LHV Bank deploys agentic AI for customer emails via Gradient Labs

LHV Bank has gone live with a fully autonomous AI agent for customer email support in the UK, built on Gradient Labs' platform, in a deployment that draws a clear line between the incremental AI adoption most banks have settled for and the agentic model now beginning to emerge in regulated financial services.

The agent handles common retail banking enquiries end-to-end, lost cards, payment explanations, transfer queries, without a human reviewing each response before it reaches the customer. Cases judged sensitive, complex, or outside the agent's approved procedures are escalated to human staff. The bank says the design frees its customer operations team to concentrate on cases where human judgement is genuinely required, rather than approving AI-drafted replies to routine questions in a queue that still looks like the old queue.

Agentic AI versus the co-pilot model

The distinction matters strategically. Most banks deploying large language models in customer operations have taken the co-pilot route: the AI drafts a reply, a human approves it, the reply goes out. The approval step was designed to keep humans in the loop, but as LHV's interim CEO Kris Brewster frames it, the model carries its own failure mode. "When the AI is usually right, people can start approving suggestions without checking them closely," he said. "We wanted to go further and give customers an answer straight away, while keeping our people firmly in control of how the service works and stepping in whenever a case needs them."

To make autonomous resolution viable in a regulated environment, Gradient Labs pairs large language models with deterministic guardrails: hard rules governing which data and tools the agent can access, when actions must be blocked or escalated, and how complaints and interactions with vulnerable customers are handled. Customer data is not used to train or build the models, and is accessed only during live interactions. The approach is designed to satisfy the Financial Conduct Authority's growing expectations around AI governance in retail banking, where operational resilience and consumer-duty obligations create significant liability for uncontrolled model behaviour.

Gradient Labs' Collaborate tooling adds a software-engineering discipline to the agent's operating procedures: LHV's operations, product, risk and compliance teams can draft procedural changes in a controlled branch environment, test them against representative customer cases, and approve or roll back changes with a full audit trail. Dimitri Masin, Gradient Labs' CEO, said the end-state is "trusted agents taking responsibility for defined work, supported by deterministic controls, clear escalation points and human judgement where it matters."

The convergence angle: fintech infrastructure meets compliance automation

The LHV-Gradient Labs deployment sits at a broader intersection. Regulated-sector AI adoption has long lagged consumer and enterprise contexts because the cost of a compliance failure is asymmetric: a hallucinating agent that misadvises a retail banking customer on a vulnerable-customer case carries reputational and regulatory consequences that a chatbot confusing a takeaway order does not. The emergence of deterministic guardrail architectures, blending the generative flexibility of foundation models with rule-based constraint layers, is the engineering response to that asymmetry, and it is beginning to unlock agentic deployment in finance, healthcare and insurance simultaneously.

For investors tracking this space, the capital signals are notable. Gradient Labs was founded in 2023 by alumni of Monzo's AI and data functions, and is backed by Octopus Ventures, CommerzVentures, Redpoint Ventures and LocalGlobe, a VC roster that spans UK fintech generalists and European strategic capital. The company says its agents now serve more than 32 million end users across its client base, a scale metric that places it credibly in the tier of infrastructure providers rather than point-solution vendors. The next competitive pressure point will be channel expansion: LHV's roadmap includes live chat, voice, back-office workflows, and KYC and AML automation, the last of which represents a considerably higher-stakes regulatory frontier than email triage. How Gradient Labs' guardrail architecture performs when applied to financial crime detection, where false negatives carry criminal and regulatory liability, will be the real proof of the platform's ambitions in regulated AI infrastructure.