Generative artificial intelligence (AI) can astound with human-like conversation and research superpowers, but it can also be disastrously wrong. That unreliability is a major barrier to adoption as AI begins to control physical machines, payments and other high-stakes domains.
Neurosymbolic AI is a rising technique that aims to solve the hallucination problem. It’s a hybrid approach that marries gen AI’s neural networks for pattern recognition with symbolic reasoning’s logic that can prove whether outputs are correct, with mathematical precision.
The goal is to maintain the flexibility of generative AI while making its outputs trustworthy.
Trust is a key factor for greater adoption of generative and agentic AI, according to a July 2025 PYMNTS Intelligence report. While chief financial officers (CFOs) have warmed up to generative AI, they are still unsure whether they could trust agentic AI. Just 15% of executives are considering deployment.
Neurosymbolic AI could help. Its roots stem from the early 20th century, when mathematicians developed ways to formalize reasoning using symbolic logic, laying the foundation of modern computing. Out of that tradition grew automated reasoning, a subset of symbolic reasoning that uses logic to prove the correctness of computer programs.
Amazon has become one of neurosymbolic AI’s most prominent adopters, according to The Wall Street Journal. Its Automated Reasoning Group, founded more than a decade ago, built tools to verify security policies in the AWS cloud. Those methods now underpin new systems such as the Vulcan warehouse robots, which combine neural networks for perception with automated reasoning for precise planning.
Advertisement: Scroll to Continue
Amazon also applies neurosymbolic AI to customer-facing products. The Rufus shopping assistant, for example, uses large language models for conversation but leans on automated reasoning to ensure recommendations align with rules and policies. In August, Amazon announced an automated reasoning feature to minimize hallucinations, claiming it can identify correct model responses with up to 99% accuracy.
“This is emblematic of where the field is going,” Imandra Co-founder and co-CEO Grant Passmore said in an interview with PYMNTS. “Fundamentally, just large language models on their own cannot be trusted to reliably reason. On the other hand, you have this incredible and hitherto underutilized field of AI — which is automated reasoning — completely built on logic, completely precise” and independently audited.
See also: The Two Faces of AI: Gen AI’s Triumph Meets Agentic AI’s Caution
Financial Services Industry Needs Certainty
Amazon’s automated reasoning group was founded by Byron Cook, its former distinguished scientist. About 20 years ago, he and Passmore crossed paths at Cambridge University in the U.K.
In 2014, Passmore and Co-founder and Co-CEO Denis Ignatovich — a college friend who ran Deutsche Bank’s equities trading desk in London — founded Imandra to apply neurosymbolic logic to finance.
Ignatovich’s constant concern was that a developer working on the trading system could inadvertently omit a line of code that causes the system to violate a regulation — and it becoming a big issue. At the time, Passmore was working on new techniques for reasoning in the U.K.
“We realized we could actually commercialize this stuff,” Passmore said. “Finance needs it. Financial regulators need it, and finance can pay for it.”
Founded in 2014 and incorporated in the U.S. in 2019, Imandra has raised $23 million from investors including Citi, Green Visor Capital, Albion, IQ Capital and Austin’s LiveOak Venture Partners. Goldman Sachs is a client, and Citi led its most recent round.
One use case is automating the FIX (Financial Information eXchange) protocol, which lets organizations send electronic trading instructions to each other, say from Goldman Sachs to BlackRock or the New York Stock Exchange. “Every computer in capital markets trading uses a version of this protocol, but everybody is allowed to customize it,” Passmore said.
Typically, these quirks are described in 100-page PDFs. To connect, companies had to interpret the manual, build their own client code and then test repeatedly to make sure both sides can connect without errors.
Imandra built a mathematically precise language for specifying the FIX protocol. Instead of relying on hundreds of pages of PDFs, clients can now use automated reasoning to ensure precise communication each way.
“This very error-prone process at the plumbing level of the markets used to take sometimes upwards of six months” for a new participant to be onboarded, Passmore said. “Now they can be onboarded within three days with mathematical guarantees.”
To scale its approach, Imandra launched Imandra Universe, a platform it describes as the first marketplace for neurosymbolic agents. Similar to Hugging Face’s repository for machine learning models, it hosts symbolic reasoning engines specialized in domains from geometry to logistics. A “Reasoner Gateway” lets developers plug them into agent frameworks, augmenting AI systems with logical checks.
One flagship tool, Code Logician, targets the flood of AI-generated code. Once installed in coding assistants like Cursor, Code Logician builds a mathematical description of what the AI generated code is doing and use Imandra to verify it. Passmore said around 60% of AI-generated code contains bugs and Code Logician can make 96% of the code correct within three iterations.
Imandra is expanding Code Logician beyond Python to Java and even COBOL, reflecting enterprise demand for code migration. It is also developing agents for geometry reasoning, among other pursuits.
Read more:
AWS Turns to Ancient Logic to Tackle Modern AI’s Hallucinations
AI’s Dual Nature: Reasoning Models Emerge as Key Differentiator for Business
Report: Reasoning AI Models Fail When Problems Get Too Complicated
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
