From: aidotengineer
Due diligence in finance workflows is a complex and demanding process, particularly in competitive deal environments where speed and accuracy are paramount [00:00:27]. This often involves digesting vast amounts of content, identifying critical risk factors, and building conviction quickly [00:00:33].
Challenges in Manual Due Diligence
The nature of financial due diligence tasks makes them “non-trivial” for human analysts [00:00:45]. Challenges include:
- Large Data Volumes Thousands of pages of content need to be reviewed in data rooms [00:00:33].
- Tight Deadlines Junior analysts are often tasked with impossible goals under severe time constraints [00:01:37].
- Cognitive Demands Understanding market dynamics at both sector and individual ticker levels, or sifting through hundreds of vendor contracts to spot early termination clauses, is cognitively demanding and time-intensive [00:01:02], [00:01:10], [00:02:30].
- Human Cost The manual nature of this work can put professionals “in a meat grinder” due to the high stakes and human cost involved [00:01:32], [00:02:04].
Examples of Due Diligence Scenarios
- Competitive Deal Process Assessing thousands of pages in a data room pre-term sheet to gain conviction quickly and spot risk factors that could diminish asset performance [00:00:27].
- Mutual Fund Analysis During earnings season, analyzing calls, transcripts, and filings for a universal coverage of 80 to 120 names [00:00:47].
- Confirmatory Diligence Reviewing hundreds of vendor contracts to identify early termination clauses or understand thematic negotiation strategies across a portfolio [00:01:08].
The Role of AI in Financial Due Diligence
AI, particularly knowledge agents, are transforming financial due diligence by allowing professionals to bring greater sophistication to analysis [00:02:50]. Companies like BrightWave build research agents specifically designed to digest very large corpuses of financial content [00:00:22].
This shift mirrors the impact of spreadsheets on accounting, where the computational tools allowed for more effective and efficient thinking, elevating the job beyond mere “running numbers” [00:02:22].
Benefits of AI Agents
AI systems can:
- Digest Volumes of Content Rapidly process vast amounts of data that would be impossible for humans [00:03:10].
- Accelerate Efficiency Perform meaningful work that accelerates efficiency and time to value by orders of magnitude [00:03:14].
Design Considerations for AI in Due Diligence
A key design consideration for these systems is how to reveal the AI’s thought process to a human user in a useful and legible way, especially when the AI has considered thousands of pages of content [00:03:40].
- Transparency The final form factor for presenting AI findings is still evolving, as simple chat interfaces are likely insufficient [00:03:57]. Products need to scaffold workflows and shape system behavior to specify user intent, removing the burden of complex prompting [00:07:31].
- Mimicking Human Decision-Making Autonomous agents should mimic the human decision-making process, decomposing tasks like finding public market comparables, assessing relevant document sets, distilling findings, and enriching/error-correcting them [00:08:00].
- Intermediate Notes Capturing intermediary notes—what the model believes based on its findings—is extremely useful for transparency and auditability [00:08:55].
- Self-Correction Models can self-correct for factual accuracy or entity recognition by asking “is this factually entailed by this document?” or “is this actually an organization?” [00:09:22].
- Synthesis AI can weave together disparate fact patterns across many documents into a coherent narrative [00:09:55]. However, current models have limitations in producing very long, coherent, and novel outputs, making it beneficial to decompose research instructions into multiple sub-themes to achieve higher-fidelity, more information-dense results [00:13:31], [00:14:27].
- Granularity Being more granular and specific in instructions leads to higher quality outputs [00:14:24].
- Limitations State-of-the-art models still face limitations in managing complex real-world situations, such as temporality (e.g., changes after a merger and acquisition or contract addendums) and combining disparate fact patterns across many documents [00:15:39], [00:15:57].
- Human Oversight and “Taste-Making” Human oversight remains crucial. Analysts can nudge the model with directives, pull interesting threads, and leverage their access to non-digitized information (e.g., conversations with management, portfolio manager opinions) to guide the AI [00:10:04], [00:10:19]. This “taste-making” is where powerful products will lean [00:10:31].
- User Interface Interfaces should offer a continuous surface where users can click on citations for context, interact with structured outputs, highlight any passage to ask for implications, and “drill in” for details on demand [00:17:37]. This provides an audit trail for the system’s findings [00:18:59].
The Latency Trap and User Experience
The “latency trap” highlights a challenge with agentic systems: if the feedback loop for user interaction is too long (e.g., 8-20 minutes), users cannot develop proficiency with the system quickly, leading to low faculty with the product [00:12:50]. The product experience should allow users to refine their mental model of how their prompts elicit behaviors from the models [00:12:43].