SATIRARASI | INTERVIEW · English edition
Interview by Mustafa Azizoğlu · 14 October 2026
Read the original interview in Turkish / Türkçe röportajı okuyun
Andy Simpson, Founder and CEO of SIGMA AI, who has nearly 30 years of experience in financial markets, says the biggest challenge in using artificial intelligence in finance is not technology but reliability. “A demo only has to work once. A production system has to work every time,” says Simpson, warning that AI will not eliminate accountability for financial decisions: “Regulators will take people to court, not AI.”
Simpson is not merely an industry executive making predictions about the future of AI. SIGMA AI, the company he founded, develops technologies that analyze news feeds and financial market data in real time to detect unusual movements and generate predictions in areas such as market volatility and the probability of trade execution.
Simpson is also Head of AI and Innovation at Trading Technologies (TT), a global financial trading technology company, where he works on bringing AI applications into the operations of major financial institutions.
Speaking to Satırarası during Istanbul Fintech Week 2026, Simpson argues that financial institutions should not overlook the importance of traditional machine learning and reliable data while focusing on the capabilities of large language models. Explaining that SIGMA builds much of its analytics on repeatable and auditable systems, he emphasizes that the real competition will be less about developing more powerful AI models and more about using them to solve real business problems.
Simpson, who also sees Turkey as an important growth market for SIGMA, discussed the limitations of AI in finance, real-time market analytics and future applications of the technology.
You have nearly 30 years of experience in financial markets. What is the difference today between an impressive generative AI demo and a system that can genuinely be trusted?
We have seen many waves of technology in financial markets before. Early distributed ledger technology (DLT) projects were often technologies looking for a problem to solve. The applications that survived, such as tokenized repo transactions, addressed real problems. We need to be careful not to get caught up in the same kind of excitement with generative AI.
A demo only needs to work once. At the proof-of-concept stage, you begin demonstrating that a system is scalable. But once you move into production, the situation changes completely. The system must work consistently and deliver the results the customer expects.
And producing results is not enough. Those results must be verifiable, auditable, repeatable and compliant with regulatory requirements. It is easy to replicate a simple large language model application, but much harder to trust it.
Why, then, do companies so often fail to move beyond the pilot stage?
Because they often do not start by defining the expected business benefit. They have a wonderful technology and are now looking for a problem that fits it. It should be the other way around.
Traditional project management had two very important concepts: return on investment and benefits realization. Both had to be measured and monitored. We should apply the same approach to AI.
Of course, we can build a demo and demonstrate what the technology can achieve. But ultimately, we need to solve a real business problem and deliver a measurable improvement.
Last year, you said that data quality was one of the biggest challenges and that a single trusted source of data was essential. Has this become less important as models have improved?
No. Today, we need to distinguish between three things: the quality of the data, the model itself and the inference capabilities we derive from the model.
Models have become extremely powerful. Of course, they continue to improve, but we have reached a point where differentiation does not always come from the model itself. What matters is which data you use and what you can derive from it.
There is an old rule: garbage in, garbage out. The same applies to prompts. You cannot expect a high-quality result from an incorrect or poorly defined prompt. That is why a reliable, accurate and consistent source of data remains fundamental.
Can generative AI handle all these processes? Do we still need traditional machine learning?
We should not think of generative AI as the entirety of artificial intelligence. Machine learning, including neural networks, and agent-based systems are other important components of AI. In financial markets, some tasks require deterministic results, meaning results that can be repeated according to defined rules, while others require probabilistic outcomes.
Large language models are becoming increasingly capable in analytical tasks. But we should remember that they are fundamentally systems based on predicting words.
A machine learning system developed for a specific task and trained on an institution’s own data may perform that task much better. I see these capabilities not as alternatives, but as technologies that should be used together.
How does this division of labor work within SIGMA’s AIMA platform?
Our approach is roughly 80–20. About 80 percent of our analytical processes are carried out by repeatable, auditable machine learning and rule-based systems that we have developed and control. We use large language models for approximately 20 percent.
At this stage, we use large language models primarily for natural language generation. That is because we know they still have weaknesses in mathematical operations, executing algorithms and hallucinations.
AIMA functions both as an AI assistant and a report-generation tool. It also incorporates sentiment analysis, the examination of market behavior and predictive models. Our goal is to uncover insights that humans could not produce at the same scale.
What kind of data infrastructure do you use for sentiment analysis?
We have an infrastructure that analyzes approximately 1.4 million news stories and articles every day. We continuously analyze content from various data providers, including Opoint.
The goal is not simply to collect news. It is to understand which themes these stories point to, how they influence market sentiment and how they can be linked to behavior.
We do this using machine learning, sentiment analysis algorithms we have developed and other techniques. This allows users to monitor changes in the markets much more closely rather than simply receiving a report about the past.
You say that you assign limited tasks to large language models. How do you decide which model to use?
Our platform is not dependent on any particular large language model. We can use different models. We access various models through AWS Bedrock and select them according to our needs.
For example, we might use Anthropic’s Haiku model for a short response. Sonnet or Opus might be more suitable for another task. We carefully evaluate their capabilities and costs.
Even if the cost of an individual token falls, total inference costs can rise because models are being used much more intensively. So it is important not only to deliver good results to customers, but also to do so economically.
