Artificial intelligence is changing the way securitization practitioners operate, but work remains to make sure the technology is deployed safely.
Its capabilities — to code, summarise and aggregate information — promise to speed up chunks of manual work. It has already transformed software development. Work that developers would have taken months to do three years ago can now be completed in minutes.
“If it takes three people in the engineering team six months to build a tool and after all that it doesn’t really do what you want, that’s a bad look,” says Matthew Moniot, co-head of credit risk sharing at Man Group in London. “AI allows you to just have an idea, maybe only partially formed and roll it out nearly instantly. It turns the process of engineering on its head. You go from needing to be very pedantic and process driven to being able to iterate super fast.”
Others report the same experience.
“One of my issuer clients has millions of data points on their portfolio,” says Charles Cornelis, founder of CCSF Advisory. “They can iterate so quickly and produce useful insight in just a couple of hours using AI to write the code for the analysis.”
Agentic age
Structured finance professionals and their advisers are now using AI for more than writing code. Agentic systems — tools that can respond to inputs and take decisions over time — were too unreliable to be useful a year ago. That is changing.
“In a typical scenario, we could use an AI agent to take a previous deal from a shelf, the term sheet, and some client instructions for the next deal, and turn out a decent first draft of the documents,” says Thomas Quorrol, securitization partner at Linklaters and chair of the firm’s AI programme.
“It’s not perfect, and it’s not really the standard that a junior associate would provide, but it gets you a good chunk of the way there in a relatively short space of time.”
Given how fast AI systems are improving, it is worth looking at what they could make possible in the future.
Overworked investors could screen every deal they are shown. Structurers could make tweaks to deal documentation in seconds. Analysts could aggregate information from hundreds of sources. Issuers could respond to investor requests for information at speed regardless of format.
On February 5, Anthropic released its first Claude Opus model with a 1m token context window, five times bigger than previous versions. A model’s context window is the amount of information it can use while generating a response.
“The 1m token context window has made a really big difference for us. I think it was transformative for people in structured finance,” says Quorrol. “We went quite rapidly from being a bit frustrated with the output because you couldn’t cover the size of documents that we deal with, to be able to do something useful.”
The result of this should be that the cost of transacting falls, according to Luca Borella, who is CEO and co-founder of Algoritmica.ai and runs an annual structured finance hackathon to coincide with Global ABS.
“As one example, in the past you may have had to hire consultants for $100k to sanitise your data,” he says. “There are already some solutions in the market now that help you do that with just a few prompts.”
The possibilities excite Borella, but he also shows some caution.
“It’s easy to build a pilot. You can easily showcase what can be done,” he adds. “But especially if you are a bank, there’s a lot of work to be done before you can use AI in production.”
The consequences of getting it wrong depend on the use case, but, in general, building a product that a business can deploy safely requires humans to assess risks and build guardrails to mitigate them.
For more contained deployment, the mitigation can be as simple as having a human check the AI output. Building in automatic controls is also necessary.
“Best practice in software engineering is that it doesn’t matter if you believe your product works or not,” says Julian Mateu, founding engineer of Arc Analytics. “You need some fundamental automated tests to tell you if it passes or fails your specified criteria. That goes for both humans and AI.”
“Either you trust AI fully or you have a junior go back and double check that everything was grounded in the documents or in the data,” says Sam Griek, founder of Prospekto App. “For that, building an audit trail is super critical. You have to be able to go right back to docs and see where things came from.”
Time horizon of software tasks different LLMs can complete 50% of the time
The task-completion time horizon is the task duration (measured by human expert completion time) at which an AI agent is predicted to succeed with a given level of reliability. The 50%-time horizon is the duration at which an agent is predicted to succeed half the time. The graph below shows the 50% time horizons for frontier AI agents, calculated using their performance on over a hundred diverse software tasks.
Source: Model Evaluation & Threat Research (METR). Last updated May 8, 2026.
Repeated testing
Agentic systems require a new way of thinking about testing.
“You need to keep testing,” says Borella. “We are moving away from test and then deploy, and we are moving into a new era where you do continuous testing. Even the end user becomes a tester because of the non-deterministic nature of these models.”
Using this approach brings extra benefits. New generations of models can take the place of older versions, reducing dependence on a single AI source.
In certain sectors mistakes are intolerable and a stricter approach is required.
Luca Primerano is founder of Arc Comply, a firm that provides tools to detect various compliance risks like anti-money laundering, sanctions, fraud, and trade compliance.
The business uses AI in various forms, but Primerano says it must be deployed carefully.
“From a compliance point of view, you cannot afford to have black-box models, where recommendations come from something that you cannot explain,” he says. “We have developed our own models that provide additional information whenever there is a compliance risk. Models have to be explainable, it has to be clear which data are used to train the model, and the model has to be unbiased.”
Primerano says that means choosing how to approach problems carefully.
“The first layer is often defined using policies, which are deterministic, replicable, rule-based criteria,” he says. “Machine learning can slowly remove noise and explain why it has removed the noise. Depending on the problem, a traditional machine learning model can be enhanced with a language model within a broader system.”
Man or machine
Frontier large language models (LLMs) are progressing quicker than many businesses can build safety infrastructure around them. Firms are incentivised to deploy cutting-edge models ahead of their competition, but letting autonomous AI agents loose can be risky.
“The things we’re focused on [concerning] risk and security are around the agent inadvertently doing something that we do not want it to do,” says Quorrol. “How do you sandbox things? We’re increasingly going to come to agents that have access to your inbox.
“If an email comes in, they will read the email. There’s the risk of a prompt injection, meaning will somebody who’s nefarious put an instruction in the email which the agent acts on? It’s those risks that we’re really focused on, because I think they’re a real concern. It comes down to kind of which tools do you deploy, and how do you deploy them?”
Accidents in other sectors highlight the importance of sandboxing. The Guardian reported in April that an AI had deleted the entire production database of a firm called PocketOS.
The concerns around trust are just one of the reasons humans and AI will be working alongside each other for the near future.
“We have spent 10 years obsessing about our data,” says Man Group’s Moniot. “We have really good high quality data. You’ve got to have people that really know their stuff to teach the AI. There should be no expectation that as good as it is, if you haven’t trained it to understand how to read documents, that it’s just going to do it perfectly.”
Conversely, if deployed correctly, AI can help humans maximise their opportunities.
“There are cases where founders of funds are using Claude directly, and building cool, bespoke reports,” says Griek. “When you have the founder doing that, it’s a big waste of their time. They should be making the key decisions.”
Working together
Indeed, good human judgement remains a key part of deploying AI at every stage of the process.
“I think it’s a skill to decide whether a project is a good one to delegate to AI or not,” says Arc’s Mateu. “I don’t want to delegate thinking to AI. I want it to help me think better.”
Businesses have to strike a balance. Move too slowly and competitors may be able to muscle in but being hasty or thoughtless is risky.
“When I talk with clients, I like to find out what parts of their system they consider their edge,” says Griek. “It’s important to protect that IP. Whatever we do, they need to own that completely. For example, we wouldn’t want to put that in a third-party tool.”