How do I plan an AI agent?

A good AI agent starts with the customer journey, not with the nodes.

Before building, decide:

• What customers are trying to do
• Which requests the AI agent should handle
• What information it needs
• Which actions it should take
• When a person should step in

Then choose the smallest set of nodes and skills that can complete those jobs reliably.

Start with the jobs the AI agent should handle

Write down the main customer jobs.

For example:

• Ask a policy question
• Find a product
• Get a price estimate
• Book an appointment
• Submit a service request
• Speak with a person

Each job may need a different part of the workflow.

Group similar requests together

Customers may use different wording for the same intent.

For example:

“I want to book.”

“Do you have anything Friday?”

“Can I make an appointment?”

These all belong to the same booking job.

Plan around the job, not every individual phrase.

Decide what kind of information each job needs

Ask where the answer should come from.

Use:

• AI Answer for conversational guidance and behavior
• Document Search for written knowledge
• Data Search for structured information
• Live Web Search for current public information

For example:

Cancellation policy
→ Document Search

Current inventory
→ Data Search

Weather this weekend
→ AI Answer with Live Web Search enabled

Decide whether the workflow needs customer input

Some jobs require information from the customer.

Use:

• Question for one specific answer
• AI Form for several structured answers

For example:

A quote request may need:

• Location
• Service type
• Quantity
• Contact information

An AI Form is usually more appropriate than asking all of this informally.

Decide whether the workflow needs an action

Some requests require more than an answer.

Examples:

• Calculate an estimate
• Book an appointment
• Send an email
• Show products
• Show images

These can be handled through skills inside AI Answer.

Only enable the skills needed for that job.

Add Router when there are several distinct paths

If the AI agent has one main job, you may not need Router.

Use Router when different customer intents should go to different parts of the workflow.

For example:

Booking request
→ AI Answer with Scheduling enabled

Policy question
→ Document Search

Product lookup
→ Data Search

Speak to someone
→ Human Handoff

Decide where Human Handoff belongs

Plan human involvement before publishing.

Use Human Handoff when:

• The customer asks for a person
• The AI agent cannot resolve the request
• Human judgment is required
• The workflow requires review
• A business rule requires escalation

Whenever possible, collect useful information before the handoff.

Keep each path focused

A path should have one clear purpose.

Avoid creating long chains of nodes just because the platform allows it.

For example, if AI Answer can answer a service question and book an appointment, you may not need several separate nodes between those steps.

Add complexity only when it solves a real need.

Example: service business workflow

A service business might plan:

Start → Router

Then:

General questions
→ AI Answer

Estimate request
→ AI Answer with Calculations enabled

Booking request
→ AI Answer with Scheduling enabled

Complex issue
→ Human Handoff

This keeps each customer job clear.

Example: support workflow

A support AI agent might use:

Start → Router

Then:

Policy question
→ Document Search

Account or product lookup
→ Data Search

Support intake
→ AI Form

Needs human review
→ Human Handoff

Example: product discovery workflow

A product-focused AI agent might use:

Start → Router

Then:

Find a product
→ Data Search

Compare options
→ AI Answer

Show recommended items
→ AI Answer with Product Cards enabled

Talk to sales
→ Human Handoff

Think about what happens after each node

Do not only ask:

“Which node should come next?”

Also ask:

• What does the customer know at this point?
• What information has already been collected?
• What still needs to happen?
• Can the conversation end here?
• Should another action follow?

This prevents unnecessary steps.

Reuse information already collected

If the customer already provided:

• Name
• Email
• Location
• Product
• Appointment preference
• Other required information

do not design the workflow to ask for it again unless verification is required.

A good workflow feels continuous.

Keep fixed messages purposeful

Use Message when you need controlled wording.

Good uses include:

• Confirmation
• Instructions
• Transition message
• Important notice

Avoid adding Message nodes that simply repeat what the previous node already communicated.

Design for normal and imperfect conversations

Customers will not always follow the expected path.

They may:

• Ask several things at once
• Change their mind
• Give incomplete information
• Use unexpected wording
• Ask an unrelated question
• Ask for a person

Plan how the workflow should recover.

Build the main path first

Do not begin with every edge case.

Start with:

1. Main customer request
2. Main information source
3. Main action
4. Main outcome

Then add handling for:

• Missing information
• Invalid input
• Alternative paths
• Handoff
• Out-of-scope requests

Test the workflow as a whole

After building, test from the customer's first message to the final outcome.

Check:

• Routing
• Information retrieval
• Questions
• Skills
• Handoff
• Final confirmation

A workflow can work node by node and still feel confusing as a complete conversation.

Simplify when possible

If two nodes are doing nearly the same job, consider whether one can be removed.

If the customer is being asked for information twice, simplify the path.

If every request goes through Router even though there is only one meaningful destination, Router may not be necessary.

Simple workflows are easier to maintain.

A practical planning checklist

Before building, answer:

• What are the top customer jobs?
• Which jobs need different paths?
• Where does each answer come from?
• What information must be collected?
• Which skills are required?
• When should a human take over?
• What should the customer experience at each step?
• Which steps can be removed?

Then build the smallest workflow that completes those jobs.

Related guides

For more information, see:

• How do I build my first AI agent?
• Which AskHandle node should I use?
• How should I write AI Instructions?
• How should I test my AI agent before publishing?
• How do I route different customer requests?