Neal Lathia, Co-Founder and CTO of Gradient Labs – Interview Series

Neal Lathia, Co-Founder and CTO of Gradient Labs, is a machine learning and data science leader with nearly two decades of experience spanning financial technology, consumer platforms, and academic research. Before co-founding Gradient Labs in 2023, he spent more than five years at Monzo, progressing from Senior Data Scientist to Director of Machine Learning and Staff Machine Learning Engineer, where he built machine learning infrastructure and helped scale the discipline to support applications across operations, financial crime, and product. Earlier, Lathia worked as a Senior Data Scientist at Skyscanner and held research positions at the University of Cambridge and University College London, where his work explored recommender systems, behavioral data, smartphone sensing, and personalized digital services. At Gradient Labs, he now leads the technology behind AI agents designed for complex, regulated financial-services workflows.

Gradient Labs is a London-based AI company building specialist autonomous agents for financial services, with a focus on automating customer operations that traditionally require substantial human involvement. Founded by former Monzo leaders Dimitri Masin, Neal Lathia, and Danai Antoniou, the company develops agents for areas including lending and collections, disputes, Know Your Business (KYB), onboarding, insurance claims, and customer service, operating across voice, chat, and email while incorporating compliance guardrails designed for regulated environments. Gradient Labs says its technology now serves more than 32 million end users across customers including Wise, Zego, Current, Stash, and Rho. In June 2026, the company announced it had expanded its Series A financing to $26 million as it works toward a broader vision of automating complex banking and fintech operations with interconnected specialist AI agents.

You spent more than five years building and leading machine learning at Monzo before co-founding Gradient Labs in 2023. What did you see firsthand at Monzo that convinced you there was an opportunity to build a new company around autonomous AI agents, and why did the arrival of large language models make that opportunity possible?

At Monzo, I spent years building ML systems that needed to work in a regulated environment, where mistakes have very real consequences and everything requires an audit trail. What influenced me most was the need to develop custom-built infrastructure, because nothing pre-made could perform safely under those strict constraints. That experience left me with two convictions. The first was that technology could finally lift banks out from under the operational burden that has held the industry back for decades. The second was that general-purpose, horizontal AI wouldn’t make a dent in it — the nuance of regulated work is too specific. When LLMs became capable enough to reason through multi-step, sometimes ambiguous customer interactions, it started becoming achievable to build agents that could manage entire regulated workflows end-to-end, rather than just assisting a human to do it. That gap, between what regulated companies actually needed and what already existed, is why we started Gradient Labs and why we bet on vertical AI built specifically for finance.

You’ve described what consumers are experiencing as the “jagged edge of experience”: an AI agent can perform an extraordinarily complex task while another system fails at something as simple as distinguishing a phone number from a name. What actually causes such a dramatic gap between AI experiences when the underlying models may be similarly capable?

I’d say that the gap is more about how much engineering goes into the model than about the raw model competence instead. If a system fails at a simple task, such as mistaking a phone number for a name, it means it has not invested in the framework around it, like validation, fallback logic, structured data handling, etc. The more impressive agent has likely been engineered precisely for its task. It’s the same fundamental model family with a completely different amount of stringency around it, and that is exactly what leads to the jagged edge.

As frontier models continue to improve, why do seemingly basic AI failures persist? Are these primarily limitations of the models themselves, or failures in the surrounding system architecture, data, workflows, evaluation, and product design?

It is mainly the system and not the model. Frontier models continue to improve at reasoning. However, companies often latch them to systems that were not developed for probability-based behaviors. There are fragile integrations and incomplete data in the systems, with no real evaluation loop before changes are shipped. So a basic AI failure within production is not really an AI failure at all, but a failure to invest in evaluation, monitoring, and workflow design processes.

Many companies appear to optimize AI customer service around speed, containment, or ticket deflection. What metrics should companies be using instead if they want to measure whether an AI agent is genuinely improving the customer experience?

For this, we would have to consider resolution accuracy. Speed and containment are used to assess whether you were able to get a customer off the phone, not if you were able to resolve their issue. Resolution accuracy, which should be used to measure an AI agent’s capability, will consider if the agent did the right thing, as opposed to just responding quickly. This will work together with other signals like repeat contact rates, the volume of complaints, and how often a human has to intervene afterward. You will know you are optimising for the wrong outcome if deflection goes up, but repeat contacts and complaints also increase.

Gradient Labs focuses on regulated financial services, where an incorrect response can have much greater consequences than a typical customer-service mistake. How do you determine when an AI agent is reliable enough to autonomously handle processes involving areas such as lending, disputes, onboarding, or Know Your Customer checks?

The default route that most teams would rather take is that of co-pilot: someone approves every action because it feels safer. And most would assume it is safer. But in reality, as the AI is getting most things correct, reviewers approve things without really checking, and safety is lost anyway. This does not do anything to minimise the existing workload. It just makes it faster and reduces the value you get. 

This is why the concept of an AI agent being reliable enough is quite process-specific at Gradient Labs. For us, it has to be developed through staged autonomy and not be seen as just one goal to tick off. An agent can be given greater independence on things like KYC or disputes only after it has been benchmarked against a large, living set of real cases, with human review on a sample of its decisions even after it has gone into production. Interestingly, this then confirms that reversibility also really matters here. If something can be undone, it will gain autonomy more quickly than something that cannot.

Guardrails are increasingly presented as the solution to making AI agents safer, but adding more rules can also make systems rigid or prevent them from completing legitimate tasks. How do you balance autonomy with guardrails without reducing an agent to another highly constrained chatbot?

