Decisional Knowledge Engines are now live on our website at

https://www.decisional.com/

Over the past few months, we've conducted in-depth conversations with hundreds of financial professionals to understand the real barriers preventing AI adoption in high-stakes decision-making. Through these discussions, we gained critical insights into specific use cases, the complexity of unstructured financial data, existing workflows, and the exacting quality standards required for generating actionable insights.

What We Learned

The most profound lesson was recognizing the extraordinary level of complexity and nuance that financial professionals navigate daily. This understanding fundamentally shaped our approach and reinforced our commitment to building Decisional as the definitive AI-native workspace for financial work.

A new way for financial professionals doing in-depth work to collaborate with AI agents.

TLDR

Knowledge engines are AI native workspaces built for financial professionals in private markets to accelerate due diligence, fundamental research, and portfolio management in a secure sandbox environment with full citations and audit trails.

Preview video here -

THE MOUNTAIN OF FILES

Financial professionals spend hours doing grunt work that includes copy pasting tables, screenshotting charts and doing data entry and have been unable to use AI effectively due to a lack of precision and transparency in using LLMs in their workflows.

The Decisional knowledge engine allows you to add and view sources to so that you can

1. Write your own thoughts into Memos

2. Ask AI questions related to the sources you upload

3. Generate / Edit high quality tables, charts and metrics based on the underlying materials with your AI Assistant Dex

Decisional features a powerful document processing system that handles multiple formats (excel, pdf, ppt, docx) with OCR for tables, charts, and figures. Our agentic RAG implementation grounds AI responses in your sources, delivering up to 4.1x greater accuracy with fewer hallucinations and the most transparent citations. Choose from top models by Anthropic, xAI, OpenAI, and Meta.

But that isn’t why Finance pros should use Decisional.

1.You start from SCRATCH

Don’t become a consensus thinker, use AI only when you need it - to do the grunt work not the real decision making. Upload sources you trust and view it on the left, iterate on your workup in the middle and ask AI for assistance on the right, ONLY when you need it.

2.Deep Research & Agentic Editing with Finance Grade Tables, Charts and Writing

We’ve tuned Dex to be thoughtful, analytical and rigorous like any financial professional worth their salt. It spends time thinking before writing to improve depth and understanding. Get Dex to refine outputs, answer follow-ups, and iterate on your materials. Write inline when you need to onto a collaborative editor with Dex. It can even do financial modeling and scenario analysis.

3.Full transparency

Every insight links back to original sources with complete audit trails. Imagine an analyst that can read hundreds of thousands of files in seconds but tells you the exact page with a bounding box around where it got a figure from. Inline citations ensure you don’t have to second guess whether the data is accurate, see for yourself without yielding all your time savings to fact check AI.

All this with enterprise grade APIs to embed AI agents within your workflow and deployments that work with your data security policies.

Try Decisional free for a week with no credit card

There are multiple AI Agents for finance but none offer the level of transparency and control that decisional does. Most focus on sourcing or search, Decisional focuses on deep thought and work.

No smoke and mirrors, just results you can trust.
Choose from one of our public samples where we have pre loaded documents for you:

• AI Infrastructure Build Out

• US Gambling Restrictions

• Tesla SEC Filings (10 years)

• Coffee Retailer Data Room

OR login and upload your own documents here - https://app.decisional.com/sign-in

Reach out to us at founders@decisional.com if you want to chat.

Overall, 77% of professionals now predict AI will have a high or transformational impact on their work over the next five years, up 10 percentage points from last year. This comprehensive guide compares the top AI financial analysis platforms helping Family offices, PE firms, investment banks, and wealth managers transform their workflows in 2025.

The Dual Reality of AI in Finance

1. There is a lot of hype about AI.

The financial services industry is at a critical juncture with AI adoption, as some aggressive innovators make AI usage mandatory for all employees, offering $25,000 bounties for new AI applications .

2. AI is going to transform the way financial services are delivered.

Specialized platforms have emerged that promise dramatic efficiency gains for deal-making, research, and client communications.

We live in a strange time where both these statements are true.

The Implementation Paradox

It has never been easier for financial services firms to transform the way they work.

Yet it has never been harder to pick the right tools and stack for the job, as the proliferation of AI platforms demands a strategic approach to selection and implementation.

The Decisional team has been speaking to hundreds of financial professionals from private equity, family offices, venture capital, and wealth management sectors, who are consistently keen to learn about new tools available.

Each sector has unique needs: private equity firms conducting intensive due diligence benefit from document processing capabilities; investment banks find integration of 65+ million sources valuable; while wealth management firms often start with more accessible platforms like ChatGPT for client communications.

How AI Financial Analysis Actually Works (And Why It's 100x Faster)

AI financial analysis uses machine learning, natural language processing, and predictive analytics to automate complex financial tasks that traditionally require hours of manual work. Here's what makes it impactful:

AI vs Traditional Financial Analysis Comparison

These AI performance metrics in financial analysis were validated through time-motion studies comparing traditional and AI-powered workflows. Processing speed (100x faster) was measured by tracking analysts spending 4+ hours on financial models versus AI completing them in 2-3 minutes. The 65+ million data sources reflects AI's ability to simultaneously access Bloomberg, Reuters, SEC filings, and other financial databases versus manual sequential searching. Error reduction (88%) came from quality audits showing manual analysis contained errors in 88% of cases versus <12% with AI. Forecast accuracy improved 13% based on backtesting (53% to 60% accuracy). Document processing increased 300x from manual rates of 50-100 pages daily to AI processing 1,500+ pages hourly, validated through workflow analysis of earnings reports and regulatory filings.

Key Capabilities Transforming Financial Workflows

1. Automated DCF Modeling: Generate sensitivity analyses in seconds

2. Intelligent Document Processing: Extract data from 1,500+ documents simultaneously

3. Real-Time Variance Analysis: Identify anomalies across massive datasets instantly

4. Predictive Financial Forecasting: Leverage ML for 60% accuracy improvements

5. Natural Language Querying: Ask complex questions, get cited answers

Preparing for an AI-Powered Future

Louis Pasteur, the French microbiologist and chemist, famously said

In the context of financial AI, preparation means starting with clearly defined use cases, prioritizing platforms with robust verification capabilities, and maintaining human oversight for all critical decisions.

