Egor Kraev sees LLMs as components in existing data pipelines, especially for converting messy language into structured fields that traditional machine learning can use.
2
Egor Kraev's MotleyCrew framework lets teams combine agents from different frameworks and enforce validation through tools instead of trusting an agent's prompt instructions.
3
Egor Kraev argues that autonomous organizations need leaders to explain the team's purpose and remove obstacles, while giving people room to choose and pursue useful work.
Summary
Egor Kraev discusses how Wise uses machine learning across financial crime, treasury, customer support, and marketing. He sees the most reliable LLM applications as small pipeline steps that turn unstructured text into structured information, such as classifying complaints or extracting contract rates. He explains how causal inference can estimate customer-level marketing effects, improve targeting, and test hypothetical campaign rules from a prior randomized experiment. The conversation then moves to agents. Kraev describes MotleyCrew, an open-source framework for mixing agent frameworks and enforcing tool-based validation. He is skeptical of agent hype: the difficult work is data quality, permissions, taxonomies, and production integration. He also describes a possible analyst tool that combines quantitative segments with relevant company context through dialogue. The final section covers autonomy at Wise. Kraev says leaders should provide a clear story about the team's purpose and clear obstacles, rather than direct every task. He connects this structure to future AI-assisted organizations.
LLMs are most useful when they turn messy language into structured data
Egor Kraev says one of the strongest uses for large language models is bridging unstructured and structured data. Instead of first converting text into vectors or matrices and then struggling to interpret the result, an LLM can work directly with raw text. In production, he often sees the model classify a customer complaint, extract a rate from a contract, or produce a score that becomes one input to a conventional model. He estimates that at least three quarters of the production LLM applications he has seen follow this pattern. The model is one additional component in a larger pipeline, rather than a system that does everything on its own.
Technology adoption inside an organization can take two years
Kraev says the main obstacle to adopting AI is usually not explaining the technology. Organizational structure and the invisible walls in people's heads matter more. When a team already has a workflow, a new technology can add clear value and still take a long time to become normal. He says the period from introducing a technology to people treating it as an ordinary part of their work is a minimum of two years. A prototype can exist in two weeks, but people need time to understand how it works, build new habits, and stop seeing it as risky or unusual. He advises setting expectations around that longer period.
Wise combines traditional machine learning with newer language-model applications
At Wise, Kraev describes classic machine learning for fraud detection, anti-money laundering, and financial crime. These systems use wide tables of data, XGBoost, and hyperparameter tuning. Treasury uses machine learning to estimate currency flows, manage cash placement, and handle the risks of holding currencies. Customer support is a newer area because much of its information is text-based. Kraev says LLMs now make interactive support possible. A model can assess whether a customer message contains enough information, then suggest what the system should ask next. This turns the model into a part of a controlled support workflow rather than the sole decision-maker.
Causal inference can estimate marketing impact at the customer level
Kraev explains that causal inference differs from ordinary prediction because the impact of sending email A instead of email B cannot be directly observed for the same customer. Wise uses causal inference models to estimate those effects from an A/B test population. The team developed a way to score models out of sample, which allows model selection and hyperparameter tuning even when individual treatment effects are not directly observable. The resulting estimates can support targeting and segmentation. Instead of slicing an A/B test into noisy subgroups, analysts can examine which customers are likely to respond to each treatment. Kraev says the same data can also evaluate hypothetical targeting rules without running a new test for every rule.
Wise Pizza started as a tool for growth analysis. Given customer dimensions such as region, device, currency pair, and product, it searches for simple segments that explain why a metric changed between periods. Kraev says the tool can identify which groups contributed to a change in average revenue or growth. Analysts then investigate those groups instead of searching blindly through pivot tables. The team can fit on one part of the sample and check significance on another part. Kraev is also working on an agentic extension that would bring in relevant information from internal wikis, chats, and Slack channels. He stresses that this future tool would support analysts in telling a data-supported story, rather than fully automating the interpretation.
Agent systems need permissions, taxonomies, and data quality around them
Kraev says the hard parts of building agents are the surrounding systems. Data quality, production engineering lead times, and a clear taxonomy of supported questions matter more than the agent itself. A customer-support agent still needs an underlying list of question types that the business can handle. For internal data, retrieved snippets should carry metadata showing their source and associated permissions. The system can filter out a snippet when the user is not allowed to read the source, such as a private Slack channel or restricted wiki page. Kraev says current internal tools keep humans in the loop and often produce drafts for review. He expects agents to become an ordinary software pattern, similar to databases, rather than a special technology.
MotleyCrew lets agents from different frameworks work together
Kraev created MotleyCrew after finding that existing agent frameworks did not support the mix-and-match model he wanted. Its premise is that teams should be able to combine agents from AutoGen, LlamaIndex, LangChain, CrewAI, and other systems, using the strengths of each without entering a walled garden. The framework wraps common agent types and supports interfaces that allow them to work with LangGraph. Its forced-validation pattern makes an agent return results through a validation tool. If generated Python code is invalid, the tool sends the error back and the agent tries again. If the agent attempts to answer directly, it is told to use the tool. Anything returned through the tool has passed the required validation.
Autonomous organizations depend on stories, trust, and mixed teams
Kraev describes autonomy at Wise as an organizational design choice, not only a matter of personal intent. A leader explains why the team exists, how it fits into the wider organization, and what matters most. The leader also removes obstacles the team cannot resolve itself. People gain more freedom as they build a record of work that others can understand and value. Seniority grows through the ability to convince others to join an idea, first around an individual project and later across the team. Kraev compares autonomy-centered organizations to gardens, where different people and specialties grow together, and says functional verticals can undermine autonomy when they override product-focused teams.
"The vast majority of use cases for LLMs that actually work in production are not just in comes the magical LLM, in comes the magic agent and does everything."Egor Kraev09:33
Who should watch
You are deciding where LLMs belong in an existing machine learning or data pipeline and want examples from fintech operations.
You are building agent systems and need practical guidance on validation, data permissions, source metadata, and production constraints.
You are designing teams around autonomy and want a concrete account of what managers still do when they stop assigning every task.
