From Chat Box to Control Tower
Designing agentic workflows around deterministic analytics
Trading and risk software has long come in two forms, a rigid standard product or a custom build that grows outdated quickly. Today’s AI chat tools add another limit. They start from an empty box and wait to be asked. Six senior Numerix practitioners see a new approach taking shape, where agents gather and propose, a deterministic engine calculates, and a person approves the result.
In this paper, the practitioners catalog eight agent patterns, from term sheets read in seconds to a fleet of agents monitoring convertible bond events, and how far along each one is. They also set out what has to be true at the seams, including how an agent is identified and scoped, what evidence it carries back, and where the release gate sits.
Read practical insights on:
- Eight agent patterns, from those already in daily use to those still in planning, and where a person signs off on each
- What it takes to connect AI tools to an analytics engine, including the Model Context Protocol (MCP), clean APIs, and documentation written for a machine reader
- Eight governance questions and the practices behind them, covering agent identity and scope, prompt injection, supply chain risk, and the kill switch
- Why the integration layer often takes more engineering effort than the calculations, and what AI-assisted development, unowned code, and rising model costs mean for the firms building it
- What a personalized, proactive workspace could look like in 12 months, with agents that know the portfolio, follow the news around its underlyings, and recommend which risks to recheck
“The future of AI in our space is as a wrapper and a facilitator around deterministic analytics that makes them easier to use, easier to challenge, and ultimately easier to trust.”
-Russell Goyder, PhD, Chief Analytics Officer, Numerix
In this series…
This white paper series, Trust, Verified, follows trust through the three places it needs to be won in capital markets. The first paper, Trust by Task, maps where AI reasoning has earned its place at the practitioner’s desk. The second, Beyond Plausible, sets out what it takes to prove a pricing or risk number when AI is in the workflow. This paper, From Chat Box to Control Tower, describes how firms are composing agents around their analytics and what has to be true at the seams.
Frequently Asked Questions
1. What is the focus of this paper?
This paper describes how capital markets firms are composing AI agents around their analytics, and what has to be in place to deploy those agents responsibly. It draws on interviews with six senior Numerix practitioners and catalogs eight agent patterns, each labeled by maturity, along with the governance and security practices applied at the seams.
2. What is an AI agent?
An agent is software that uses a model to choose and carry out steps through the tools it is given, within a bounded task. In the patterns the paper describes, agents gather information, translate between forms, propose, and monitor. A deterministic engine calculates, and a person approves what is released.
3. Does the paper suggest AI should replace the pricing engine?
No. Russell Goyder, PhD, Chief Analytics Officer, considers it a grave mistake to rely on AI for quantitative analytics with today’s models and tomorrow’s. The practitioners see AI as a wrapper around deterministic analytics that makes them easier to use, challenge, and trust, and every working pattern in the paper depends on the engine’s numbers.
4. Which agent patterns are already in use?
Patterns in use include translating a term sheet into a priced trade, desk and risk agents that query stored analytics results, client-built tools connected to the engine, and the translation of an agreed scenario into one the engine can run. Numerix also runs a fleet of agents that monitors convertible bond events for its Kynex products. Exposure analysis with external context and specialist agents with scoped tools are emerging, and an orchestrator over a mixture of experts is a target state.
5. How do the convertible bond agents work?
The agents have been taught the key sources to search and how to read indentures and term sheets. They report market events several times a day. They cannot write to the production database, because they sometimes hallucinate or present old information as new. A person reviews everything before it enters the system, which takes a few minutes a day compared with the full-day manual search the team used to run every seven to ten days.
6. What role does the Model Context Protocol (MCP) play?
MCP is an open standard for connecting models to tools and data sources, introduced by Anthropic in November 2024 and placed under a Linux Foundation body in December 2025. Numerix has put MCP servers in front of several of its applications. The practitioners found the server itself straightforward to build. The harder work is a clean API, clean documentation, and authentication set up properly.
7. What has to be in place before an AI tool connects to an analytics engine?
The practitioners point to four requirements. These are a clean API, documentation written for a machine reader, an identity for the agent with permissions enforced by the server, and a mark the engine applies to every number it produces. That mark lets a user see where the tool queried the engine and where it added an interpretation of its own.
8. What security risks do agentic systems introduce?
Nestor Nelson, Chief Information Officer, separates ordinary software faults from malicious activity. For faults, he wants visibility into what each agent is doing and a kill switch. For malicious activity, he names prompt injection, LLM poisoning, and supply chain risk, and the paper describes controls including a prompt proxy under evaluation, permissioned access channels, a governed process for third-party packages, and least-privilege access.
9. Will AI agents be trusted to execute trades?
Not yet, according to the practitioners. Satyam Kancharla, Chief Product Officer, compares delegating trading to an agent to granting power of attorney, and notes that many agents today can be jailbroken. Nelson expects AI-driven analysis to shape a strategy that is then executed through the traditional trading path, and Ilja Faerman, Head of Product Strategy, still sees a human in the loop at clients today.
10. What do the practitioners expect over the next 12 months?
They expect more agentic flows, with agents that analyze a portfolio, its market data, and its positions and flag issues before a user asks. Kancharla expects clients to want an experience tailored to their own strategy, assets, and counterparties. Across all three papers, the practitioners describe the engine as the trusted core, with a person still deciding what is released.
