BNY put AI governance inside the interface, not around it

The interface is the control plane

BNY did not describe Eliza as a chat window bolted onto frontier models. It described a front end where prompting, agent development, model selection, and sharing all happen inside the same governed environment. That design decision is the whole story.

Eliza embeds governance at the system level. It standardizes permissions, security, and oversight across all models and tools, ensuring every workflow meets the same level of protection.Montana Labs

Deputy General Counsel Watt Wanapha's phrasing matters for anyone building enterprise AI surfaces. Tagging, telemetry, approval flows, and access controls are enforced by the interface itself — the source says this happens 'without burdening end users with manual steps.' The user builds an agent; the front end quietly records who built it, what model it used, and who can see it. Governance is not a gate the user walks through. It is the material the tool is made of.

Sharing scoped to ten colleagues

The sharing model is a concrete front-end constraint worth noting. Eliza initially allowed only private agent builds. It now lets agents created by certain teams and roles be shared with up to ten colleagues. That cap is a deliberate product boundary — enough to create reuse, small enough that a shared agent stays inside a knowable blast radius.

The result BNY reports is a reuse pattern where 'one team's agent often becoming another's foundation,' feeding more than 125 tools in production across major business lines. The interface treats an agent as a shareable artifact with an owner and an access list, which is how a private experiment becomes a departmental asset without a separate provisioning process.

Building for non-developers, measured by habit

The front end's real test is whether people who are not engineers will build with it. BNY's evidence is specific: 20,000 employees actively building agents, 99% of the workforce trained on generative AI, and a Sales hackathon where, per Head of Sales Ed Fandrey, 'there were no IT or tech folks present, but everyone felt like a developer.'

BNY also tied interface adoption to a behavioral program. 'Make AI a Habit Month' ran daily seven-minute trainings on prompting and agent building, and Global Head of Talent Michelle O'Reilly reports it drove 'a 46% increase in the number of agents people were building.' The lesson embedded here is that a low-friction builder surface still needs a deliberate onboarding cadence before non-technical staff treat it as a daily tool rather than a novelty.

From front end to 'digital employees'

BNY extends the same surface into what it calls 'digital employees' — agents with identities, access controls, and dedicated workflows handling tasks from payment instruction validation to code security. The reported early wins are narrow and measurable: a Contract Review Assistant cut legal review from four hours to one, a 75% reduction, across 3,000+ annual vendor agreements.

The implication for teams building AI front ends is that the interface is not a thin wrapper over a model — it is where identity, permissioning, and telemetry get defined for every agent an employee creates. BNY chose to extend its existing legal and compliance frameworks into that tooling rather than write AI-specific governance from scratch. For a firm with $57.8 trillion under custody, the front end is not the last mile of the system; it is where accountability starts. That is the reusable engineering choice, independent of which frontier model sits behind it.