Dimitri Masin is the CEO and Co-Founding father of Gradient Labs, an AI startup constructing autonomous customer support agents specifically designed for regulated industries comparable to financial services. Prior to founding Gradient Labs in 2023, Masin held senior leadership roles at Monzo Bank, including Vice President of Data Science, Financial Crime, and Fraud, and previously worked at Google. Under his leadership, Gradient Labs has quickly gained traction, reaching £1 million in annual recurring revenue inside five months of launch. Masin’s focus is on developing AI systems that mix high performance with strict regulatory compliance, enabling secure and scalable automation for complex customer operations.
What inspired you to launch Gradient Labs after such a successful journey at Monzo?
At Monzo, we had spent years working on customer support automation, typically targeting modest 10% efficiency gains. But in early 2023, we witnessed a seismic technological shift with the discharge of GPT-4. Suddenly, it became possible to automate 70-80% of manual, repetitive work completely autonomously through AI.
This technological breakthrough we’re currently living through inspired us to start out Gradient Labs. In my profession, I’ve seen two such revolutionary waves: the mobile revolution (which happened early in my profession), and now AI. Once you recognize that you simply’re in the midst of such a metamorphosis that can completely change how the world works, you may have to seize the moment. Our team knew – that is the time.
At Monzo, you helped lead the corporate through massive hypergrowth. What were among the biggest lessons from that have that you simply’re now applying at Gradient Labs?
First, balance autonomy with direction. At Monzo, we initially assumed people simply thrive on autonomy – that it’s what motivates them most. Nevertheless, that view now seems overly simplistic. I imagine people also value guidance. True autonomy is not telling people “do whatever you select to do,” but quite providing clear direction while giving them freedom to resolve well-defined problems their way.
Second, top talent requires top compensation. In the event you aim to rent the highest 5% in your function, you will need to pay accordingly. Otherwise, major tech firms will hire them away once it becomes known you may have top talent that is being underpaid.
Third, don’t reinvent the wheel. At Monzo, we tried creating revolutionary approaches to work structures, compensation systems, and profession ladders. The important thing takeaway: don’t waste energy innovating on organizational fundamentals – 1000’s of firms have already established best practices. I still see LinkedIn posts about “eliminating all titles and hierarchy” – I’ve watched this play out repeatedly, and nearly all firms eventually revert to traditional structures.
Gradient Labs is targeted on regulated industries, which traditionally have complex needs. How did you approach constructing an AI agent (like Otto) that may operate effectively on this environment?
We took an unconventional approach, rejecting the everyday advice to release quickly and iterate on a live product. As an alternative, we spent 14 months before releasing Otto, maintaining a really high-quality bar from the beginning. We wanted to create something banks and financial institutions would trust to handle their support completely autonomously.
We weren’t constructing co-pilots – we were constructing end-to-end automation of customer support. With our background in financial services, we had a precise internal benchmark for “what beauty like,” allowing us to evaluate quality without counting on customer feedback. This gave us the liberty to obsess over quality while iterating quickly. Without live customers, we could make larger leaps, break things freely, and pivot quickly – ultimately delivering a superior product at launch.
Otto goes beyond answering easy questions and handles complex workflows. Are you able to walk us through how Otto manages multi-step or high-risk tasks that typical AI agents might fail at?
We have built Otto across the concept of SOPs (Standard Operating Procedures) – essentially guidance documents written in plain English that detail methods to handle specific issues, much like what you’d give a human agent.
Two key architectural decisions make Otto particularly effective at managing complex workflows:
First, we limit tool exposure. A typical failure mode for AI agents is selecting incorrectly from too many options. For every procedure, we expose only a small subset of relevant tools to Otto. For instance, in a card alternative workflow, Otto might only see 1-2 tools as a substitute of all 30 registered within the system. This dramatically improves accuracy by reducing the choice space.
Second, we have rebuilt much of the everyday AI assistant infrastructure to enable extensive chain-of-thought reasoning. Relatively than simply throwing procedures at an OpenAI or Anthropic assistant, our architecture allows for multiple processing steps between inputs and outputs. This allows deeper reasoning and more reliable outcomes.
Gradient Labs mentions achieving “superhuman quality” in customer support. What does “superhuman quality” mean to you, and the way do you measure it internally?
Superhuman quality means delivering customer support measurably higher than what humans can achieve. The next three examples illustrate this:
First, comprehensive knowledge. AI agents can process vast amounts of data and have detailed knowledge of an organization. In contrast, humans typically only learn a small subset of data, and once they don’t know something, they have to seek the advice of knowledge bases or escalate to colleagues. This results in a frustrating experience where customers are passed between teams. An AI agent, in contrast, has a deep understanding of the corporate and its processes, delivering consistent, end-to-end answers – no escalation needed.
