Dimitri Masin is CEO and co-founder of Gradient Labs, AI launches building autonomous customer support agents designed specifically for regulated industries such as financial services. Prior to founding Slope Labs in 2023, Masin held senior leadership positions at Monzo Bank, including vice president of data science, financial crime and fraud, and previously worked at Google. Under his leadership, the progressive lab quickly gained traction, with annual recurring revenue reaching £1 million within five months of release. Masin’s focus is on developing AI systems that combine high performance with strict regulatory compliance to provide secure and scalable automation for complex customer operations.
What prompted you to launch the Gradient Lab after such a successful journey in Monzo?
At Monzo, we spend years working on customer support automation, usually with a moderate 10% efficiency improvement. But in early 2023, we witnessed a seismic shift in the release of GPT-4. Suddenly, 70-80% of manuals can be automatically automated completely and automatically through AI, repetitive work.
This technological breakthrough we are currently stimulating us to start the Gradient Lab. In my career, I have seen two revolutionary waves: the mobile revolution (which happened early in my career), and now it is AI. You have to seize the moment when you realize that you are in the middle of this transformation, which will completely change the way the world works. Our team knows – it’s time.
At Monzo, you helped the company experience a lot of super-growth. What is the biggest lesson you have learned from your experience applying for Gradient Labs now?
First, balance autonomy with direction. At Monzo, we initially assumed that people just thrive on autonomy – that’s what motivated them. However, this view seems too simple now. I believe people also value guidance. True autonomy is not about telling people to “do whatever you decide to do”, but providing clear direction while giving them the freedom to solve well-defined problems.
Secondly, top talents need the highest salary. If you intend to hire the top 5% in the function, you must pay accordingly. Otherwise, once you know that your top talents are underpaid, major tech companies will hire them.
Third, don’t reinvent the wheel. At Monzo, we try to create innovative ways to create job structures, compensation systems and career ladders. Key Point: Don’t waste energy on organizational fundamental innovations – Thousands of companies have established best practices. I still see LinkedIn posts about “Get rid of all titles and hierarchies” – I’ve watched this game over and over again, and almost all companies end up regaining their traditional structure.
Gradient Labs focuses on standardized industries that have traditionally had complex needs. How do you deal with AI agents (such as OTTO) that can run effectively in this environment?
We took an unconventional approach, rejecting typical suggestions for quick release and iterating on-site products. Instead, we spent 14 months maintaining high-quality bars from the start before launching Otto. We need to create banks and financial institutions will trust to handle their support completely autonomously.
We are not building co-pilots – we are building end-to-end automation of customer support. With our financial services background, we have an accurate internal benchmark of a “good look” that allows us to evaluate quality without relying on customer feedback. This allows us to be free to be obsessed with quality while iterating rapidly. Without on-site customers, we can make a bigger leap, freely break things and spin quickly – ultimately delivering exceptional products at launch.
Otto is not just answering simple questions, it can also handle complex workflows. Can you guide us through how OTTO manages multi-step or high-risk tasks that a typical AI agent may fail?
We have built OTTO around the concept of SOP (standard operating procedures), which is essentially a guide file written in plain English that details how to deal with specific issues, similar to the agent you give to people.
Two key architectural decisions make Otto particularly effective in managing complex workflows:
First, we limit the exposure of tools. A common failure mode for AI agents is choosing errors from too many options. For each process, we only expose a small portion of the relevant tools to Otto. For example, in the replacement card workflow, Otto can only see 1-2 tools instead of all 30 tools registered in the system. This greatly improves accuracy by reducing the space for decision making.
Second, we rebuilt many typical AI assistant infrastructures to achieve broad thinking reasoning. Our architecture is not just about investing in the process on OpenAI or Anthropic Assistant, but also allows for multiple processing steps between input and output. This allows for deeper reasoning and more reliable results.
Gradient Lab mentions achieving “superman quality” in customer support. What does “superman mass” mean to you and how do you measure it internally?
Superman’s quality means providing better customer support than humans can achieve. The following three examples illustrate this:
First, comprehensive knowledge. Artificial intelligence agents can process a large amount of information and have detailed knowledge of the company. In contrast, humans usually learn only a small portion of information, and when they don’t understand something, they have to consult a knowledge base or upgrade to their colleagues. This leads to a frustrating experience between teams, clients pass through clients. By contrast, AI agents have a deep understanding of the company and its processes, providing consistent end-to-end answers – no upgrade required.