If you need to prepare for a world where AI touches every aspect of knowledge work, you need to start thinking strategically about deployment. The firms that master the balance between innovation and prudence will gain significant competitive advantages in an industry where information processing speed and accuracy increasingly determine success.

In that tune, here is a guide on some of the most popular tools we have been hearing about from financial professionals across the public and private markets. This guide is not just a perspective, but a tangible piece of research with pricing, interface screenshots and key features

Comprehensive AI Financial Analysis Tools Guide

Specialized Financial AI Platforms

1.Rogo AI

Category: Specialized Financial AI
Target Users: Investment Banks

Overview:
Rogo AI represents the premium end of specialized financial AI solutions, designed explicitly for institutional finance users requiring extensive data access via a chatbot.

Key Features:

• 65+ million integrated sources covering SEC filings, market research, and live news

• Sub-60 second query processing for time-sensitive financial analysis

• Dedicated deployment options for enterprise clients

Pricing:
Custom pricing structure varying by deployment scope and organizational requirements. The high implementation cost reflects enterprise-level integration capabilities.

Best For:
Investment banking applications requiring extensive data sources.

2.Hebbia AI

Category: Specialized Financial AI
Target Users: Legal, Financial Firms looking to parse documents in table structure

Overview:
Hebbia AI has a document-focused approach to financial analysis suited to parsing through data rooms with thousands of files.

Key Features:

• Matrix spreadsheet interface for financial data organization

• No RAG single doc in context window approach

• Unlimited document processing capacity in table format

Market Position:
Market leader with approximately 90% penetration in the private equity sector.

Pricing:
Custom $2000 per seat on average, with negotiable options down to for larger deployments.

Best For:
Financial firms dealing with huge 1500+ document data rooms.

3.Decisional

Category: Specialized Financial AI
Target Users: Private Markets focused PE, VC, Family Offices or Wealth Advisors

Overview:
Decisional offers a specialized financial AI solution that balances advanced agent capabilities like deep reasoning with the deepest verifiability through citations that go right to the source.

Key Features:

• Thoughtful AI Agent (Dex): An analytical AI agent that carefully processes information before generating outputs, designed specifically for financial professionals.

• Comprehensive Document Support: Handles multiple document formats including PDFs, PowerPoint, Word, and other formats, with unlimited document uploads and the ability to read text, charts, logos, and figures.

• Full Citation Transparency: Every insight links back to original sources with complete audit trails, allowing users to verify information.

• Interactive Collaboration: Users can refine outputs, get follow-up answers, and edit collaboratively with Dex in an integrated editor.

Market Position:
Mid-range pricing with specialized features focusing on knowledge representation and verification.

Pricing:
Starting at $99 per month with zero implementation time.

Best For:
Private Markets focused PE, VC, Wealth Management and Boutique firms with lean effective teams looking to get as much done as big bulge brackets.

General Purpose AI Tools

4.ChatGPT (OpenAI)

Category: General Purpose AI
Target Users: All Financial Services

Overview:
ChatGPT provides financial professionals with a versatile general-purpose AI solution that can be adapted to numerous financial use cases. While requiring more oversight than specialized platforms, it delivers exceptional ease of use.

Key Features:

• Document summarization capabilities for financial reports

• Code generation for financial modeling and analysis

• Earnings call analysis and synthesis

• Natural language processing for client communications

• Multiple deployment tiers to match organizational needs

Market Position:
Low to mid-range pricing with versatile application across various financial workflows.

Pricing:

• Team: $30 per month

• Pro: $200 per month

• Enterprise: Custom pricing

Best For:
Wealth management firms seeking cost-effective client communications tools and general analysis capabilities.

Research & Verification Tools

5.NotebookLM (Google)

Category: Research & Verification
Target Users: Analysts, Researchers

Overview:
NotebookLM stands out as a research-focused tool with robust source verification capabilities, making it particularly valuable for teams requiring thorough documentation of their financial analysis process.

Key Features:

• Source-grounded analysis with inline citations

• Support for up to 50 sources per notebook

• 25 million word processing capacity

• Inline citations for audit trails

• Integrated file viewer for document navigation

Market Position:
Free tier available with research-focused features; enterprise options for larger teams.

Pricing:
Free tier available for basic use, with enterprise pricing varying based on organizational needs.

Best For:
Research teams requiring audit-appropriate documentation tools with minimal training requirements.

6.Perplexity Finance

Category: Research & Verification
Target Users: Investors, Analysts

Overview:
Perplexity combines research capabilities with real-time financial data access, creating a powerful platform for investors and analysts requiring up-to-date information from premium financial databases.

Key Features:

• Support for up to 500 internal documents

• Real-time financial data streaming

• FactSet and Crunchbase integration for premium data access

• AI editorial assistant for report generation

• Enterprise search capabilities

Market Position:
Mid-range pricing with specialized real-time data features.

Pricing:
Enterprise Pro: $40 per month per seat (Enterprise Pro)

Best For:
Investors and wealth management firms require real-time data access at reasonable cost points.

Last Updated at - Jun 23, 2025

The "Vibe Test" Comes First

Here's something that caught us completely off guard: the first thing users do when they upload documents isn't ask the sophisticated, domain-specific questions we optimized for. Instead, they perform what I call a "vibe test."

Users upload a random collection of documents—CVs, whitepapers, that PDF they bookmarked three months ago—and ask exploratory questions like "What is this about?" or "What should I ask?" These documents often have zero connection to each other, but users are essentially kicking the tires to see if the system "gets it."

This led us to an important realization: benchmarks don't capture the vibe test. We need what I'm calling a "Vibe Bench"—a set of evaluation questions that test whether your system can intelligently handle the chaotic, exploratory queries that build initial user trust.

The practical takeaway? Invest in smart prompt suggestions that guide users toward productive interactions, even when their starting point is completely random.

Also just because you built your system to beat domain specific benchmarks like FinQA, Financebench, FinDER, TATQA, ConvFinQA doesn’t mean anything until you get past this first step.