How can financial institutions ensure the accuracy of AI-generated outputs?
Absolute care is required, particularly with regulatory texts. Imagine conducting a gap analysis of European financial regulations and citing specific provisions. Those references must be absolutely correct.
We cannot assume that large language models will always get this right. That is why we can use mechanisms we call AI Judges, in which independent models or algorithms check the results.
The important thing is not to trust the answer given by a single model, but to test that answer using other control mechanisms.
Who will be responsible when an AI system makes a wrong decision?
A regulated financial institution cannot delegate its compliance obligations or accountability to an AI system.
We know that large language models have weaknesses. Banks and other regulated institutions therefore need to be extremely careful about relying on them entirely.
Ultimately, a regulator is not going to take an AI model or a generative AI system to court. It will take people to court. That is very important to remember.
You have been Head of AI and Innovation at Trading Technologies for more than a year. What has changed during this period?
Much of the work carried out in large organizations is not visible from the outside. You can think of it as an iceberg. Ensuring that AI is used consistently across an organization and that employees can use these tools properly requires a great deal of effort.
We have taken a very pragmatic approach. We have focused on developing organizational maturity and consistency.
There are now many products being prepared for launch. I cannot share their details yet. But we are being particularly careful about protecting customers’ proprietary data. Ensuring that one customer’s data is not mixed with another customer’s takes time.
Turning to SIGMA’s partnership with BMLL, how do you combine deep historical order book data with real-time market data? What stage has the solution reached?
BMLL’s extensive historical order book data provides us with a benchmark for understanding what is normal. Our real-time system then identifies how current conditions differ from those historical norms.
This allows us to move beyond historical analysis and detect unusual conditions as they emerge. The combination of the two systems is extremely powerful.
Since founding SIGMA AI, we have focused on two concepts: immediacy and personalization. BMLL’s data helps us deliver exactly that.
We have a production-ready system and are working with customers on their requirements.
What tangible benefits does real-time analytics offer financial institutions?
In traditional systems, you sometimes have to wait until the next day for data to be processed and reported. Our goal is to make information available at precisely the moment it is needed.
With real-time data, we can detect unusual market movements, generate alerts and develop predictions about future conditions.
For example, we can generate predictions about volatility, bid-ask spreads, order flow toxicity or the probability that a trade will be executed. Historical data provides a benchmark, while real-time analytics tells us what has changed.
AI can also be used to predict price movements. Do you think that is where the real value lies?
The real issue is not simply predicting that a price will rise. If everyone has access to the same model and uses the same prediction, the return advantage it provides disappears very quickly.
I think what is more interesting is the ability to predict market behavior and risks. The probability of trade execution, the toxicity of market flow and valuation behavior are all very important.
Particularly in less liquid financial instruments, yield curves, swaps and risk surfaces, we can perform calculations at a scale that was not previously possible and develop predictions about future conditions.
A year from now, which AI applications do you think financial institutions will be using that they are reluctant to adopt today?
I think we will see many applications that are currently at the demo stage become real platforms and be evaluated in terms of return on investment.
Regulators will also gain greater clarity on how human oversight should be maintained, which decisions must be made by humans and how accountability should be preserved.
Institutions will have established their own AI policies, boundaries and control mechanisms. I believe the difference will come not from the new models themselves, but from the use cases implemented with them. For example, I am already genuinely impressed by the capabilities of some of the predictive models we are developing for our customers. In these models, we use generative AI not for the analytical intelligence of large language models, but to communicate our real-time updates and predictions in natural language.
I believe this approach will evolve with Inference AI, processing increasing amounts of real-time data and delivering much more powerful capabilities.
Does SIGMA have any customers in Turkey? How do you view the Turkish market?
We do not yet have any customers in Turkey. That is one of the reasons I came to Istanbul Fintech Week. I have many meetings here, and I see Turkey as an important market for SIGMA.
Turkey’s growth performance and industrial development are impressive. Activity in its financial markets is also significant. I am excited to be in Istanbul and to meet people working in this field.
What are your impressions of financial technology and the existing infrastructure in Turkey?
As in every other market where we operate, modern technologies and legacy systems are used side by side in Turkey.
Technologies such as end-to-end automation and straight-through processing are in use. At the same time, there is strong interest in investment and trading activities in Turkey.
The areas that interest me are providing investors with better analytics and predictions, improving the accuracy of hedging activities and helping people manage risk more effectively. I see significant opportunities in these areas.
Finally, how do you see the future of autonomous AI agents in the financial sector?
We have allowed machines to execute trades automatically in financial markets for many years. We already have certain control mechanisms in place for automated trading, clearing systems and straight-through processing.
I believe the same discipline now needs to be applied to agent-based AI. We should not see these systems simply as a new technology. We should draw on our existing experience to establish the necessary boundaries.
Another important development is the growing interest in sovereign AI capabilities in Europe and the United Kingdom. Where the next generation of models will be developed and how they will be used in financial systems will be among the most interesting questions in the period ahead.
About the journalist
Mustafa Azizoğlu is the founder and editor of Satırarası, an independent B2B publication based in Turkey covering banking, fintech and enterprise technology. He is a journalist and former editor at Newsweek Turkey.
