At first glance, managing context for AI agents looks like a purely technical problem. But when you dive deeper—especially in financial applications—it quickly becomes clear that it's also a regulatory and business challenge.
While building Gideon the Bull, the AI guide in my educational investment game, I ran into this issue directly.
For Gideon to provide meaningful advice, he must account for a wide range of information sources:
- 1️⃣ Client profile – demographics, risk tolerance, goals, and financial background
- 2️⃣ Portfolio and actual transactions
- 3️⃣ Interaction history – past conversations (CRM)
- 4️⃣ Reliable real-time market data from providers
- 5️⃣ The rules of the game (think of it as the trading framework)
- … and far more beyond that.
Each of these is a rich domain in its own right.
And on top of them sits the economic valuation model itself… which goes well beyond LLMs (DCF, DDM, and possibly more complex approaches).
⸻
🦾 The Challenge
AI has hard limits on the amount of context it can process efficiently.
Using all the information is inefficient on several fronts: response times get longer, costs increase (token usage), and accuracy doesn't necessarily improve—sometimes the "short blanket problem" means not all critical information fits into the usable context.
Irresponsibly narrowing context, on the other hand, risks omitting key information—leading to poor or even misleading advice.
Why is context management a regulatory challenge in AI financial services?
In the real world, giving advice without considering all relevant information can amount to providing unsuitable or incorrect guidance—and in some circumstances, even a breach of fiduciary duty. That's why context management is not just a technical issue—it's also a regulatory and business core challenge.
In the real world, giving advice without considering all relevant information can amount to providing unsuitable or incorrect guidance—and in some circumstances, even a breach of fiduciary duty.
That's why context management is not just a technical issue—it's also a regulatory and business core challenge.
⸻
🚀 Current Market Reality
No wonder we don't yet see true AI-based fiduciary investment advice in the real world.
It's not that there aren't attempts: eToro and Dr. Hedva Ber have announced AI-powered investment companions (not advisers 🤔), and Viola FinTech has backed startups in the space, such as Quinn or Magnifi by TIFIN, which operates under U.S. licensing.
Still, these are relatively few compared to what one might expect—and one of the key reasons is the context problem.
⸻
🔮 What's Next?
In my next post, I'll share how I approached this problem (and how it connects back to the real-world challenges of financial advice).
And generally—
If you have technical or business insights on this aspect, I'd love to hear them.
More about the challenges of building an AI-powered personal investment adviser in the context of my educational game and the development process can be found here:
⸻
🔗 Is AI Ready for Fiduciary Investment Advice?
🔗 RIAs: Time to Adopt AI
🔗 On the Bias–Variance Tradeoff