The Goldilocks Problem of Output Token Length

We discovered a delicate balance in response length that directly correlates with user satisfaction. Too short, and users think the system isn't intelligent enough. Too long, and they won't read it.

But here's the twist: the expected response length scales with the amount of context users provide. When someone uploads 300 pages of documentation, they expect a comprehensive response, even if 90% of those pages are irrelevant to their question.

I've lost count of how many times we tried to tell users "there's nothing useful in here for your question," only to learn they're using our system precisely because they don't want to read those 300 pages themselves. They expect us to be comprehensive because they provided comprehensive input.

Multi-Step Reasoning Beats Vector Search Every Time

This might be controversial, but after extensive testing, we found that at inference time, multi-step reasoning consistently outperforms vector search.

Old RAG approach: Search documents using BM25/semantic search, apply reranking, use hybrid search combining both sparse and dense retrievers, and feed potentially relevant context chunks to the LLM.

New RAG approach: Allow the agent to understand the documents first (provide it with tools for document summaries, table of contents) and then perform "RAG" by letting it query and read individual pages or sections.

Think about how humans actually work with documents. We don't randomly search for keywords and then attempt to answer questions. We read relevant sections, understand the structure, and then dive deeper where needed. Teaching your agent to work this way makes it dramatically smarter.

Yes, this takes more time and costs more tokens. But users will happily wait if you handle expectations properly by streaming the agent's thought process. Show them what the agent is thinking, what documents it's examining, and why. Without this transparency, your app will just seem broken during the longer processing time.

There are exceptions—when dealing with massive documents like SEC filings, vector search becomes necessary to find relevant chunks. But make sure your agent uses search as a last resort, not a first approach.

Parsing and Indexing: Don't Make Users Wait

Here's a critical user experience insight: show progress during text layer analysis, even if you're planning more sophisticated processing afterward i.e table and image parsing.

Two reasons this matters:

1. You don't know what's going to fail. Complex document processing has many failure points, but basic text extraction usually works.

2. User expectations are set by ChatGPT and similar tools. Users are accustomed to immediate text analysis. If you take longer—even if you're doing more sophisticated work—they'll assume your system is inferior.

The solution is to provide immediate feedback during the basic text processing phase, then continue more complex analysis (document understanding, structure extraction, table parsing) in the background. This approach manages expectations while still delivering superior results.

The Key Insight: Glean Everything at Ingestion

During document ingestion, extract as much structured information as possible: summaries, table of contents, key sections, data tables, and document relationships. This upfront investment in document understanding pays massive dividends during inference, enabling your agent to navigate documents intelligently rather than just searching through chunks.

Building Trust Through Transparency

The common thread through all these learnings is transparency builds trust. Users need to understand what your system is doing, especially when it's doing something more sophisticated than they're used to. Show your work, stream your thoughts, and set clear expectations about processing time. We ended up building a file viewer right inside the app so that users could cross check the results after the output was generated.

RAG isn't dead—it's evolving from a simple retrieve-and-generate pattern into something that more closely mirrors human research behavior. The systems that succeed will be those that understand not just how to process documents, but how to work with the humans who depend on them.

He could be right. But that raises a more interesting question:

What kind of knowledge worker will AI create?

Coding made nerdy hackers rich. Zuckerberg and Gates wrote code and they saw what code could become. Now we're at a similar inflection point, except the new tool isn't a programming language. It's AI.

Micheal Truell, one of the co-founders of Cursor said something revealing recently (I heard it at an event we attended hosted by Notion): "We don't consider ourselves a vibe coding platform even though we were mentioned in the original tweet."

What he meant was that they intended serious programmers use Cursor to push the ceiling higher, not to make coding feel good. For every casual "vibe coder," there's someone learning to bend AI to their will—someone who treats tokens like a new kind of capital.

I think these people will be the Agentic knowledge workers.

I was one of those entry-level workers Dario thinks will disappear half a decade ago. My first job was at a Bank, and like most knowledge work, it had a secret dirty ratio: maybe 40% real thinking, 60% grunt work. Every week I'd spend hours reconciling transactions in Excel. It wasn't why the Bank hired me—they hired me because I went to a good school and could grind through problems. But grinding was necessary.

Later, as a product manager at a Fintech, I'd read through thousands of customer tickets to find patterns. Again, grinding through data to get to insight. The company cared about the insights, not the grinding, but you couldn't get one without the other.

The grinding served a purpose though, working on the formatting made you more aware of the content, writing a feature spec makes you think more carefully. The new kind of grinding will probably also serve a purpose, like prompting and validating agent outputs.

But AI changes the ratio of grunt work dramatically. The grunt work disappears. But does that make the knowledge worker obsolete?

No—it makes resource allocation the core skill.

Before AI, knowledge workers allocated mainly one resource: their own time. You had to decide whether to spend Tuesday morning on that Excel reconciliation or on strategic planning. As you moved up, you allocated other people's time too, plus capital.

Now there's a third resource: compute. Specifically, AI agents that can do the grunt work. The question isn't whether you should spend time checking that all the fonts in your spreadsheet match—it's whether you should configure an agent to check fonts, or whether that's even worth solving.

This is harder than it sounds. When your main constraint was your own time, allocation was intuitive. You felt the cost of every hour. But compute feels free until it isn't, and delegation to AI requires entirely new judgment calls.

The best knowledge workers will be those who master this new allocation problem. They'll know when to use AI, when to do things themselves, and when to ignore problems entirely. They'll treat compute like a portfolio manager treats capital—as a scarce resource to deploy for maximum impact.

We're not just automating away grunt work. We're creating a new class of knowledge worker whose primary skill is knowing what to automate.

The talent war for Agentic knowledge workers could have already begun.

Source: https://www.bondcap.com/report/pdf/Trends_Artificial_Intelligence.pdf

Deep Research is the newest AI application of “reasoning models”. The goal is to leverage reasoning-focused large language models to generate comprehensive, detailed reports on specific topics while searching the web. Unlike typical chatbot interactions, Deep Research takes more time to respond (up to 10 minutes) because it actively searches and analyzes multiple sources. This thorough approach enables the AI to compare different perspectives and synthesize accurate, up-to-date information instead of relying solely on pre-trained knowledge.

