AI agents change who keeps routine work moving between decisions. An AI agent is software that uses a model and connected tools to choose and carry out steps toward a goal. With the right access, it can find the record, ask for missing information, follow up and bring options back to the person who decides. Staff keep the decisions that need their judgment.
Imagine a supplier emails your practice: the supplies due Thursday will arrive Monday instead. Your team needs them Friday.
Someone has to find the original order, check which location needs it, ask what the supplier can do and bring the options to whoever can approve a change. Someone also has to follow up until the problem is settled.
AI could draft a reply. That saves one step. The larger opportunity is software that can keep a task moving after its first answer: check what happened, seek missing information and return when a person needs to decide.
People use “agent” to describe different products. In the sense used here, the model interprets the request, and tools provide access to information and actions. The agent uses what comes back to choose its next step. What it can do depends on its connections and permissions. Google Cloud’s explanation covers these building blocks.
A chat window might be how you talk to an agent. The useful distinction is how much of the task it can carry forward. A fixed reminder works when the next step is always the same. An agent becomes useful when the next step depends on what comes back: a supplier answers only half the question or two records disagree. Anthropic’s guide to building agents distinguishes that flexibility from workflows with predefined steps. It recommends using the simplest system that handles the task, because more flexibility can add cost and delay.
Return to the delayed order. Imagine an agent with access to the relevant purchasing records and supplier correspondence. You ask it to prepare options for getting the approved supplies to the right location by Friday.
It finds the order and checks the supplier’s message against the date your team needs the supplies. It asks about availability through a supplier contact your organization has approved. When a reply arrives, it compares the offered quantity, pack size, delivery estimate and total charges with the original purchase. If freight is missing, it asks before presenting a complete price.
Your team receives the options together, with unanswered questions marked. The person approving the purchase can assess the choice without reconstructing the email thread. Designated staff decide whether any proposed substitute is suitable.
After an approved change, the agent could carry out the permitted follow-up and track the supplier’s response. It would need to distinguish a cancellation request from a confirmed cancellation. Site staff confirm what actually arrived.
The potential benefit is fewer routine steps for someone to remember and coordinate. Staff can focus on choices that need their judgment. That only helps if the agent’s preparation is accurate and easy to check.
The same question arises outside purchasing. AHRQ describes referral staff navigating different processes and missing information between primary and specialty care. Imagine a referral arrives without a prior test report the specialist needs. An agent with approved access could flag the gap, prepare a request for a coordinator to send and track whether the report arrives. The care team still decides what the information means and what happens next for the patient.
In both examples, work crosses records, inboxes and people. Someone must hold the context between events. Agents might help with that continuity, provided they have the right access and can show what they did. Purchasing is Manoa’s focus. Our agents and purchasing specialists do this kind of work. The referral shows how the idea could apply elsewhere in healthcare.
You don’t need to know how to build a model to spot an opportunity. You know which missing detail stalls a task, who can approve a change and what “done” means. Pick one recurring task and write down where it starts, what information is often missing, which steps staff repeat and which decisions only they can make.
Bring that map to a vendor. Our guide to why an AI demo can work while the real workflow still fails explains what to test before you rely on one.
Our purchasing examples show the decisions and follow-up inside a familiar workflow. Start with work your team knows well. You’ll be better able to see where an agent helps, where it stops and whether the handoff back to a person is clear.
