Endex builds an autonomous financial analyst on OpenAI's reasoning models

What Endex actually changed under the hood

Endex describes moving away from retrieval-augmented generation (RAG) as its core approach. Instead of fetching passages and generating from them, its agents use OpenAI's o-series reasoning models to pull facts, spot inconsistencies, and contextualize metrics the way an analyst would.

The company also reports collapsing a previously elaborate pipeline. It says it had relied on complex prompting, chained completions, and multiple verification steps, and that o1 let it simplify that without losing accuracy. Separately, o3-mini delivered what Endex calls comparable intelligence at one-third the latency per turn, which it uses for multi-step work like analyzing confidential information packages and reconciling financial models.

Those are two distinct claims worth separating: one is an architectural bet (reasoning over retrieval), the other is a cost-latency tradeoff (a smaller reasoning model for the high-volume steps). Both are the kind of decision an applied team makes only after building an evaluation harness.

The frontend is the deliverable, not a chat box

The most concrete detail for anyone building agent interfaces is what Endex ships back to the user. Agents deliver outputs as emails, documents, Excel models, or slide decks — the artifacts finance teams already work in — rather than a transcript. The named tasks are equally specific: precedent transaction overviews, earnings performance summaries, investment committee memo preparation, and data room due diligence.

CEO Tarun Amasa frames the ambition explicitly around interface rather than raw capability.

Our work goes beyond APIs – it's about building the agent-user interfaces that will change how financial analysts do work.Montana Labs

That framing matters because it names the actual product surface. The model produces reasoning; the frontend's job is to render that reasoning into a memo or model a professional can hand to a committee. The stated feature that analysts can trace agent conclusions back to their sources is a UI commitment as much as a model one — citations have to be surfaced, clickable, and tied to specific footnotes.

Trust is built through cross-checking and citation, not just answers

Endex leans on the failure modes finance teams fear: a missed EBITDA reconciliation adjustment, an overlooked change-in-control provision. Its answer is agents that flag restatements in footnotes and surface inconsistencies with targeted citations, shifting the analyst's role from manual verification to decision-making.

To measure whether the model reads the documents that actually matter, Endex built its own benchmark — Finance Agent Retrieval (FAR) — to measure context usage on tabular and chart data. It reports using o1's vision capabilities to process investor presentations, Excel models, and 8-Ks. Building a domain-specific benchmark for tables and charts is a telling signal: generic retrieval scores don't capture whether a model read the right cell in a financial statement.

On preference, the company reports that in blind user testing financial experts preferred o1's responses 70% of the time over non-reasoning models. It also describes reinforcement fine-tuning of GPT-4o mini and o1 to convert custom datasets into reasoning improvements for tasks like precedent transaction analysis and entity extraction.

The specific implication: verifiable reasoning is the product surface in finance

The lesson from Endex is not that reasoning models are better — it's where the engineering effort landed. Endex invested in a benchmark for chart and table comprehension, an evaluation framework that tracks latency, first-token time, and reasoning depth, and output formats that match existing finance artifacts with traceable citations.

For a regulated, detail-sensitive domain, the frontend contract is that every conclusion can be traced back to a source document. That constraint shapes everything upstream — model choice, the decision to reason rather than retrieve, and the choice of a lower-latency model for the repetitive steps. Teams building vertical agents should read this less as an endorsement of any single model and more as a demonstration that the interface for surfacing and citing reasoning is where the durable product work happens.