As the AI model explores more information and sources, it modifies its research plan and moves to a next set of sources spending time between searching and reading to finally synthesize an output.

Key Features:

• Generate long form narrative reports

• Web Search - Ability to search the web and find sources to compile said report

• Analysis - Use tools like python to analyse information found in sources

• Citations - Attribute answers to sources through linked citations

Now you know that you have PhD-level intelligence on tap and you want to utilize it. How do you choose between models? The AI companies provide you with the following chart, which, to be honest, is pretty hard to grok.

Where the Rubber meets the road - testing out different Deep Research products

Decisional has been working on this space for the last few months. I am extremely interested in being able to effectively deploy AI for completing my own knowledge work.

So I decided to test out the following Deep Research on a topic I am reasonably familiar with i.e Internal Combustion Engines (I was a motorhead and wanted to pick an old school topic).

Here are the current Deep Research options we have available that I checked out:

• OpenAI Deep Research

• Google Gemini Deep Research

• Perplexity R1 Deep Research

• Grok 3 Deep Research

Just like building an Eval, I started to think of things that would be super important for AI to cover while thinking from first principles BEFORE looking at the results and after speaking to a few friends I came up with the following:

• Fundamentals

  • How does an IC Engine work?

  • What are the components?

  • What are the different types of IC Engines?

  • How do you measure the power of an engine?

• Comprehensiveness

  • How much detail does the report go into ?

    • Explanation of straight six vs W series of engines

    • How vibrations impact the shape of the engine

• Visualisation

  • Diagrams

  • Compression ratios

  • Weight to power

  • Efficiency curves

• Ease of Reading

  • Is it interesting to read?

  • Does it have stories that may make someone interested to explore the topic further?

• Logical flow

  • Is the information organized in a way that makes sense

  • Does the index and table of contents make sense

  • Can someone easily navigate to the section they are interested in

• Quality of Sources

  • Are you picking up something from a blog or high trust sources

  • How many sources are you using in order to come up with information

The Test

Open AI’s Research Report

Perplexity

Grok 3

Gemini Deep Research

The Results

Insights

Models are able to capture fundamentals and make the report logically coherent but are still lacking in visualization and comprehensiveness

🟢 'Fundamentals' and 'Logical Flow' peaked at a perfect score for OAI, highlighting strong foundation and coherence across presentations.

🔔 'Visualization' recorded no variation, consistently scoring 2. This indicates a potential need for improvement across all organizations.

⚠️ 'Ease of Reading' showed the most variability, suggesting inconsistent clarity and accessibility in presentation styles.

Prompt used to evaluate all four DR reports:

"Give me a detailed report on how an IC Engine works"

• OAI - 9000 Words, 27 Sources

• PPLX - 964 Words, 23 Sources

• Gemini - 2000 Words, 122 Sources

• Grok 3 - 2000 words, 16 Sources

The Winner

And so the current winner is Open AI, but with clear areas for improvement especially on the ability to visualize information. The variability among the other models is not too high either so you may not be able to see much of a difference. Gemini is probably the best for covering a large breadth of the internet and Open AI the best in focused research while Grok seems to be the best in ease of reading and perplexity in speed of generating an answer.

Nerdy Observations

• Visualisation

  • Lack of ability to visualize or make diagrams

  • Even PPLX, which is generally good at pulling up stock market charts, didn't translate this into visual representations

  • OAI, despite excellent performance in most areas, didn't leverage diagrams to enhance understanding

• Comprehensiveness

  • None adequately covered specific engine configurations like straight-six vs. W-series engines

  • All papers provided minimal exploration of how vibrations impact engine design

  • Even OAI, with its otherwise comprehensive coverage, scored only a 4/5 on comprehensiveness due to this gap

  • This uniform weakness in very specific technical areas suggests these systems may share similar training limitations with respect to specialized engineering knowledge

• GROK's Unexpected Accessibility Advantage

  GROK demonstrates an unexpected strength in communication approach

  • It begins with a "Key Points" summary section not found in other papers

  • It scores equal to OAI on readability (4/5) despite providing less sophisticated technical content

  • Its overall score (3.3) exceeds GEMINI (3.2) primarily due to this communication advantage

  • This suggests GROK might be particularly optimized for accessibility and clear communication compared to systems often considered more technically advanced (that makes sense based on the training data being using X which is optimized for short clear communication).

The Future of Deep Research

Deep research is a great stepping stone towards helping AI apply knowledge work but it is lacking in three ways:

• Control - These are still wall of texts and you cannot control or modify the text

• Transparency - Showing you why a certain part of the report was compiled in a certain way

• Visuals - While this may not be obvious but business research often needs to be aesthetically pleasing (cue MBB consultants and their presentation formatting skills) whereas academic research generally does not

References

1. https://openai.com/index/introducing-deep-research/

2. https://www.sequoiacap.com/podcast/training-data-deep-research/

3. https://marginalrevolution.com/marginalrevolution/2025/02/deep-research.html

4. https://www.datacamp.com/blog/deep-research-openai

5. AI Enterprise tools

Could it be that it is already happening?

We have always been fielding this question since I started working on Decisional. I now have a fresh example to offer of a sneak peak of the future.

The Short Case for NVIDIA Stock [1] , a 12,000 word blog post shared on an X with over half a million views ended up being the catalyst for more than a $1 Trillion in sell off in the market in late January.

What is interesting isn’t just the short case and subsequent impact it had on markets (a great read) but the story of the individual that came up with the post Jeffrey Emanuel.

Jeffrey Emanuel is a former hedge fund analyst with over a decade of experience at top long/short firms like Millennium and Balyasny, where he specialized in both longs and shorts. Around 2010, he began focusing on deep learning when it was still highly experimental (think Restricted Boltzmann Machines and MATLAB). He then switched gears to hands-on AI development and now maintains popular open-source projects, while being a power user of models like Claude (has over 6 accounts) and ChatGPT Pro ($200 Subscription), etc.

This unique blend of market experience and technical expertise informs his analysis of AI-related companies like Nvidia. He ended up published this essay serendipitously as an after thought and it may be a piece of content (unrelated to a crisis event) that has created the single largest swing in the history of markets.

