The result was concentrated among six of the 13 models tested. Those six included both frontier and open-weight models:

The remaining seven models showed little improvement or performed worse after being given access to the same data.

hat finding is central to the study:

Access to better data creates the opportunity for better answers, but not every model can convert that access into better research.

Why the Open Web Is Not Enough

The benchmark was built from nearly 500 source-grounded financial research questions. Each question was tied to verified facts from real Aiera content, including:

• Proprietary research
• Expert interviews
• Earnings-call transcripts
• SEC filings
• Company publications

The questions were written the way a practitioner would naturally ask them. They did not identify the source or tell the model where to look.

Approximately 75% of the complete benchmark, and 78% of the proprietary evaluation subset, required information that could not be found on the open web.

That distinction matters because many financial AI evaluations assume the relevant document has already been placed in the model's context. In a real research workflow, the model must first find the right source, retrieve the relevant information, and then synthesize it into a useful answer.

The Aiera Lift evaluates that complete process.

How the Study Was Structured

Evaluation conducted across 13 models (150-question evaluation, 3,892 graded answers).

Of the 150 analyst-style questions - 116 proprietary or non-web-findable questions, 34 web-answerable control questions - each question was submitted to the same model twice, once in the baseline condition and once with access to Aiera's MCP server.

Frontier models used their providers' native MCP connectors. Open-weight models and Gemini used a common function-calling loop that connected to the same Aiera MCP endpoint.

The models were scored primarily through key-facts entailment. For every question, the expected answer was broken into individual, verifiable facts. A held-out judge then determined which of those facts were supported by the model's answer.

This creates a more precise measure than simply asking whether an answer "sounds good." A response only receives credit for facts it actually contains and supports.

Lift Does Not Follow the Frontier vs. Open-Weight Divide

One of the most important findings is that the models did not split neatly by provider type or capability tier.

Three of the six models that showed significant lift were frontier models:

• Claude-Opus-4.8
• Claude-Fable-5
• GPT-5.5

The other three were open-weight models:

• Kimi-K2.6
• GLM-4.6
• gpt-oss-120B

At the same time, Gemini-2.5-Pro, a frontier model, showed no statistically significant improvement.

The deciding factor was not whether a model was frontier or open-weight. It was whether the model could effectively use the tools and synthesize what they returned.

Tool Use Is More Than Calling a Tool

A deeper trace analysis examined how models behaved after receiving MCP access.

The results challenge a simple assumption: models do not improve merely because they make more tool calls.
Across the traced models, the number of calls had only a weak relationship with realized lift. Some of the weakest models called the tools more often than the strongest models.

For example:

• Qwen3-235B averaged 12.7 tool calls per item
• Mistral-Large averaged 9.0
• Claude averaged 7.6

Yet Claude produced substantially stronger research answers.

What mattered was whether the model completed two distinct tasks:

1. Retrieve relevant source material
2. Convert that material into a complete, fact-bearing answer

The study identified two common failure modes.

Failure Mode 1: Ignoring the Tools

Some models rarely used the available retrieval tools and instead answered from their internal knowledge.

This produced responses that were not grounded in the relevant proprietary sources.

Failure Mode 2: Retrieving Without Synthesizing

Other models called the tools frequently but failed to turn the retrieved material into useful answers.

They entered repetitive loops, exhausted the tool budget, returned thin responses, or omitted key facts even when the underlying information had been found.

Strong-performing models did both parts well. They invoked the relevant retrieval tools and produced substantial answers grounded in the results.

This leads to one of the paper's clearest conclusions:

The interface is necessary, but it is not sufficient.

The value of a data connection depends on the model's ability to use it.

Where Aiera Created the Most Lift

The largest improvements appeared in the areas where proprietary financial research adds the most value.

Cross-document synthesis
Models improved most when they needed to combine information across multiple sources rather than retrieve a single public fact.

Materiality
Aiera helped models identify why a development mattered, not simply describe what happened.

Analyst thesis
Models produced stronger answers when questions required investment perspectives, supporting evidence, or differentiated research.

Guidance changes
Lift was smaller for questions about guidance changes because that information is more likely to appear in earnings releases, transcripts, or public reporting.

The pattern supports the underlying hypothesis: Aiera adds the most value when the answer depends on information or analysis that is not readily available through public search.

The Control Questions Matter

A strong evaluation should not show improvement everywhere.

For questions that could already be answered through the open web, the expected lift was approximately zero. That is what the study found for two of the established native-web frontier models.

Claude-Opus-4.8 and GPT-5.5 already performed well on web-answerable questions in the baseline condition. Connecting them to Aiera did not improve those answers and slightly reduced the measured scores.

That is an important validation of the methodology. It indicates that the evaluation did not simply reward the MCP condition by default.

The lift appeared where the additional data was actually needed.

The paper also reports an important caveat. The open-weight and Gemini models used a shared third-party web-search tool that was weaker than the native search available to some frontier models. As a result, control comparisons are less reliable for that group.

For this reason, the proprietary-question results remain the most defensible measure of Aiera's impact.

What This Means for Financial AI Teams

The findings have several practical implications for institutions developing or evaluating AI-enabled research systems.

Data quality and model quality must be evaluated together
A strong model with incomplete information will still produce incomplete answers. High-value financial AI requires both capable reasoning and access to the right sources.

Tool-use competence should be tested directly
A model's general benchmark score does not reveal whether it can navigate a real research environment. Teams should evaluate retrieval behavior, source selection, synthesis quality, and groundedness.

More tool calls do not mean better research
Operational traces should be assessed for successful retrieval and answer quality, not activity alone.

Different models may be appropriate for different workflows
A model that performs well on summarization or numerical question answering may not be the strongest option for multi-document research, thesis development, or agentic retrieval.

Public-web evaluations are not enough
Institutional research systems should be tested on representative proprietary and point-in-time questions, not only public information that is easy to retrieve.

Methodology and Limitations

The benchmark is private because its questions and verified answers are derived from licensed financial content. Scores are published, but the underlying source material cannot be released publicly. The evaluation also has several limitations:

• MCP retrieval is stochastic, so a model may find or miss a source across different runs
• A held-out LLM judge is used for scoring, although the evaluation relies on narrow fact-level judgments rather than broad preference scoring
• The baseline search experience differs between native frontier connectors and the generic open-weight path
• The reported results use a 150-question subset of the nearly 500-question benchmark
• Model and web-search behavior may change as providers update their systems

These limitations do not eliminate the headline finding, but they define where future evaluations should expand.

Better Data Makes Capable Models Better

The Aiera Lift provides a controlled measure of what happens when foundation models gain access to professional financial content through a standard interface.

For the six models capable of effectively using the tools, answer quality increased from 13% to 32% of required key facts captured, a 2.4× better result.

The improvement was concentrated on the questions where institutional data matters most: proprietary research, expert insights, multi-document synthesis, materiality, and analyst thesis.

But the results also show that access alone is not enough.

The strongest financial AI systems will combine:

• High-quality, permissioned data
• A model capable of agentic retrieval
• Effective synthesis of retrieved information
• Transparent, source-grounded outputs

Aiera's MCP server provides the standardized data layer. The model determines how much of that value is ultimately realized.

Download the full Aiera Lift paper to review the complete methodology, model-level results, confidence intervals, controls, and limitations.