Macquarie Bank has stood up a centralised knowledge base of tightly governed enterprise data and code assets to support the development of agentic AI applications.

Built on Google Cloud, the Knowledge Platform aggregates structured data from the bank’s repositories alongside unstructured content such as PDFs, SharePoint files, Confluence pages and operational procedures.

Data is funnelled through a “highly curated pipeline” into the platform, which then feeds into Google’s Agentspace, where Macquarie Bank plans to develop an ecosystem of agentic AI applications.

Speaking at Google Cloud Summit last month, Macquarie Bank chief data officer and executive director Ashwin Sinha called the platform “the most important layer” of the envisioned ecosystem.

“You can achieve a level of accuracy or consistency based on foundational models, so just large language models by default,” he said.

“But to really make it useful in a production-ready manner, you need the underlying knowledge repository [which is ] well curated, well maintained and [has] a very clear sense of ownership.

“The curation of it becomes very important in terms of whether it is the latest version or whether there are multiple versions [of a data asset],” he added.

“How are you going to make sure that information is continuously maintained and current?”

Within Agentspace, the bank has enabled NotebookLM, Google’s AI-powered assistant that helps summarise notes, along with enterprise-wide search capabilities across Macquarie’s data repositories.

According to Sinha, this search engine is already connected to several data sources via connectors, with plans to expand those integrations further.

These initiatives support Macquarie Bank’s ambition to develop a comprehensive ecosystem of AI agents, with several currently in pilot stages.

According to a conceptual framework outlined by Sinha, these agents would start from personal assistants, which function as “simple no-code automation of individual tasks”.

At the next level are enterprise agents, purpose-built specifically for Macquarie’s Banking and Financial Services (BFS) division.

Meanwhile, tools such as GitHub Copilot may be tapped to develop integrated development environment (IDE) agents that will “enhance software developer productivity”.

“We have found these to be incredibly useful for our engineering team in terms of driving their productivity and their pace of delivery,” Sinha said.

To complement this, Sinha envisions adopting a set of “core” or reusable agents, where “aspects of software engineering or software development could be automated or semi-automated”.

A “less mature” idea in the works is to create a marketplace for third-party agents, which, according to Sinha, “will drive a lot of productivity or tasks in terms of things which have to be done here”.

A year of experiments

These initiatives evolved through a concerted effort across the entire bank to explore and trial AI technologies over the past 12 months, from the top down.

“We created a culture of experimentation and very rapidly learning about many of these technologies,” Sinha said.

“We also went on an extensive exercise of training all of our leadership, as well as all of our staff. We trained over 2,500 people in prompt engineering.

“We trained the top 150 leaders across the bank on this topic, and that created a lot of good ideas.”

This training culminated in a session for Macquarie Bank’s wider leadership team, which sparked ideas for “130-plus agents, which can change how we are running our business”, Sinha said.

This employee and stakeholder engagement also reflected Macquarie Bank’s strategic view of AI as a “whole-of-business change”.

“The way we see generative AI is that it’s more of a general-purpose technology, which means a lot of people who are outside technology areas are still able to use it, make really productive use of it and come up with very good ideas that can drive productivity and better customer outcomes,” Sinha said.