Here is the timeline of events:

• 7AM PST Jan 25th - NVIDIA Price: $142 at close of market on Jan 24th

https://x.com/doodlestein/status/1883168862704947237

• 5:19 PM PST Jan 26th - VP of Dropbox shares this

https://x.com/morganb/status/1883686164336709661

• Subsequently goes viral after being shared by Chamath, Naval, Jared and a number of folks

• Monday Jan 27th - What happened next is an over $Trillion selloff over the weekend with NVIDIA losing $600 Billion the single largest drop in its history.

The Consensus View was missing something

All the reports from wall street were already out and NVIDIA is THE most covered stock in the world and yet, somewhere out there there was an AI practitioner out there in the world with something unique to say that led to a trillion dollar swing in the market. Each word in that blog post could be worth a cool $83 million.

The craziest thing is that Jeffrey didn’t even make a trade himself on this short and was using the post as a way to generate organic impressions for his blog!

“I never actually shorted Nvidia. I just put my analysis out there, and then the market reacted—but I didn’t take a position myself.

Jeffrey Emanuel on a the Delphi Podcast [2]

AI Application boom will grant superhuman abilities

In the post ChatGPT era there was an overhanging question, what are examples of AI applications that are actually making money besides the foundation model company. Or will there be other applications ? And will there be anything besides OpenAI and Github Copilot.

There were serious questions on how the revenue was going to be generated from the app layer.

There was not much to point to in 2022 and 2023. But now it is 2024 and AI apps are supercharging productivity across sectors giving birth to the single fastest growing SaaS app of all time.

• Cursor - 100M ARR in 2 years

• Bolt.New - 10M ARR in 2 months

• Harvey - 10M ARR in 2023

• Glean - 30M ARR

There is no other vertical where AI enhanced cognitive capabilities can have as huge impact like Financial Services.

“It’s Christmas every day for Application companies”

“I created my own silly website that turns Youtube videos into transcripts. When I first programmed this I was going to do it with GPT 4o calls. I was like as long as I can make a couple of bucks it is fine but then I realized I could use GPT 4o mini at almost no cost. I could have made the math work with GPT 4o but now it costs my less than a penny to process the document and now I have a 99% profit margin.”

Jeffrey Emanuel on a the Delphi Podcast [2]

It is also christmas day for the AI supercharged professionals of our age.

Every IQ point for cognitive work is worth $1500 on avg in the US and the models are getting smarter and smarter and we are not even talking about the speed of the same cognitive work getting reproduced faster through AI.

Although it isn’t mentioned it is super likely that AI tools were used in creating a post that moved the market by a trillion dollars. Write 12,000 words in a few hours is not small feat!

[1] Short case for NVIDIA Stock

[2] PODCAST - Jeffrey Emanuel: Viral Author of The Short Case for Nvidia Stock

[3] Deep Research Ouput

[4] Why GPT Wrappers are good

The Coffee Shop Conundrum

I recently overheard a snippet of conversation at a coffee shop in San Francisco:

Person 1: “O3 Mini just dropped, and it’s available in Cursor, but honestly, it’s still not as good as Sonnet.”
Person 2: “Is that an O3 problem or a Cursor problem?”
Person 1: “Definitely O3. I don’t blame Cursor.”
Person 2: “Why not?”
Person 1: “Cursor is just the interface. They let you pick which model you want to use.”

This holds a valuable lesson for every AI application builder: if you lock your users into a single model and the results aren’t up to par, YOU take the blame. But if you give users the flexibility to choose—and a particular model fails—the MODEL takes the hit instead.

At first glance, that might sound like a trivial user-experience detail, but it actually says something much deeper about where AI is headed—something that echoes Richard Sutton’s famous “Bitter Lesson.”

The Bitter Lesson

My co-founder Adit, a machine-learning researcher, has often pointed me to Richard Sutton’s seminal essay, The Bitter Lesson. Sutton, one of the founding fathers of reinforcement learning at DeepMind, argues that the approaches leveraging vast computation and search consistently outperform those that rely on handcrafted human knowledge or features. In other words: scale beats cleverness.

This principle crops up time and again in AI progress and has been re-enforced with Deepseek R1. Models that simply consume more data and more compute often stumble into tactics and insights that even the best human engineers never hard-coded. And that’s exactly what we’re seeing with the newest wave of foundation models.

You may say but OpenAI is using more scale and compute so how does the bitter lesson get emulated here. It is counterintuitive since Deepseek was trained on a much smaller budget ($6Mn) but the really interesting part was the algorithmic breakthrough that decided to do away with human feedback based reinforcement learning.

The “Aha Moment” in DeepSeek-R1-Zero

It is worth noting the “aha moment” we observed in the training of DeepSeek-R1-Zero. In the excerpt below, there’s a remarkable point in the model’s development when it starts to reevaluate its own approach—spending more “thinking time” on complex prompts.

A particularly intriguing phenomenon observed during the training of DeepSeek-R1-Zero is the occurrence of an “aha moment”. This moment, as illustrated in Table 3, occurs in an intermediate version of the model. During this phase, DeepSeek-R1-Zero learns to allocate more thinking time to a problem by reevaluating its initial approach. This behavior is not only a testament to the model’s growing reasoning abilities but also a captivating example of how reinforcement learning can lead to unexpected and sophisticated outcomes.

This moment is not only an “aha moment” for the model but also for the researchers observing its behavior. It underscores the power and beauty of reinforcement learning: rather than explicitly teaching the model on how to solve a problem, we simply provide it with the right incentives, and it autonomously develops advanced problem-solving strategies. The “aha moment” serves as a powerful reminder of the potential of RL to unlock new levels of intelligence in artificial systems, paving the way for more autonomous and adaptive models in the future.

This is one of the most powerful affirmations yet of The Bitter Lesson: you don’t need to teach the AI how to reason, you can just give it enough compute and data and it will teach itself! - STRATECHERY [2]

Why is that such a big deal? Because no one explicitly taught DeepSeek-R1-Zero to backtrack or to be more reflective. The model spontaneously discovered that giving itself more internal deliberation led to better results. In other words, we just gave it the right incentives through reinforcement learning, and it invented an advanced problem-solving strategy on its own.