Second, non-lazy lookup – AI collects information quickly. When humans try to save time by asking questions before inquiring, AI proactively checks account information, flags, alerts, and error messages before the conversation begins. So when a customer vaguely says “I have a problem with X”, AI can provide a solution immediately instead of asking multiple clarification questions.
Finally, patience and quality consistency. Unlike people who face the pressure of a certain number of replies per hour, our AI maintains consistent high quality, patient and concise communication. It answers patiently without haste.
We measure this primarily by customer satisfaction scores. For all current clients, we achieve an average of 80%-90% CSAT scores – usually higher than their human teams.
You deliberately avoid connecting Gradient Labs with a single LLM provider. Why is this choice important and how does it affect customer performance and reliability?
Over the past two years, we have observed that whenever OpenAI or Anthropic releases something faster, better or more accurate, our biggest performance improvement is our ability to switch to the next best model. Model agility has always been key.
This flexibility allows us to continuously improve quality while managing costs. Some tasks require more powerful models and fewer models. Our architecture enables us to adapt and evolve over time, selecting the best model for each situation.
Ultimately, we will support private open source LLM hosted by customer infrastructure. This will be a direct transition due to our architecture, which is especially important when serving banks that may have specific requirements for model deployment.
Gradient Labs is more than just building chatbots – your goal is also to handle the backend process. What are the biggest technical or operational challenges when automating tasks like AI?
There are two different categories of processes, each with its own challenges:
For simpler processes, this technology already exists. The main challenge is integration – connecting to many custom backend systems and tools used by financial institutions, as most customer operations involve many internal systems.
There are still significant technical challenges for complex processes. These processes usually require humans to be employed and trained for 6-12 months to develop expertise, such as fraud investigations or money laundering assessments. The challenge here is knowledge transfer – how do we provide the same domain expertise for AI agents? This is a tough problem and everyone is still trying to solve it.
How does Gradient Laboratories balance the demand for AI speed and efficiency with strict compliance requirements in regulated industries?
Of course, it’s a balance, but at the conversation level, our agents only need more time to think. It evaluates multiple factors: Do I know what the client is asking? Am I giving the correct answer? Does the customer show a sign of vulnerability? Do customers want to file a complaint?
This intentional approach increases the latency – our median response time may be 15-20 seconds. But for financial institutions, it is a fair trade. The 15-second response is still much faster than the human response, and quality assurance is more important to the regulated companies we work with.
Have you foreseeed a future where AI agents not only support but also trusted in higher bet decision-making tasks within financial institutions?
Financial institutions have already used more traditional AI technologies to make high-risk decisions before the current wave of generative AI. The real opportunity I see now is in orchestration – not making decisions, but coordinating the process.
For example, a customer uploads documents, and the AI agent routes them to the verification system, obtains validity confirmation, and then triggers appropriate operations and customer communications. This orchestration feature is where AI agents perform well.
I don’t see much change in the short term for the highest decision itself. These models require interpretation, bias prevention and approval through the Model Risk Committee. In these cases, large language models will face significant compliance challenges.
How do you think AI will reshape the customer experience of banks, fintech companies and other regulated departments in the next 3-5 years?
I see five main trends that reshape the customer experience:
First, true omnichannel interactions. Imagine launching a chat in your banking app and then seamlessly switch to voice with the same AI agent. Sound, call and chat will blend into a single continuous experience.
Second, the adaptive UI minimizes navigation within the application. Instead of looking for specific features through the menu, customers simply express their needs: “Please increase my limit” – the action is immediately carried out through conversations.
Third, better unit economics. Support and operation are a large cost center. Reducing these costs can enable banks to serve previously unprofitable customers or pass on savings to users, especially in underbanked segments.
Fourth, large-scale excellent support. Currently, startups with a few clients can provide personalized support, but as the company grows, the quality often decreases. AI makes strong support scalable, not just possible.
Finally, customer support will shift from frustrating necessity to truly useful services. It will no longer be seen as a labor-intensive infrastructure cost, but rather a valuable, effective customer touchpoint that enhances the overall experience.
Thanks for your excellent interview, and readers who hope to learn more should visit Gradient Labs.