The mistake is treating guardrails as a wall the agent simply bounces off. In our system, they do two things at once. We run guardrails on every single turn of a conversation — some inspect what the customer is saying, to catch things like vulnerability, financial difficulty, or a complaint, and others inspect what the agent is about to say, to keep it compliant. But when one fires, it doesn’t just block: it re-routes the agent into the right procedure, and that decision is transparent in the agent’s reasoning, so operators can see why it acted. The deeper guardrails are also built into how the agent is taught to think about a task, the data it can access, and the tools it can use. That combination is what keeps it safe without making it rigid: the agent understands what the guardrails are and why they exist, so it completes a legitimate task instead of refusing anything that merely resembles an impermissible one.

Here, I think the issue is seeing guardrails as a brick wall for the agent to bounce off of. More specifically, we run two different types of guardrails at every single turn a conversation takes. First, we have the customer guardrails that inspect what the customer is saying and are supposed to pick up on vulnerabilities, financial difficulties, customers making complaints, and others alike. Second, we have agent guardrails that work the other way around and check what the agent will say before sending the message, in order to keep it compliant.

This means that, instead of completely blocking an action when one of these guardrails is fired, it will reroute it to the right path. This way, it maintains transparency regarding the agent’s decision-making processes and provides human operators with the reason for taking said action.

We are also able to do this because not every aspect of an agent’s reasoning process is forced to pass an LLM. Guardrails such as these are deterministic because they won’t only fire at the end of a conversation, but will trigger throughout the agent’s lifetime. This is how we are able to trust the agent with sensitive communication and allow customisation of specific agents for use cases at scale and in a predictable manner.

What matters the most is how the agent is taught to think through a task in the first place, ensuring that it can access relevant data and know the tools it needs to use. This ability for the agent to complete a legitimate task instead of refusing to take on anything that seems as though it’s not permitted is what helps improve safety and avoid rigidity.

Where should humans remain in the loop as AI agents become increasingly autonomous? Are there certain decisions or customer interactions that you believe should continue to require human judgment regardless of how capable the underlying models become?

A human needs to be involved if a decision calls for honest discretion, has significant repercussions for the customer, or has to do with an outcome that the agent has not yet been evaluated against. The human does not necessarily have to do the whole thing, but they should review or approve as needed. This is probably even true as models get more capable, because it often comes down not to a question of capability, but of accountability and of a customer’s right to a human decision-maker where things such as credit outcomes or disputes are involved.

A human needs to be involved if a decision calls for honest discretion, has significant repercussions for the customer, or has to do with an outcome that the agent has not yet been evaluated against. I would even take this a bit further than the typical fallback model to argue that agents are increasingly earning their place in the organisational chart alongside humans. In turn, this is shifting where human efforts are heavily applied towards escalations and judgement calls instead, leaving agents to handle the coordination and routing work.

But for this to be really successful, agents need the same institutional context that a human would have too, so they know who to loop into certain issues and at what time. This is why we built Collaborate. To let operators, engineers, and agents work together as peers with version control, evaluations, and continuous learning at the heart of it. This ensures that humans stay fully in the loop, and the loop just changes shape to account for how well the agent and human coordinate.

Gradient Labs has focused on specialist agents designed for specific financial-services workflows rather than a single general-purpose agent. Do you believe the future of enterprise AI will primarily consist of networks of specialized agents, or will increasingly capable foundation models eventually make that specialization less important?

I think even as base models improve, specialization still matters. People will not be paying just for a smart model, but for evaluation specific to the workflow, guardrails, and data integration built around it. And that work doesn’t go away just because the underlying model gets better. I would see this being a network of specialist agents all built on top of a common and strong base model rather than one general-purpose agent doing everything.

AI agents increasingly need to learn and improve after deployment, but in regulated environments even a small behavioral change can introduce new risks. How can companies continuously improve an agent while ensuring that updates do not create regressions, compliance problems, or unexpected behavior?

The principle we work to is that every update is treated as a new model release, not an incremental configuration change, so every change goes through a full evaluation suite before it ships. That includes regression tests run against precedents where a mistake would have made a real difference, and nothing goes live to every customer at once: we roll out gradually and monitor, so that if drift starts to occur we catch it on a limited slice of traffic rather than at full scale. This is exactly what our recent Collaborate release is built around. Collaborate lets operators, engineers, and the AI agent work together as peers on the same living definition of how the agent should think, with version control, evaluations, and continual learning built directly into the workflow. So when someone improves how the agent handles a case, that change is versioned, tested against past conversations, and rolled out under the same controls as any other release. It means an agent can keep getting better after deployment without the improvement itself becoming the thing that introduces a regression or a compliance problem.

Looking ahead, as access to powerful foundation models becomes increasingly commoditized, where will the real competitive advantage in AI applications come from? Will the winners be those with the best models, or those that are best at engineering reliable systems and delivering consistently good experiences around them?

Now that more and more foundation models are beginning to converge, the upper hand will move almost entirely in favor of whoever is best at engineering more reliable systems around these models, including evaluation, guardrails, data, and workflow design. The model is becoming a commodity input, but it is consistency and dependability in production that are the actual product.

Even beyond that, I see the advantage coming more from how deeply an agent has been embedded in a company’s operations. Agents that have the ability to embed in many different parts of a company’s operations will win over tools that can only handle specific problems or only reach the frontline. When an agent can work across frontline and back office systems, it can bring more context to interactions and handle more complex problems end-to-end, just like a human would. An agent being able to fit with how a company actually works is where I expect the lasting win to come from.

Thank you for the great interview, readers who wish to learn more should visit Gradient Labs.