@robert.myers, Head of the io.Intelligence Unit, was the guest on the latest Waters Wavelength Podcast by WatersTechnology.
Listen from the link or below.
If you don’t have time to catch the full episode on a Friday, here are some key takeaways.
AI is Becoming Proficient in Finance-Related Tasks
Frontier models can now pass rigorous financial exams like CFA Level 3 with high accuracy. This signals that AI is no longer just a novelty, it’s capable of performing entry-level analyst tasks, such as:
- Running discounted cash flow (DCF) models
- Reviewing earnings reports
- Finding comparable valuations
This means AI can augment junior roles in finance, freeing up human analysts for higher-level strategic work. However, Bob cautions that while models are impressive, they’re not perfect with some benchmarks still showing only ~55% accuracy, so human oversight remains essential.
Too Many Tools Can Overwhelm AI
Giving AI access to hundreds of tools (via APIs or MCP) actually reduces its effectiveness.
- AI struggles to choose the right tool when options are too broad.
- Outcomes become non-deterministic - the same request might yield different results each time.
The solution? Human-curated tool bundles. By composing tools into logical, task-specific sets, AI can perform more reliably. There’s a “sweet spot” around 50 tools => beyond that, performance drops.
Interoperability is Foundational for AI Success
Organizations that have already invested in interoperable systems (e.g., using FDC3 standards) are ahead of the curve. These systems:
- Declare app capabilities in directories
- Pass context and intents between apps
This metadata is exactly what AI needs to function effectively. So, rather than starting from scratch, companies can build on existing infrastructure to integrate AI smoothly.
AI Should Enhance Existing Workflows, Not Replace Them Blindly
Don’t view AI as a business process automation tool. Instead, it should:
- Accelerate validated workflows
- Help users navigate complex systems
- Provide assistive intelligence, not full autonomy (yet)
For example, AI can guide users to the right app or dataset, but humans should still validate outputs, especially in high-stakes environments like finance.
Centralized Copilots Are Better Than Fragmented Chat Experiences
A major UX concern: if every app has its own AI chat window, users will be overwhelmed. Bob recommends:
- A single enterprise copilot that spans all apps
- Unified access to tools and data
- Controlled, permission-aware interactions
This mirrors how companies centralized alerts and notifications in the past. A centralized copilot improves governance, usability, and trust.
Visualization Still Matters
Chat interfaces are not ideal for exploring large datasets. Often the better path is:
- Directing users to trusted visual tools (e.g., dashboards, charts)
- Using AI to guide (not replace) data exploration
Instead of summarizing a time series in text, AI should open a charting app with the relevant data. This preserves data integrity and improves decision-making.