Agentic AI meets the autocallable: From term sheet to priced trade in 18 minutes
Agentic AI has rapidly moved from novelty to real-world implementation across the capital markets. Evidence of this shift can be seen in the Cambridge Centre for Alternative Finance's 2026 Global AI in Financial Services Report, which found that 52% of industry respondents are already actively adopting agentic AI.
For many quant teams, the focus is shifting from whether to use the technology to how to use it within a pricing workflow that must be accurate, auditable, and reproducible. In a recent Numerix webinar, Twenty Years of Autocallables Meet Agentic AI, Dr. Ping Sun, SVP and Head of Quantitative Research at Numerix, put that question to a live test.
Sun asked an AI agent to turn an autocallable term sheet, chosen by the audience and not read by him in advance, into a working, priced trade template using the Numerix CrossAsset Python SDK. The results showed how much of the build an agent can handle today, and exactly where quants still need to step in.
A live build from an unread term sheet
Attendees chose a publicly available term sheet listed on Borsa Italiana. Sun gave Claude a short natural-language prompt asking it to read the PDF, create a CrossAsset Python SDK template using available sample market data, price the trade with a local stochastic volatility (LSV) equity model and a two-factor Hull-White rates model, unpack the calculation for inspection in Excel, and write a walkthrough document.
About 18 minutes later, the agent had identified the note as a Barclays-issued Phoenix autocallable, written the payoff in the Numerix payoff scripting language, built and calibrated the requested models, produced a present value, and generated the documentation. Reviewing the output live, Sun confirmed the payoff logic was accurate for the coupon, autocall, and final knock-in observations. He also flagged what was missing. The agent omitted the LSV mixing fraction and fell back to a default, and used placeholder correlation values instead of calibrated ones. Sun described both as easy fixes once the template exists and noted that agents often skip functionality that is not explicitly requested.
Built from the documentation clients already have
Prior to the webinar, Sun gave Claude the same documentation and examples a Numerix client receives, including the CrossAsset Object Reference, Advanced Risk Guide, Scripting Guide, Python SDK API references, and SDK Quick Start examples. Sun was surprised by how well it seemed to understand the library from those documents alone.
From those inputs, the AI model had produced a knowledge base that distilled product and API knowledge into plain-language files, a set of validated code examples with explanations, and a library of skills. The skills defined step-by-step workflows, such as how to write a working payoff script by defining observation events, the underlying indices, the payoff logic, the trade data, and the pricing call. Sun said building that foundation took several days and cost less than $150 in AI usage. Because every file is human-readable, quants can correct or extend it the way they would review a colleague’s notes.
Sun noted that AI models already follow many good coding practices, but not always consistently. In his early testing, the agent explained its payoff scripts with comments only some of the time. To make every script easy for a quant to review, Sun updated the skill to require an introductory summary at the top of each script and inline comments throughout.
The engine still produces the number
An early audience question asked whether the pricing itself was AI-based, and Sun confirmed it was not. The agent builds the template and calls the library, while the valuation runs on CrossAsset's established numerical methods, in this case Monte Carlo simulation. Sun noted that Numerix is researching neural network-based valuation for autocallables, but that work is not yet in the product.
Greg Murray, SVP of Growth Marketing at Numerix and the webinar’s moderator, noted that some firms are trying to build AI-based pricing engines, and that several people at Numerix have tried as well, only to find it much harder than expected. What Numerix is seeing across the market, including from regulators and audit teams, is that AI is best used on top of a deterministic engine whose numbers can be audited and reproduced. That principle is at the core of Numerix's Trust, Verified white paper series, which includes Trust by Task, Beyond Plausible and From Chat Box to Control Tower.
Guardrails, hallucination, and the last 10%
When an audience member asked whether the AI agent typically invents inaccurate product terms, Sun said he had not seen any hallucination in recent work. In his view, the library-specific documents and examples act as guardrails, and once the agent captures the formulas from the term sheet there is little room to improvise. He noted, however, that trouble may arise with library functionality that has no worked examples for the AI model to learn from. The one case he encountered in the previous six months involved a function that uses a Brownian bridge to monitor barriers continuously. The agent could not make it work from documentation alone, recognized the failure, and kept trying until Sun supplied guidance. Once a reviewed example exists, the agent can absorb it into the knowledge base for future work.
Sun estimated that at least 90% of a generated Python script is accurate on the first pass. The remaining 10% is where quant and financial engineer expertise is key, catching desk preferences and nonstandard requirements that a PDF term sheet does not spell out. Even with that review, he said, the work shrinks from days or weeks to a few hours at most.
What all this means for quant teams
The demo compressed new product onboarding, pricing, and documentation into a single session. The agent’s walkthrough document, which summarized the trade terms, payoff structure, model setup, and pricing steps, drew particular notice. Asked whether a human financial engineer would document a valuation that thoroughly, Sun said he would be impressed if a human quant could do it in two days, while the AI agent did it in minutes. Sun added that the same agent can run scenario analysis from a plain-language request, such as repricing across a grid of dividend shocks, work that previously required a custom script.
For structured products desks, where issuers regularly introduce new features to meet client demand, turning days of work into hours has real value from a time-to-market perspective. The session also made it clear that speed and control go together only when a validated analytics library is the foundation of the calculation and a quant reviews everything the agent produces.
Get deeper insight
Watch the full replay of Twenty Years of Autocallables Meet Agentic AI to see the complete live build, from term sheet to priced trade, along with the full Q&A.
For additional insight into the autocallable modeling portion of this webinar, read our blog on Sun’s discussion of autocallable structures, risks, and model choice.
Frequently Asked Questions
1: How do structured products desks onboard a new autocallable without weeks of custom scripting?
Structured products desks face new autocallable features from issuers regularly, and onboarding each one has traditionally taken days or weeks. In the Numerix webinar, Twenty Years of Autocallables Meet Agentic AI, Dr. Ping Sun, SVP and Head of Quantitative Research, had an AI agent turn an audience-chosen term sheet he had not read in advance into a priced trade template using the Numerix CrossAsset Python SDK in about 18 minutes. Sun says that even with quant review, the work shrinks to a couple of hours at most.
2: How accurate is an AI-generated autocallable pricing script, and where does a quant still need to step in?
Quant teams need AI-generated pricing code that is accurate, auditable, and reproducible. According to Dr. Ping Sun of Numerix, at least 90% of a generated CrossAsset Python script is accurate on the first pass. The remaining 10% needs quant review for desk preferences and nonstandard requirements that a PDF term sheet does not spell out. In the live build, the agent omitted the local stochastic volatility (LSV) mixing fraction and used placeholder correlations, which Sun described as easy fixes once the template exists.
3: Should AI price an autocallable directly, or sit on top of a deterministic valuation engine?
Many firms are trying to build AI-based pricing engines, and Greg Murray, SVP of Growth Marketing at Numerix, said several people at Numerix found it harder than expected. In the webinar Twenty Years of Autocallables Meet Agentic AI, the agent built the template and called the library, while CrossAsset Monte Carlo simulation produced the valuation. Numerix is seeing across the market, including from regulators and audit teams, that AI is best used for now on top of a deterministic engine that can be audited and reproduced. Neural-network valuation is still in research.