Building AI Agents That Can Be Trusted in the World’s Most Regulated Workflows

You spent more than five years building and leading machine learning at Monzo before co-founding Gradient Labs. What did you see firsthand that convinced you there was an opportunity to build a new company around autonomous AI agents?

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 experience 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 causes that gap?

I’d say that the gap is more about how much engineering goes into the model than about raw model competence. 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: validation, fallback logic, structured data handling. 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 limitations of the models themselves, or failures in the surrounding system architecture?

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 in production is not really an AI failure at all — but a failure to invest in evaluation, monitoring, and workflow design processes.

Many companies optimize AI customer service around speed, containment, or ticket deflection. What metrics should they actually be using?

For this, you have to consider resolution accuracy. Speed and containment measure whether you got a customer off the phone, not whether you resolved their issue. Resolution accuracy asks: did the agent actually do the right thing? This works together with other signals like repeat contact rates, volume of complaints, and how often a human has to intervene afterward. You 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. How do you determine when an AI agent is reliable enough to autonomously handle processes like lending, disputes, or KYC checks?

The default route most teams would rather take is co-pilot: someone approves every action because it feels safer. But in reality, as the AI gets most things correct, reviewers approve things without really checking — and safety is lost anyway. This doesn’t 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, not 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, 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 make systems rigid. How do you balance autonomy with guardrails without reducing an agent to another 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.

Where should humans remain in the loop as AI agents become increasingly autonomous? Are there certain decisions that should always require human judgment?

A human needs to be involved if a decision calls for honest discretion, has significant repercussions for the customer, or involves 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 like credit outcomes or disputes are involved.

I’d also argue that agents are increasingly earning their place in the organisational chart alongside humans. Human effort is shifting toward escalations and judgment calls, leaving agents to handle coordination and routing work. But for this to really succeed, agents need the same institutional context that a human would have — knowing who to loop in on certain issues, and when.

Gradient Labs has focused on specialist agents for specific workflows rather than a single general-purpose agent. Do you believe the future of enterprise AI will be 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. Companies will not be paying just for a smart model, but for evaluation specific to the workflow, guardrails, and data integration built around it. 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.

In regulated environments, even a small behavioral change can introduce new risks. How can companies continuously improve an agent while ensuring updates don’t create regressions or compliance problems?

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 Collaborate release is built around. It 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. An agent can keep getting better after deployment without the improvement itself becoming the thing that introduces a regression or a compliance problem.

As access to powerful foundation models becomes increasingly commoditized, where will the real competitive advantage in AI applications come from?

Now that more and more foundation models are beginning to converge, the upper hand will move almost entirely in favour 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 from how deeply an agent has been embedded in a company’s operations. Agents that can work across frontline and back-office systems 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.