Second, non-lazy lookups – AI is quick to collect information. While humans try to save lots of time by asking customers questions before investigating, AI proactively examines account information, flags, alerts, and error messages before the conversation begins. So, when a customer vaguely says “I even have a difficulty with X,” the AI can immediately offer an answer as a substitute of asking multiple clarifying questions.
Finally, patience and quality consistency. Unlike humans who face pressure to handle a certain variety of replies per hour, our AI maintains consistently prime quality, patience, and concise communication. It answers patiently so long as needed without rushing.
We measure this primarily through customer satisfaction scores. For all current customers, we achieve CSAT scores averaging 80%-90% – typically higher than their human teams.
You’ve got deliberately avoided tying Gradient Labs to a single LLM provider. Why was this alternative necessary, and the way does it impact performance and reliability to your clients?
Over the past two years, we have observed that our biggest performance improvements got here from our ability to modify to the following best model each time OpenAI or Anthropic released something faster, higher, or more accurate. Model agility has been key.
This flexibility allows us to constantly improve quality while managing costs. Some tasks require more powerful models, others less. Our architecture enables us to adapt and evolve over time, choosing the optimal model for every situation.
Eventually, we’ll support private open-source LLMs hosted on customers’ infrastructure. Due to our architecture, this can be an easy transition, which is particularly necessary when serving banks that will have specific requirements about model deployment.
Gradient Labs is not only constructing a chatbot — you are aiming to handle back-office processes too. What are the most important technical or operational challenges in automating these sorts of tasks with AI?
There are two distinct categories of processes, each with its own challenges:
For less complicated processes, the technology largely exists already. The fundamental challenge is integration – connecting to the numerous bespoke backend systems and tools that financial institutions use, as most customer operations involve quite a few internal systems.
For complex processes, significant technical challenges remain. These processes typically require humans to be hired and trained for 6-12 months to develop expertise, comparable to fraud investigations or money laundering assessments. The challenge here is knowledge transfer — how will we give AI agents the identical domain expertise? That’s a tough problem everyone on this space continues to be trying to resolve.
How does Gradient Labs balance the necessity for AI speed and efficiency with the rigorous compliance requirements of regulated industries?
It’s actually a balance, but on the conversation level, our agent simply takes more time to think. It evaluates multiple aspects: Am I understanding what the client is asking? Am I giving the proper answer? Is the client showing vulnerability signs? Does the client wish to file a grievance?
This deliberate approach increases latency – our median response time is perhaps 15-20 seconds. But for financial institutions, that’s a good trade. A 15-second response continues to be much faster than a human reply, while the standard guarantees are vastly more necessary to the regulated firms we work with.
Do you foresee a future where AI agents are trusted not just for support but in addition for higher-stakes decision-making tasks inside financial institutions?
Financial institutions were already using more traditional AI techniques for high-stakes decisions before the present wave of generative AI. Where I see the true opportunity now’s in orchestration – not making the choice, but coordinating your entire process.
For instance, a customer uploads documents, an AI agent routes them to a validation system, receives confirmation of validity, after which triggers appropriate actions and customer communications. This orchestration function is where AI agents excel.
For the highest-stakes decisions themselves, I do not see much changing within the near term. These models require explainability, bias prevention, and approval through model risk committees. Large language models would face significant compliance challenges in these contexts.
In your view, how will AI reshape the client experience for banks, fintech firms, and other regulated sectors over the following 3–5 years?
I see five major trends reshaping customer experience:
First, true omni-channel interaction. Imagine starting a chat in your banking app, then seamlessly switching to voice with the identical AI agent. Voice, calls, and chat will mix right into a single continuous experience.
Second, adaptive UIs that minimize navigation throughout the app. Relatively than hunting through menus for specific functions, customers will simply voice their needs: “Please increase my limits” – and the motion happens immediately through conversation.
Third, higher unit economics. Support and ops are massive cost centers. Reducing these costs could let banks serve previously unprofitable customers or pass savings to users — especially in underbanked segments.
Fourth, exceptional support at scale. Currently, startups with few customers can provide personalized support, but quality typically degrades as firms grow. AI makes great support scalable, not only possible.
Finally, customer support will transform from a frustrating necessity to a genuinely helpful service. It’ll not be viewed as a labor-intensive infrastructure cost, but as a priceless, efficient customer touchpoint that enhances the general experience.
