Seven jobs AI agents do at a benefits agency

How does an AI executive assistant help an agency owner?

An AI executive assistant opens the day with what’s due, what’s overdue, and what the other agents need from you. I used to try to leave myself notes at night, a habit I called “close to open,” after the restaurant shift that sets the tables for brunch. “The only problem is I’m forgetful, so a lot of times I would open up, and no one had closed.” Now the agent closes for me, and a digest lands at 7:00 every weekday morning. More in [AI executive assistant](/articles/ai-executive-assistant).

What does an AI agent team do that one assistant can’t?

An AI agent team splits the work so each agent does one job well and hands off the rest. When I built my video editor agent, I’d plan the design with my executive assistant. The assistant handed my specs to the video editor, and while it built the backgrounds, we moved on to the next design. Handoffs between my agents are action items only, never FYI. More in [AI agent team](/articles/ai-agent-team).

How does a research agent prep a sales meeting?

A research agent builds a brief on the person and the business before you walk in. I used to Google a business about 15 minutes before a meeting and wing it. Now the brief takes about 5 minutes and gives me six questions, built on the Sandler pain funnel, to ask before I leave. Every fact cites where it came from, and anything it can’t find is marked “unknown, ask.” More in [AI research agent](/articles/ai-research-agent) and [sales meeting prep](/articles/sales-meeting-prep).

What does an inbox agent do with email?

An inbox agent sorts incoming email and drafts a reply for a person to approve. Mine reads 10 mailboxes without changing anything in them. It sets aside out of office notices, auto replies, and bounces, then sorts real replies by intent: interested, questioning, objection, rejection, referral, or unclear. “If it’s not one of a few set issues that we have a pretty clear template on, my judgment is the bar that needs to be met.” Only my click sends. More in [AI email assistant](/articles/ai-email-assistant).

Why does shared memory matter?

Shared memory is one business brain every agent reads, so nobody explains yesterday twice. Before my 4 level business brain, I spent about 10 extra minutes at the start of every session catching the agent up. Now every session starts with this week’s decisions, who I am, and what’s due. After every conversation, a short note lands in a daily log, and a Sunday cleanup archives old notes without deleting anything. More in [AI second brain](/articles/ai-second-brain).

What is a custom agent good for?

A custom agent is built for one job with its own written rules. My business coach agent owns pricing, positioning, and go to market, with nine research files as its standing knowledge. I needed it because the operations consultant I talk to had never used AI for operations processes, so the coach owns my blind spots. Finn, a benefits Q&A chatbot, is another: it answers from about 300 emails’ worth of real questions and answers plus all of an agency’s plans. More in [custom AI agent](/articles/custom-ai-agent) and [benefits chatbot](/articles/benefits-chatbot).

What do human approval and guardrails catch?

Human approval and guardrails decide what an agent can draft, what it can send, and what it can never touch. I rejected 23 of 26 follow up drafts because I hadn’t built the part that attaches the deck I’d gone over in the meeting. The voice rules catch the words I never want in anything with my name on it. Reply agents can read and write drafts, and I took away their send permission. More in [human in the loop AI](/articles/human-in-the-loop-ai).

Where should an agency start?

Start with shared memory and one research agent. Memory is what every other agent reads, and a meeting brief is useful at your very next meeting. The inbox agent comes after the memory knows your agency.

Quick answers

What does shared memory save?

For me, about 10 minutes at the start of every session.

Can agents send email on their own?

Mine can’t. Every draft waits for a click.

Do I need all seven?

No. Most agencies should start with one or two.

What do they run on?

Claude, the AI model from Anthropic.