Applied vertical AI works best when an agent handles a narrow task for one industry rather than trying to do everything.
2
The hard part is judging outputs in fields such as finance and pharma, where engineers lack the domain training to spot weak reasoning.
3
Hiring domain experts creates the data, prompts, evaluations, and feedback loop that can improve an application and justify its cost.
Summary
Ayush Bhardwaj compares his work at a hedge fund with his current role at Allos AI, a pharma technology company. The industries have different timelines and standards, but the applied AI work follows much the same process. Teams need to define a narrow task, find proprietary data, encode an expert's process in prompts, and observe the agent. Building the agent is relatively easy. Judging whether its output is useful is much harder when the engineers are not traders, scientists, or chemists. Bhardwaj tried using an LLM as a judge and found that it produced plausible jargon without understanding concepts such as alpha. His answer is to hire the user, then build a learning loop around that domain expert. Error analysis is the cheapest place to start, with preference training and reinforcement learning later if needed. He argues that model infrastructure is a commodity. Domain expertise and private data are what make a vertical AI product defensible.
Applied vertical AI simulates a specific job in one industry
Bhardwaj defines applied vertical AI as applied AI built for one specific industry. Its goal is to simulate a person's job in that field. Google Translate is a general-purpose example that can support many industries, while Allos AI focuses specifically on building drugs with AI. Legal technology and real estate technology are other examples. The common feature is a focused industry context, rather than a general assistant that can be used anywhere.
The industry changes, but the applied AI job stays similar
Bhardwaj expected a major change when he moved from a hedge fund to pharma technology. Hedge funds value speed and can tolerate being wrong on individual bets if the overall result is positive. Pharma can take much longer and demands that work be right. Despite those differences, the core applied AI work stayed the same. Conversations with people in legal AI and real estate technology led him to see a shared process that could be broken into a series of steps.
A production agent matters only when it works and pays for itself
Bhardwaj rejects the question of whether teams are putting agents into production. He says that even very young people can deploy agents. The useful questions are whether the systems work, save or make money, and justify their return on investment from end to end. In his experience at both the hedge fund and the pharma company, the agents either saved money or made more money. Shipping an agent is therefore an incomplete milestone.
A narrow task gives an agent something it can actually do
The first step is to formulate a narrow problem. A finance agent should not be asked to find the top three market opportunities in general. Instead, the team could specify US equities, choose an industry, and rank stocks using defined factors such as capital expenditure or AI investment. Pharma work can be broken into smaller steps and handled through pointed questions. Bhardwaj sees no reason to make one agent do everything when teams can build separate agents for separate jobs.
Private, curated data differentiates a vertical application
Public sources such as sell-side reports, preprint servers, PubMed, and research papers are available to everyone. Bhardwaj says the advantage over ChatGPT or Claude comes from proprietary data. Organizations often have years of useful material that is still unstructured, and an LLM workflow can turn it into structured data. In finance, trade theses explain what worked and why. In pharma, failed experiments are especially valuable because successful experiment data is easier to obtain.
Building an agent is easier than knowing whether its answer is good
Bhardwaj says the first four steps, including prompts, data, and observability, can fit on one screen. The difficult part begins when the team tries to iterate. He could build an agent but could not tell whether its output was sound because he was neither a trader nor a biology or chemistry expert. Engineers can judge generated code because years of practice give them a mental model for quality. They do not automatically have that judgment for trade theses or drug candidates.
An LLM judge cannot replace missing domain judgment
Bhardwaj tried to solve the evaluation problem with an LLM judge and calls it a stupid mistake. He says the model can jargon its way through an evaluation without understanding what alpha means or how value is created. Verifiable rewards work well for mathematics and code because answer keys exist, code can compile, and results can be checked. In finance and pharma, errors can compound across multiple steps, while the model has no reliable way to verify its own reasoning.
The missing training data is withheld for economic and legal reasons
The data needed to teach models to reason in finance and pharma is often unavailable by design. Hedge funds must disclose qualifying long positions quarterly, but competitors can reverse engineer those positions and reduce the fund's returns. Pharma companies are required to disclose clinical trials, yet Bhardwaj says 30% do not, and that the FDA reminded more than 2,000 sponsors in 2026 to release unfavorable results. Frontier labs do not have this private data, and hiring traders to annotate it is expensive and restricted by nondisclosure agreements.
Hiring the user turns expert judgment into an improvement loop
Bhardwaj's answer is to hire the person who will use and buy the product. At the hedge fund, the trader was already his boss and collaborator. At the pharma startup, a group of young engineers hired a senior scientist. The scientist changed the tools' trajectory, and the tools began to make sense to large pharma companies because they spoke the industry's language. The expert then curates reliable sources, improves prompts, decomposes problems into a logical sequence, and judges outputs.
"The question to ask is whether they actually work, whether they actually make or save money, whether they justify their ROI."02:57
Who should watch
You are building an AI product for finance, pharma, legal work, or another field where engineers cannot easily judge the output.
Your team has deployed an agent but still cannot explain whether it works or pays for itself.
You need a practical order for using domain experts, proprietary data, observability, and error analysis.