For researchers, seeing a model evolve its own meta-strategies feels like a revelation. But it’s also a testament to how letting scale and search run wild can yield emergent behaviors—and how little we often need to guide the process. This is precisely the demonstration of the Bitter Lesson: you don't teach the AI how to reason, you give it the platform (enough data and compute) so it can figure reasoning out on its own.

The New Reality for AI Startups

Over the past couple of years, AI startups have gotten used to picking a single foundation model—say GPT4, or a specialized alternative—and then contorting their application and guardrails around it. This was fine when new models arrived once a year; you could thoroughly fine-tune or customize each time. But now, with cutting-edge releases dropping every few months (and model performance often leapfrogging prior state-of-the-art), that one-model approach has become a liability.

Model Timeline Acceleration

• Closed-source leaps: Larger, well-funded entities keep rolling out proprietary models with impressive capabilities.

• Open-source momentum: Meanwhile, the open-source community produces more models, some surprisingly close in performance to commercial ones—especially at specific tasks.

The gap shifts almost monthly, making it hard to bet everything on one system. If your application is tightly coupled to a single model, you end up re-engineering your entire product whenever a new contender emerges. That’s a recipe for near-constant disruption and wasted development cycles.

Pick Your Side—Wrapper or Model Provider

To paraphrase Bill Gurley, technology markets often favor either those who own the transformative tech or those who aggregate it best. If you’re an old-school model-training shop, you once asked customers for proprietary data in exchange for specialized AI solutions (like Datarobot). But today, you’re going head-to-head against foundation models trained on nearly the entire internet. They enjoy scale advantages—both in pre-training and in the sheer resources behind them as well as test time compute.

On the other hand, if you build on the application layer—aggregating, orchestrating, or “wrapping” multiple models—your advantage is flexibility. You can slot in the newest or most cost-effective model as soon as it becomes viable. But being in the middle can be tough: you risk repeated context switching and re-tuning if you aren’t designing your system from the outset to be model-agnostic.

Practical Implications

1. Model-Agnostic Infrastructure: We’ve reached a point where it’s dangerous to rely on too many idiosyncrasies of any single model. Instead, design an agent that can route tasks to the best model for each job—or at least let your users choose, as the coffee shop conversation illustrated.

2. Cost vs. Performance: If a new open-source model can match 80% of the performance of a top-tier commercial offering at half the cost, that might revolutionize the unit economics of your startup. Don’t overfit your entire pipeline to a single model you can’t swap out.

3. Guardrails at a Higher Level: Don’t bake your application logic so deeply into one model’s architecture that you can’t pivot. Instead, handle domain-specific guardrails and user flows above the model layer. This also creates a more transparent, modular design that fosters user trust.

4. Continuous Evaluation: The days of annual revalidations are gone. You’ll need a robust, automated system to continuously test new releases and measure them against your core metrics. That’s the only way to keep up with the breakneck tempo of new model launches.

Final Thoughts

All of this points to a broader theme underscored by Ben Thompson’s aggregation theory and the swift shifts we see in AI markets: the value is consolidating at two ends—those who command the greatest scale in training (the model providers) and those who are best at orchestrating these models for specific verticals or user scenarios (the wrappers). The companies that remain stuck in the middle, or married to a single approach, risk being overshadowed the moment a new model with better performance or economics comes along.

For AI entrepreneurs, there is a clear choice: either build the next foundation model (a daunting, massively capital-intensive task) or become the best aggregator—one that always picks the right model at the right time for the right user need. It’s reminiscent of the “aha moment” we saw with DeepSeek-R1-Zero. Don’t waste time crafting overly specialized “hand-coded solutions.” The real challenge is building frameworks flexible enough to leverage that emergent, unstoppable wave of scale.

As Dario Amodei notes in his reflections on DeepSeek, the questions around how these models were trained—or whether you can replicate them with a $6 million budget—are fascinating but secondary to the practical realities for most builders. The train of progress is moving fast. Your job is to pick a side.

References

1. Deepseek & Export Controls – Dario Amodei

2. Deepseek FAQ – Stratechery

3. Short Case for NVIDIA – YouTube Transcript Optimizer

Decisional (An AI Agent for Deep Financial Work) went from 29/150 to 119/150 on Financebench over the past three months (primarily without using o1) through various optimizations on RAG and Tool use. This post highlights the learnings and what you can do to achieve 4X boost in accuracy.

Why RAGnarok?

Ragnarok signals the end of an old world and the start of a new one. For the world of Finance, the explosion of unstructured data - PDFs, earnings calls, market news - is its own form of “data Ragnarok”. This chaotic information overload needs an army of analysts to sift through. Unlike the myth however, it isn’t the end but the start of a new era - the bionic financial professional powered by AI.

As firms and individuals started their explorations with AI - the most common problem they arrive at is hallucinations and lack of context with AI. These are problems that are addressed with Retrieval Augmented Generation architectures that help make an AI system stateful and grounded in order to be able to deal with thousands of pages of complex documents and excel files. However these problems multiply when you need to create Agent system that chains multiple tools and actions together.

You will find this post helpful if you are trying to apply any of the following technologies to your AI stack:

• RAG

• Agents

• LLMs & VLMs

• Vector Databases

• Evals

• Document Ingestion & Parsing

AI Agents need to be built on top of a strong cognitive architecture

RAG - A crucial Tool in the Agent toolkit

Deep learning is a field inspired by biology, with our brain being the most efficient learning system that we started trying to mimic. This biomimetic approach was revolutionized through the contributions of pioneers like Geoffrey Hinton, & Claude Shannon. Similarly an AI Agent system should be mimic an expert financial analyst. This means you need to give it Tools that it can use to replicate human analyst behaviour or skills. At the same time introducing too many tools or making a tool complex really affects the ability to be able to successfully chain together a series of tools to achieve complete a workflow.

A Tool is a specialized component within an AI Agent system that replicates a specific capability or function. Tools are designed to be simple enough to chain together effectively in workflows, while being sophisticated enough to handle complex tasks like RAG queries, data extraction, or table parsing.

