Big AI moves at Lendi Group, brokers and customers benefit

Lendi Group has shifted more than 70% of its development work to AI products and is rebuilding core mortgage technology around agentic systems, moving beyond pilots to production tools designed for its 1,300 brokers.

AI rebuild

"If I look at our development, over 70% of our development now is on AI products, compared to a lot of other organisations that will have proof-of-concept teams and that sort of thing," said Travis Tyler, Chief Product Officer, Lendi Group.

The shift forms part of Aurora, a programme designed to reshape both customer-facing products and the underlying technology. Lendi began building its first AI products in April 2023 before moving into agentic systems capable of handling end-to-end workflows rather than adding isolated assistants to existing processes.

That work extends below the application layer. Lendi is modernising data models, business rules, APIs and workflow services so agents can access structured and unstructured information without depending on logic embedded in legacy systems.

Its existing environment includes Salesforce for workflow management, while the new architecture moves business rules into a domain-service layer organised around customers rather than Salesforce opportunities.

Lendi is also consolidating about 130 APIs into roughly 20 domain services. Tyler said one partner integration built using the new approach was running in three days, compared with about three months under the previous method.

The architecture includes Amazon Bedrock for model access, the open-source Agno framework for agent orchestration and runtime, MongoDB and Kafka, alongside a wider data-platform overhaul involving Snowflake and Databricks.

The initial greenfield build involved 20 people, including one Product Manager, a Business Analyst and a Designer, with most of the remaining team made up of engineers. Lendi also used external partners early in the programme to provide AI engineering skills it did not yet have internally.

Aurora has since been divided into separate streams. Aurora Central is responsible for the future-state platform, while Aurora to Core brings components from that environment back into the existing business. Another programme focuses on making AI tools available to staff, including enterprise products and Lendi's own platform, which incorporates the company's authorisation and security controls.

Guardrails first

"Where there is anything that is regulatory or needs to be 100%, it's deterministic," added Tyler.

Lendi designed the system so AI can present options before handing the process to a broker. Especially where regulated obligations apply.

The agent platform can access about 350 tools covering functions including serviceability, borrowing power and product selection. Lendi has about 4,500 mortgage products available for selection, but decisions requiring certainty are routed through deterministic processes rather than left solely to a language model.

Compliance was also embedded in the development process. Lendi placed its Head of Compliance in the team building the end-to-end system so regulatory requirements could be addressed from the design stage.

The same approach underpins Lendi's policy agent. The system consists of a manager and 28 agents aligned with lenders on the company's panel and draws on about 10,000 pages of policy material. Brokers can ask about circumstances such as overseas or casual income and identify which lenders' policies may apply.

Lendi also added links to source policy documents after testing showed some brokers wanted to verify generated answers against the underlying material. The design keeps AI output as a tool for brokers rather than treating it as a final regulated decision.

Broker rollout

"Our brokers will write more deals. We're 6% of the market, so we write 6% of Australia's loans every single week; we're going after the 94%. In terms of number of people, number of brokers, it will be the same. But our goal is to significantly increase our deals per head," added Tyler.

Lendi expects AI to increase broker capacity rather than remove brokers from the mortgage process. Tyler said brokers account for about 80% of the Australian mortgage market and that customer behaviour will play a major role in determining how quickly more of the process becomes digital or automated.

The policy agent has been developed with direct broker involvement. Twenty brokers helped build and test the agents, while 50 were using the system when Tyler described the rollout. Lendi planned to extend access to all 1,300 brokers within the following weeks and use the wider deployment to assess how the technology is used in day-to-day work.

Early testing has shown two broad behaviours: some brokers use the agent to confirm what they already believe, while others want to inspect the full lender policy. Lendi responded by allowing users to open the supporting policy page directly from an answer.

The wider product strategy extends beyond mortgage origination. Lendi has expanded into buyer's agency and conveyancing and is considering a broader product set organised around the customer rather than treating each loan application as a standalone opportunity.

That shift is also reflected in its new data model, which treats the customer as a core entity.

Its Guardian product follows a similar approach. While the interface can resemble a chatbot, it connects to deterministic services and proprietary data, including a database covering about 11 million properties and attributes such as sale status and valuations. The aim is to bring those services into the agent workflow rather than allow a model to generate answers independently.

Cost control

"We use 20 or 30 models now. We pick the right model for the right task," added Tyler.

Lendi uses Amazon Bedrock as its main model-access layer and can route to models outside Bedrock when required. It has also added open-source models and avoids relying on a single model, selecting them according to task requirements, performance and cost.

Model choice has become a financial issue as well as a technical one. AI usage has introduced a variable cost line into a technology budget where much of the existing spend is contract-based, requiring closer monitoring of token consumption and tighter platform controls.

The need became clear during an early enterprise rollout, when Lendi used about 60% of its available credits in six weeks. It later renegotiated the contract and introduced tighter usage controls.

Tyler receives daily alerts on spending thresholds for the core platform, while engineering and staff usage is tracked to identify both excessive costs and employees who are not using the available tools.

Lendi is also moving its measurement beyond activity alone. Five group-level AI metrics are reported to the board each month: revenue associated with Guardian products, progress towards agentic software development, AI interactions per team member, interactions by user group and the number of autonomous workloads.

AI initiatives are also assessed across four areas: business outcomes, quality, governance and productivity.

Aurora was originally approved through a dedicated business case but has since been incorporated into business-as-usual planning. AI initiatives now have to align with the group's objectives and key results, while expected AI-driven revenue uplift has been incorporated into the budget.