Desia Team
Product
November 13, 2025
Nobel laureate Daniel Kahneman famously described two modes of human thinking in his book Thinking, Fast and Slow: System 1 (intuitive, fast, automatic) and System 2 (reflective, deliberate, analytical). System 2 is our internal engine for planning, reasoning, and self-correction. It’s what separates a guess from a strategy.
For years, Artificial Intelligence has been stuck in System 1.
Until recently, large language models (LLMs) operated like a runaway train of thought. They generated text token-by-token, with each new word a probabilistic outcome based on the sequence that came before. When an LLM made a mistake or hallucinated, it would double down, producing outputs that were confidently and comprehensively wrong.
For finance, where precision is paramount, this was a non-starter. You can't build a valuation model on a hallucination.
This is where a new generation of thinking LLMs marks a fundamental shift. By incorporating techniques for self-correction and multi-step reasoning, these models can now engage in a System 2 approach. They can reason through steps, detect their own errors, and revise their logic before producing an answer.
At Desia, we’ve embedded this capability directly into our platform. Our users can choose between instant, intuitive responses (System 1) for relatively straightforward questions - e.g., "What was NVIDIA's revenue for fiscal year 2023?" - and deep, reflective reasoning (System 2) for more complex problems - e.g., "Compare the sentiment on supply chain challenges in the latest earnings transcripts for both Apple and Samsung.". The thinking mode trades a few seconds of latency for a leap in accuracy, making it essential for complex retrieval tasks.
But true intelligence isn't just about thinking; it's about acting. Analysts do not just reason about a problem but use tools to solve it, and this is where the thinking modes of LLMs, while powerful, hit a ceiling.
To break through this limitation, we don't need a single, monolithic super-tool. Instead, we need a new approach to how an AI can compose and combine a vast, ever-growing toolkit on the fly. We need a system where individual tools are simple, reliable specialists that can be chained together to produce powerful, emergent workflows.
This is the principle behind one of the most elegant and powerful philosophies in computing: Unix. Its core idea is simple but profound: make each program do one thing well, and combine them to achieve complex outcomes.
Consider this classic command to count unique students in a file:
cat students.txt | sort | uniq | wc -l
Each tool (cat, sort, uniq, wc) is a simple, reliable specialist. The pipe (|) links them into a powerful workflow.
This is the breakthrough of Agentic AI. The agent's own reasoning ability, its System 2 thinking, becomes the intelligent, dynamic pipe. Agents can analyze a goal, select specialist tools from their environment, execute them in a logical sequence, and use the output of one tool to decide which tool to use next.
Coding agents first sparked momentum by proving an AI could manage a complex workflow of developer tools. Now, we are applying the same power to orchestrate to the toolchain of a financial analyst.
We built Desia Agent Mode to extend beyond conversation and into end-to-end task execution, from understanding a user’s goal, to exploring data sources, to planning and carrying out the steps required to deliver a final output. Desia agents execute complex workflows by chaining together specialized tools built for modern finance, tackling challenges not as a single monolithic task, but as a sequence of precise, targeted actions.
One of the most immediate and powerful applications of this is in information retrieval. Where traditional approaches focus on text and semantic search, Desia agentic retrieval hones on the interaction.
Our agents interact with your data ecosystem as a human analyst would. They can browse folder structures, autocomplete file names, and learn organizational patterns; understanding, for example, that a specific folder likely holds the business logic for a query even if the file's name or content isn't an obvious keyword match.
Crucially, this solves one of the most persistent problems in knowledge management in finance: version control. Standard search engines often fail here, returning stale information because they can't distinguish between report_v2.docx, report_v10.docx, and report_vfinal.docx. As humans, we navigate this complexity easily, but traditional search can't. Agentic retrieval solves this by reflecting not only on the content but also on the context provided by filenames and folder structures.
This is the next frontier of search: an interactive, intelligent dialogue with your data.
Each new tool added to the Desia platform doesn't just add one new capability. It expands the agent's potential combinatorially, unlocking thousands of new, sophisticated workflows.
Ask a chatbot why a stock moved and you’ll get a paragraph. Ask Desia agents, and you get a plan, execution, and files you can audit.
We start with one prompt:
Analyze Cogent Biosciences (COGT) share-price movement post-Q3’25. Crawl the past week for analyst reactions, press releases, and clinical updates. Extract all quantitative metrics from the Q3 10-Q vs Q2, flag >10% moves. Package everything into an Excel with: Variance Analysis, External Reactions, Key Catalysts.
The agent’s first move isn’t to write, it’s to set a trail. It picks the source of record: research Q325 earnings and management commentary in our filings and transcripts database, while firing parallel pulls for analyst reactions and market data.
Before it stitches anything, it checks coverage: do we have ratings, price targets, and earnings data for the window? If something’s missing, the run stops here instead of papering over gaps. Only when the set is complete it drafts a report, with inline citations. That’s the first file: a clickable, source-anchored narrative you can review line by line.
Signals widen next. A finance search fetches fresh rating changes; a crawl gathers press releases, trial updates, and street notes. The agent waits for each research agent to finish, then merges drafts into a single report.
Now the numbers. Instead of re-typing tables, it pulls any metric with >10% quarter-over-quarter variance, cross-checking the cited 10-Q pages. If a subtotal doesn’t tie, it jumps back to the source page, retries a different parser, and re-calculates until the math is consistent.
Only then does it build what analysts actually use: an Excel workbook with three auto-generated tabs: variance analysis (Q3 vs Q2),external reactions and key catalysts.
What you get: a clear line from news to numbers to why-it-moved, plus the Excel you actually use. The point isn’t that the agent can write; it’s that it can decide. When to trust a filing over a blog. When to stop and fix a missing target. When to keep conflicting broker views instead of forcing one story.
We're at an inflection point. The leap from System 1 to System 2 AI isn't incremental, it's fundamental. When agents can think, reflect, and orchestrate tools like analysts, we move from AI that assists to AI that truly works for you.
The Unix philosophy taught us that elegance comes from composability. Now, that same principle is changing financial workflows. Every new tool we add doesn't just extend capabilities, it multiplies them. The platform you adopt today becomes exponentially more powerful tomorrow, without changing how you work.
Bring us your messiest workflow: the one with seventeen file versions, three data sources, and that formula no one wants to touch. We'll show you how Desia Agent Mode handles it.
Our team includes former bankers and analysts who've lived your pain. Reach out at info@desia.ai
Our large-scale natural language processing allows you to search, generate documents,and uncover answers to complex questions in minutes.
In Q3 2023, Company XYZ faced a major data breach, exposing customer information. The incident led to a 30% churn rate among enterprise clients over the next two months. Consequently, annual recurring revenue dropped from $30M to $23M by year-end 2023.
Revolutionize your firm's operations and join the leading institutions trusting Desia.