EVOLUTION IN COMPUTING USHERING IN THE AGE OF THE AGENT

Core Tools in the AI Financial Analyst stack

An AI Agent for Financial work is built with a Cognitive Architecture that resembles the mental models of a good financial analyst

• Ability to read widely - A good analyst is not constrained by context windows but can go through thousands of pages of information making notes and summaries along the way to build a strong mental model.

  We built our AI Agent in a way that can read thousands of documents and is not limited to the context window of an LLM supported by RAG.

• Ability to read deeply - For some information you need to be able to go through every piece of information, i.e read deeply.

  The AI Agent must be able to deeply analyse any page of a document including the diagrams and tables and not be limited by OCR or text layer extraction.

• Be Quantitative - Keen eye on metrics instead of ideas.

  We built a metric index that specifically tunes answers on numbers linked to entities within documents and not lose context of important metrics that help guide financial analysis.

Decisional’s Agent Tools Inspired by Experts

Based on this Decisional built the following Tools in order to be able to support these expert financial analyst capabilities better:

• Extract Custom Table - A tool that is able to pass a schema of a table and fetch the outputs in a format that is predefined

• Query - Runs a RAG workload across a set of pre uploaded data sources. Combines search and synthesis over thousands of data sources.

• Extract Chunks - A tool that is able to deeply parse a data source into chunks of text that is contextual (i.e specifies graphs, charts, tables and other important context)

Building the Agent

Decisional Cognitive Architecture

The goal of the Decisional’s Agent System is to take two kinds of input tokens:

• Data Source Input Tokens

• Tool Run Input Tokens

The Data Source input tokens are ingested and preprocessed so that it can be used at Tool run execution time where a new set of tool run input tokens are passed on to the system in order to be able execute a workflow.

Since the tool run execution and data source preprocessing is decoupled it is possible to use sophisticated Vision Language Models to process the each data source independently instead of just text layer extraction.

When the workflows are executed the Agent system has the ability to make multiple tool calls as well as more sophisticated techniques to refer to the data sources using a query router in order to analyse the underlying data sources.

Specific Optimizations that led to a 4X increase in performance

• Garbage In, Garbage Out: Unlike systems that rely solely on text layer extraction and basic chunking strategies, we developed custom ingestion pipelines that preserve the context of images, diagrams, and tables - crucial elements in financial documents that often contain key insights. Our chunking strategy involves using LLMs to intelligently partition a page and create detailed, specific chunks. We also attach document and section level metadata for each chunk - an enhanced version of Anthropic’s Contextual Retrieval.

• Metric-Aware Parsing: We built a specialized metric index that is used during ingestion and inference that helps LLMs maintain precise context of numerical data. This ensures that critical financial metrics aren't lost or misinterpreted during analysis.

• Sophisticated Query Planning: Our system excels at complex financial queries through advanced query rewriting and planning, particularly for multi-hop queries that require broader search and retrieval patterns.

• Time-Aware Reranking: Financial data is inherently time-sensitive. Our reranking system labels all metrics with timestamps to maintain temporal context and relevance.

Design Tools to be modular and obvious

The Anthropic blog has good advice here:

Put yourself in the model's shoes. Is it obvious how to use this tool, based on the description and parameters, or would you need to think carefully about it? If so, then it’s probably also true for the model. A good tool definition often includes example usage, edge cases, input format requirements, and clear boundaries from other tools.

Source [4] - https://www.anthropic.com/research/building-effective-agents

Modular tool design in financial AI systems follows the principle of specialized components working together seamlessly. Each tool functions as a focused expert - like a table extraction tool that excels at parsing financial statements but doesn't attempt sentiment analysis. This specialization enables tools to be chained together effectively, similar to how analysts collaborate on complex projects. By keeping interfaces simple and predictable while maintaining robust error handling, these modular tools can be combined into sophisticated workflows while remaining easy to test and validate individually.

Monitor Tool Performance

Tool performance is fundamental - even more critical than overall workflow effectiveness. Our evaluation strategy operates on two levels: First, we validate against Financebench [3], an open-source suite of 150 finance-specific questions that provides standardized performance metrics. Our DRAG system achieves a score of 119, demonstrating a 4.1x improvement over baseline implementations.

However, excelling at public benchmarks is just the starting point. The true test comes from developing custom evaluations that mirror real customer use cases, ensuring our tools perform effectively in practical scenarios. This combination of standardized testing and customer-specific evaluation helps us build tools that are both technically robust and practically valuable for financial analysis.

Conclusion

• The financial industry's "RAGnarok" represents not an ending but a revolutionary beginning, where AI and RAG technologies are transforming how we handle the overwhelming volume of unstructured financial data that previously required massive human effort to process. This shift marks the emergence of a new kind of financial professional - one augmented by AI capabilities.

• Our progress is both measurable and meaningful, with systems like DRAG showing a 4.1x improvement over baselines. However, the true value lies not in benchmark numbers but in how these tools are enabling financial professionals to process, synthesize, and act on information at scales previously impossible, while maintaining the crucial human element in decision-making.

• The future of financial analysis will be built on the foundation of modular, specialized AI tools working in concert with human expertise. By focusing on augmentation rather than replacement, and maintaining strict principles around tool design and evaluation, we're creating systems that amplify rather than substitute human intelligence in financial decision-making.

[1] https://stratechery.com/2024/nvidia-waves-and-moats/

[2] https://arxiv.org/pdf/2406.04744

[3] https://arxiv.org/abs/2311.11944

[4] https://www.anthropic.com/research/building-effective-agents

Welcome to “Neural Markets”

This is our very first post on the Decisional blog called “Neural Markets” and we’re incredibly excited to share our journey and learnings. The name is an ode to deep learning and neural networks being a transformative technology applied to private financial markets.

I’ve been pondering about what I heard from Gwen (the anonymous writer who predicted the scaling laws) on the Dwarkesh podcast: by writing, we contribute to the corpus of text that AI models learn from, effectively influencing how these systems evolve. Gwen emphasizes that this is a critical moment for writers, thinkers, and professionals to “vote” for the values, perspectives, and insights they want AI to carry forward. I think we should share this. By telling our story, we not only reflect on our experiences but also ensure that our voice, our principles, and our vision for the future of AI make it into the training data.

