Individual access
The relevant distinction is no longer simply between a chatbot and an application. An employee can give a model access to tools, reusable instructions and working information, and delegate a meaningful task. That can be much more useful than asking for a draft or a summary.
But the employee may still decide when work starts, assemble its context, transfer the result and establish what happens next. Those activities can remain the constraint on the wider service even when the delegated task performs well.
Process design
A client instruction needs an identity, a purpose, the appropriate authority and a record of progress. The operation must distinguish an approved update from one that has actually reached the destination system. A changed source may invalidate a conclusion or require an approval to be obtained again.
Those requirements cannot be left to whichever interaction an employee happens to construct. They belong in the process the firm operates, with shared state across people and applications. Individual agents and reusable skills can perform work within it, but their presence does not establish the process.
Execution and coordination
Consider a reporting discrepancy. A basic tool flags the difference and hands it to the team. A more capable system can inspect the supporting records, test plausible explanations, identify the affected period and prepare or complete the permitted correction. It should refer the matter when it reaches a genuine limit of information, authority or capability.
The difference is commercially significant. Treating every unusual input as a human exception leaves the system performing the easy work while staff retain the investigation and coordination that constrain capacity.
Operating design at different scales
A single workflow can establish a useful shared record and a clear account of completion. A larger implementation can extend those foundations across a function. Neither requires a separate bespoke application for every model task, nor a complete replacement of the firm's specialist systems.
The design should specify the result and operating requirements without needlessly fixing every intermediate decision. Some paths are best implemented in ordinary code. Others benefit from a model that can select tools and revise an investigation as new information arrives.
Measurement
The relevant test is the effort required to complete the service at the required standard, including the employee's supervision and correction. A firm can have impressive individual AI usage while leaving its marginal servicing cost largely unchanged.
Halyard's implementation proposition is to address that remaining work directly. The objective is a process that performs more of the service, not simply a better-equipped employee at every existing handoff.