How to Write Instructions for an AI Phone Agent

Good AI phone agent instructions are short, specific rules that tell the agent its goal, what information to collect, when to transfer or take a message, and what it must never do. Write them as rules, not a script; keep facts (hours, prices, policies) in a separate business facts section; and put the most important rules first.

Instructions vs business facts

Separate the two:

Instructions (behavior)	Business facts (knowledge)
"Collect name and callback number before ending."	"We're open weekdays 8 to 5."
"Transfer if the caller reports a gas smell."	"Service area: Dayton, Kettering, Centerville."
"Never quote prices that aren't listed."	"Drain cleaning starts at $149."
"Answer in one or two sentences."	"We don't service commercial buildings."

Facts change often; behavior changes rarely. Separating them means updating a price never risks breaking a rule. See AI agent business facts sheet.

A structure that works

Use five short blocks:

1. Goal: one sentence on the agent's job.
2. Collect: the information it needs from every caller.
3. Answer: what it can answer from the facts, and what to do when it can't.
4. Hand off: exact triggers for transfers and messages.
5. Never: hard limits.

Example: receptionist for a home-services company

Goal: Help callers get service from Harbor Plumbing or leave a message the team can act on.

Collect: Name, callback number, service address, the problem in the caller's words, and whether water is actively leaking.

Answer: Hours, service area and services using only the business facts. If asked something not covered, say the team will follow up and take a message.

Hand off: Transfer immediately if water is actively leaking, there's sewage backing up, or the caller smells gas (also tell gas callers to leave the building and call the gas company or 911). Transfer if the caller asks for a person. Otherwise schedule a callback.

Never: Quote prices not in the facts. Promise arrival times. Diagnose problems. Collect card numbers.

That's under 150 words and covers what matters.

Writing rules that the agent follows

Be specific about triggers

• Weak: "Transfer urgent calls."
• Strong: "Transfer if the caller says it's an emergency, mentions flooding, no heat below freezing, or a gas smell."

Give a fallback for every uncertainty

• "If you don't know, say the team will follow up, and take a message."
• "If the caller's request doesn't fit any of these, take a message."

Without a fallback, a model tends to guess. See stop an AI agent from making things up.

Ask for one thing at a time

Phone callers can't scroll back. "What's the best number to reach you?" then "And the address for the service?" works better than asking for both at once.

Require repeat-back for critical details

"Read back the phone number and address before ending the call." Speech recognition errors are most costly on numbers and addresses.

Keep answers short

"Answer in one or two sentences, then ask if there's anything else." Long answers cost minutes and lose callers.

Things to put in the "Never" list

• Quote prices, discounts or availability not in the facts
• Promise outcomes, arrival times or refunds
• Give legal, medical or financial advice
• Collect full card numbers, security codes, Social Security numbers or passwords. PCI DSS rules restrict storing card security codes after authorization, and recordings count.
• Claim to be a human
• Discuss competitors

Honesty about being AI

Tell the agent how to answer "Are you a real person?" A clear rule: "If asked, say you're an AI assistant for the business, and offer to connect them with the team or take a message." California's bot disclosure law and similar rules target bots that mislead people about being artificial, so never instruct an agent to deny it. See AI disclosure laws for phone agents.

Instructions for outbound calls

If the agent places calls, add:

• The purpose of the call, stated in the first sentence
• That it's an AI assistant calling on behalf of your business
• What to do on voicemail: short message, callback number, hang up
• What to do if the person asks not to be called again: confirm, end the call, and flag it for your team

The FCC treats AI-generated voices as artificial voices under the TCPA, so consent is required before calling. See AI outbound calls and the TCPA.

Testing your instructions

After writing, test against these scenarios:

• The caller asks something not in the facts
• The caller describes an emergency
• The caller asks for a person
• The caller gives an answer the agent doesn't understand
• The caller asks for a price not listed
• The caller asks if it's a real person
• The caller tries to get a discount

If the agent fails one, add or sharpen a rule. Then retest everything, because new rules can conflict with old ones. See test an AI receptionist before launch.

Common instruction mistakes

Mistake	Result	Fix
Facts mixed into instructions	Outdated facts linger	Move facts to the facts section
Vague triggers ("if important")	Inconsistent transfers	List exact conditions
No fallback	Agent guesses	"If unsure, take a message"
Too many rules	Key rules ignored	Cut to what matters; important rules first
Contradictions	Unpredictable behavior	Read rules together for conflicts
Word-for-word scripts	Robotic, brittle calls	State goals, let the agent phrase

Before and after: rewriting weak instructions

Before: "You are a helpful assistant for Bright Dental. Be friendly and professional and help callers with whatever they need. Answer their questions and book appointments. If it's urgent, let us know. Make sure to be polite."

Problems: no list of what to collect, no fallback, "book appointments" implies confirmed bookings the agent can't make, "urgent" is undefined, and nothing says what's off-limits.

After: "Goal: help callers request appointments or get answers about Bright Dental. Collect: full name, callback number, whether they're a current patient, reason for visit, and two preferred days and times. Answer: hours, location, insurance accepted and services, using only the business facts. If something isn't covered, say the office will follow up. Hand off: if the caller reports severe pain, swelling or bleeding that won't stop, mark the message urgent and transfer. If they ask for a person, transfer. Never: give medical advice, promise a specific appointment time, or discuss costs not in the facts. Close: confirm name, number and that the office will call to confirm the time."

The second version is only slightly longer but removes every ambiguity.

When to split one agent into two

If your instructions grow past a page because one agent is doing reception, sales and support, split it. Separate agents with focused instructions are easier to test and less likely to confuse rules from different jobs. See choosing an AI agent role.

Writing instructions in Callata

Callata's agent editor has separate fields for instructions and for "Business facts the agent can share," and the agent is told never to state facts that aren't in that field. Each role preset (Receptionist, Sales rep, Support rep, Appointment setter, Custom) fills in starting instructions you can edit. Callata also adds fixed rules to every agent: be concise, ask one question at a time, never invent prices, availability, policies or promises, say it's an AI assistant if asked, and confirm the caller's name and callback number before ending.

Save as a draft, test, and go live when ready. AI minutes are $0.25 per minute of talk time; Callata Office is $99 per month and includes five users. Write your first agent's instructions.