A leveraged buyout model built by an AI system can look finished. The statements tie together, the debt schedule rolls forward and the returns land in a range any sponsor would be happy to present. By most automated measures, the work passes.
To someone who has spent years building those models by hand, it can still be obviously wrong. The capital structure may be one no lender would offer. The interest rates may sit far from where the market is pricing debt today. An unsupported EBITDA adjustment may be quietly inflating the valuation built on top of it. Every formula can be correct while the transaction itself would never get done.
That gap, between financial work that looks right and financial work that would survive a senior banker’s review, is the problem Taron Arshakian has built Proxion around.
“Often, the more convincing an AI-generated analysis looks, the harder its underlying mistakes can be to detect,” said Arshakian, the company’s founder and chief executive.
Grading the answer is not enough
When the Bank of England and the Financial Conduct Authority surveyed 118 UK financial firms in 2024, 75% were already using AI. The most common way those firms monitored model effectiveness was through accuracy-style metrics such as precision and recall, cited by 88% of them, ahead of operational efficiency and model robustness. Among firms using or planning to use AI, only 34% said they had a complete understanding of the technology.
Arshakian argues that accuracy, however carefully measured, cannot on its own show whether financial work can be trusted. A model may calculate the returns perfectly without questioning whether the debt structure is realistic, the underlying assumptions are defensible or the transaction makes economic sense. In finance, work is trusted because assumptions withstand scrutiny and uncertainty is acknowledged rather than overlooked. Put a polished AI-built model in front of an investment banking VP with five years’ experience, he says, and they will spot assumptions that do not hold up, even when all the numbers tie out.
The stakes are rising with what AI is being asked to do. Systems that once pulled figures from annual reports are now building financial models, analysing transactions and running multi-step workflows. Once that happens, an error stops being a wrong answer on a screen and starts flowing into decisions. In transactions of the size institutional finance handles, even small mistakes can do real damage.
Capturing the banker’s reaction
New York-based Proxion, founded in April 2026, builds training data and evals for AI labs and financial AI companies, drawing on a network of more than 1,000 senior finance professionals across investment banking, private equity and asset management.
The idea is to write down what those professionals check instinctively. A proper test of a deal model would ask whether the leverage is sustainable and the financing reflects current market terms, not only whether the spreadsheet balances.
The goal, as Arshakian frames it, is to reduce errors at the source rather than catch them afterwards: trained on detailed datasets built by senior finance professionals, models can learn how transactions are actually structured, instead of simply producing financial models that look convincing on paper.
“We’re training models to make better decisions and testing whether their work can be trusted,” he said.
Why finance needs its own standard
Arshakian’s perspective was shaped on the sell side. As an associate at Deutsche Bank, he worked on more than 45 transactions, including over $80 billion in investment-grade debt issuance. Much of that work meant evaluating capital structures, financing capacity and downside risk. Later, as a finance expert reviewing AI models, he recognised an opportunity to bring that same level of financial judgement into how AI systems are trained and tested.
Broad AI data platforms play an important role, he acknowledges, but finance cannot be handled as one more vertical inside a general-purpose operation. Financial decisions turn on transaction structure, accounting treatment, legal terms, materiality and risk appetite, and the standards for each have to be set by people who understand how institutional finance operates. Each project, he adds, leaves behind more calibrated reviewers and a growing record of how models fail on financial tasks, knowledge that compounds over time.
Improving the work, not removing the professional
Arshakian is deliberate about how he describes AI’s effect on finance roles. He sees the near-term change as delegation: analysis that once took hours or days can be produced in minutes, freeing professionals for judgement, client relationships and the decisions that carry the most weight. That only holds if the output can be trusted.
For banks, asset managers and investment firms, the question is shifting from whether AI can do financial work to whether anyone can prove it does that work reliably. Arshakian’s answer is that the proof has to come from the people who know what a defensible deal looks like, built into how models are trained and tested from the start.