Moreover, since our own product offers source-grounded insights, we’re having a lot of fun using Decisional itself to craft these posts. It feels like the perfect way to bring our technology full circle—building an AI Agent while benefiting from its capabilities as we write.

Recap: The AI Agent for Financial Work

At Decisional, we’re building an AI Agent designed for deep financial work in private markets. We founded Decisional under the conviction that AI would soon spread into every imaginable domain—as a consequence of having “intelligence on tap.” Our founding team has worked at banks, fintechs, and investment funds; wherever we looked, the same story repeated itself: tiresome “pass the butter” tasks are widespread in financial services, and they drain resources and morale.

We saw that finance was starved for rapid, on-demand intelligence. As AI evolves from a 100 IQ assistant to a 200 IQ powerhouse and more close to AGI, the back office and front office will both transform, providing teams with superhuman capabilities that older paradigms of Bloomberg + Excel + MS Office trinity simply won’t do justice.

Tech Eating Financial Services from the Outside In

Technology has long approached the financial services space from the outside in changing business models, platforms but leaving internal business processes untouched. By harnessing the Internet and mobile technology, disruptors like Stripe, Robinhood, Venmo, and PayPal ate into the stack of traditional banks and financial firms. But their impact mostly stayed consumer-facing. Internally, old processes and “pass the butter” work continued unabated. Even the most forward-looking fintechs still rely on large teams for customer support and onboarding.

AI Will Now Devour Unstructured Work

There is little value in making people sift through hundreds of files, reconcile line items, or scan contracts for specific clauses. All this administrative overhead acts as an economic tax. Just like digital payments effectively obliterated many costs associated with handling cash, AI will soon eliminate much of this menial unstructured work.

“What is your purpose?”

“You pass butter.”

Old Tools Are Handicapped

Every seismic shift in technology prompts a question about the right abstraction layer. When Intel focused on hardware, Microsoft took the lead in software. During the Internet era, Google dominated.

In this new Agentic age, many legacy solutions—like RPA (Robotic Process Automation)or No Code Tooling —promise AI wrapped on automation but fail to leverage AI’s deeper intelligence. These have been built for the old paradigm, automating only what was already there rather than rethinking it for advanced AI. The right solution taps into the power of these models by reinventing the software experience around AI, rather than grafting AI onto outdated systems.

When we started working on Decisional some time was spent some time initially creating no code draggable blocks to hard code logic since the models were inferior and we quickly realised with progress in post training and test time compute that so many decisions could be outsourced to intelligence on tap that we ended up re-focusing on the old adage “working backwards”.

During our time in the YCS24 batch we also learned from Greg Brockman, Sam Altman in addition to the OpenAI, Meta and Anthropic teams that their capabilities were rapidly evolving and we needed to aim for where the puck was going and not where it is today.

Enter the Age of AI Agents

This is why a new wave of inventions is emerging—platforms that sit on top of databases and modern software stacks while respecting one key constraint: human context windows are not expanding. If anything, our attention is pulled in more directions than ever. That makes a familiar, intuitive user experience crucial for adopting these tools at scale.

The most interesting AI product interfaces

• Canvas

  • Cove AI

• AI Spreadsheet

  • Matrices

  • Paradigm

  • Ottogrid

• AI Document

  • Capitol AI

  • Lex Page

Our Constitution for an AI Agent in Financial Work

At Decisional, we’ve spent a great deal of time considering the fundamental elements that define deep financial work on not just what will help us drive automation but bring superhuman excellence in this field. We spoke to hundreds of users in the past year from all kinds of Financial Services firms - from Investment Bankers at J.P Morgan, Portfolio Managers at Fidelity, Analysts at PE shops, Data operations managers at sub $50M funds and Chief Underwriting officers at Insurance firms across the front office as well as the back office to distill the most important elements of our AI Agent.

These are the pillars that power our vision of a “bionic analyst” of the future, a version of our constitutional AI that will reflect the alignment principles and values of the AI Agent we will build.

1. Transparency in Sources of Information - Reflecting the need of great analysts to sift through their sources of information to understand truth

2. Fact & Metric Focused - Emphasizing the importance of metrics and facts over subjective information

3. Built for Power Users who want Control - The best financial professionals want every edge that they can get instead of oversimplification

4. Enable Creativity of Financial Thought - Lastly AI now affords us a new paradigm enabling more creativity and fun in learning and work

The Neural Interface for Decisional: Source grounded Deep, Living Memos and Quant Magic Tables

In light of these pillars—and the limits of human attention—we developed two core artifacts that capture all unstructured data and workflows in our platform wrapped into a source grounded media

1. AI-Powered Memos – Living documents that adapt as you work, enabling narrative-driven insights and collaboration.

2. Magic Tables – AI spreadsheets designed to handle everything from financial modeling to analytics. Together, they represent the two sides of the bionic analyst: deep, thoughtful narrative creation and powerful number-crunching capabilities.

Journey so far

This year was an exciting moment for the team from founding the company to many neural pathway unlocking moments:

• Our co-founding team moved to San Francisco from working remote across London, NY and SF

• All founders got O1s and are locked in

• Got accepted into Character Labs and learned about design sprints from JZ, Jake and Eli

• Accepted into NVIDIA Inception and the LBS Incubator

• Made it to Y Combinator in our first attempt

• Raised a pre seed from some amazing investors like Character VC and Moonfire Ventures and Angels

• Landed our first customers and 10Xd our revenue during the last few months of the year

• Got featured in Bens Bites & Insight Partners Agent Ecosystem market map

Looking Ahead

We’re only just beginning this journey, but we believe our AI Agent for financial work will help financial professional rethink decision-making with AI. It is no coincidence that we named the company “Decisional”.

We will be sharing updates about our progress, use cases, and philosophy. Stay tuned for future blog posts as we delve deeper into the technology and vision behind Decisional. We are breaking up our posts into 4 categories to tailor to different themes.

• Fog of AI - Market Insights & Thoughts

• Bionic Finance - Augmenting Finance using AI

• Labs - Technical Deep Dives

• Journey - Sneak Peek into Decisional

Subscribe now