Design Your Custom Agent

Find the workflow

Identify the work worth giving to an agent. The best candidates have clear patterns, a specific domain, and enough volume to be worth automating.

What work does your team handle repeatedly?

Example: Refund Exception Agent The billing team handles ~200 refund exception requests per week. Each case requires an agent to review order history, check refund eligibility, apply policy rules, and either approve or escalate to a supervisor.

Why is this worth targeting? What would it mean for your team if this ran on its own?

Example: Refund Exception Agent Each case takes 8–10 minutes, requires 3 trained agents, and approval delays are a top driver of customer escalations. Automating even 60% of cases would free up significant capacity.

Describe it

Map out how the work actually runs — who's involved, what decisions get made, and where it gets complicated.

What does this workflow look like in practice?

Example: Refund Exception Agent A customer submits a refund request. An agent reviews the order history, checks the refund policy, determines eligibility, and either approves or escalates to a supervisor.

Who does this work today, and roughly how long does it take?

Where does it get complicated? What requires judgment, exceptions, or context that changes case by case?

Check the opportunity

Custom Agents work best when a job requires context, judgment, exception handling, or interpreting information from multiple sources and formats — not just fixed rules. The strongest opportunities are focused, repeatable capabilities that can be defined once and reused across workflows, AI Agents, Agent Copilot, or human-agent experiences.

Does this work involve different types of information — such as written requests, structured records, images, attachments, or customer history?

Example: Refund Exception Agent Yes — agents review written refund requests alongside order records, payment history, and sometimes attachments like receipts or screenshots. No single source tells the whole story.

How often does this happen, and what's the current cost? Think: volume per week, time per case, headcount involved, errors, customer wait.

Define the agent's job

Be specific about what the agent does — and doesn't do.

Who does the agent help?

What should the agent do for them?

Example: Refund Exception Agent Review incoming requests, surface relevant policy, recommend a resolution, and draft a response for agent review.

What should the agent never do? Decisions, actions, or communications that always stay with a human.

What does success look like? How will you know it's working?

Example: Refund Exception Agent Handle rate increases, time-to-resolution decreases, escalation rate holds steady or drops.

What the agent needs

Define the inputs that make the work possible.

What data or systems does the agent need access to?

Example: Refund Exception Agent Order management system, refund policy documentation, customer history, CRM notes.

What rules or guardrails should it follow?

Example: Refund Exception Agent Never approve refunds over $500 without human review. Always check for active disputes before resolving.

What happens at the edges

Define where the agent's job ends.

When should the agent stop and hand off?

Example: Refund Exception Agent When a request falls outside policy, when a decision needs manager approval, or when the case requires human judgment to resolve.

Who or what does it hand off to — and what does it pass along?

What should always stay with a human?

Your Custom Agent Brief

Ready to share.

Drop this into Slack, a ticket, or a Confluence page as your starting point.

Agent name

Agent type — Does it tell someone what to do, summarize what it found, or complete the action itself?

Who it helps

What it does

What it needs access to

Where it stops

How we'll know it's working