industry · product
How to Write AI Agent Instructions for Customer Service
Fewer rules and no contradictions: a practical guide to the instructions that decide how your AI agent talks to customers and when it hands over to your team.
Good AI agent instructions for customer service describe behavior, not facts, and they never contradict each other. The reason is measurable: in the IFScale benchmark, which tested 20 models from seven providers, even the best frontier models followed only 68% of instructions when given 500 at once, and the most common failure was simply skipping a rule. Every rule you add competes with every other rule for the model's attention. Write the few rules that matter, in plain language, and move everything else to the knowledge base.
This guide is for the person who opens the "instructions" box of an AI agent and wonders what to type. It sits next to two posts we have already written: how to train an AI agent on your business knowledge covers what the agent knows, and how to stop your AI agent from hallucinating covers the safety stack around it. This one covers how the agent behaves.
What is the difference between instructions and knowledge?
Instructions tell the agent how to act. Knowledge tells it what is true.
| Put it in the instructions | Put it in the knowledge base |
|---|---|
| Tone and voice ("warm, short, no exclamation marks") | Prices, packages, and service menus |
| What to do when a customer is angry | Opening hours and holiday closures |
| When to hand over to a person | Return and refund policy wording |
| What the agent must never promise | Product specs, sizes, ingredients |
| Which details to collect before booking | Addresses, parking, directions |
The most common mistake we see is pasting the whole price list, the FAQ page, and the cancellation policy into the instructions. It feels thorough. It makes the agent worse, for two reasons.
First, facts in a prompt go stale silently. You change a price on your website, forget the prompt, and now the agent quotes the old number with full confidence. Second, long prompts dilute attention. Chroma's 2025 "Context Rot" report tested 18 models and found that performance degrades as input length grows, even on simple tasks. A Stanford-led study published in TACL found models use information best at the beginning or end of a long input and worst in the middle. Your one critical rule, buried on line 140 between two price tables, is exactly where a model is weakest.
How many rules should an AI agent's instructions have?
Fewer than you think. The IFScale numbers are worth looking at closely because they show how agents fail as rules pile up:
- Accuracy drops with density. Several strong models stayed above 94% at 100 instructions, then fell below 50% at 500.
- Earlier rules win. The authors found a consistent bias toward instructions that appear earlier in the prompt.
- Omission is the main failure. Models mostly did not break rules; they quietly dropped them.
One honest caveat: IFScale used simple "include this keyword" instructions in a report-writing task, and the authors say plainly the results may not transfer to every domain. Customer service rules are harder to follow than keyword inclusion, so if anything, expect real-world adherence to be worse, not better.
The practical takeaway for you: if your instructions read like a 40-item policy manual, the agent will follow most of it most of the time, and you will not know which rules it dropped until a customer finds out. Aim for a short list of rules that would genuinely cost you money or trust if broken. Put the most important ones first.
How do you write AI agent instructions, step by step?
1. Start with the job, in one sentence
"You answer Instagram and WhatsApp messages for a dental clinic, help people book cleanings and consultations, and hand anything clinical to the front desk." That sentence does more work than a page of adjectives. Anthropic's own prompting guide suggests treating the model like a brilliant but new employee who lacks context on your norms. Write what you would tell a new hire on day one.
2. Describe the voice with examples, not adjectives
"Friendly and professional" means nothing; every business writes it. Instead, give two short sample replies in the voice you want. Show the length too. If your customers write three words on WhatsApp, a five-paragraph answer reads as a robot.
3. Give the reason behind each hard rule
"Never quote a final price for implants" is a rule. "Never quote a final price for implants, because the price depends on an X-ray the dentist has not seen yet; offer a consultation instead" is a rule the agent can apply to cases you did not foresee, like crowns or bridges.
4. Write the refusals and handovers explicitly
Decide what the agent does when it should not answer: medical or legal advice, complaints about a specific staff member, refund disputes, anything involving a payment problem. For each, say what it tells the customer and what happens next. "Tell them a team member will reply within business hours, then hand the conversation to a person" is clear. "Escalate when appropriate" is not.
5. Say what to collect, not how to interrogate
If a booking needs a name, a phone number, and a preferred day, list those three fields. Let the agent gather them naturally across the conversation rather than scripting a fixed question order. Customers volunteer information out of sequence, and a rigid script breaks the moment they do.
6. Check for contradictions before you save
This is the step almost everyone skips. Instructions grow over months: someone adds "always offer the 10% first-visit discount" in March, and someone else adds "never mention discounts in DMs" in July. The agent now has two rules that cannot both be true, and it will pick one, inconsistently. Read the whole prompt top to bottom after every change, not just the line you edited.
7. Test with the messages you actually get
Pull ten real messages from last month, including the awkward ones (the angry customer, the one written in Spanish, the person asking for a discount, the "are you a bot?"). Run them through the agent and read every reply. Change one thing, then run the same ten again.
What do bad and good instructions look like side by side?
| Weak instruction | Stronger version |
|---|---|
| "Be helpful and friendly." | Two sample replies showing the exact tone and length. |
| "Don't give wrong information." | "If the answer is not in the knowledge base, say you will check with the team and hand over." |
| "Escalate complaints." | "If a customer is upset about a past visit, apologize once, do not argue, and hand over to a person." |
| "Our hours are 9 to 6, a cleaning costs..." | Move hours and prices to the knowledge base. |
| "Always upsell." | "After a booking is confirmed, mention the whitening offer once. Never repeat it if declined." |
Why do instructions matter legally?
Because what your agent says is what your business said. In Moffatt v. Air Canada (2024), a British Columbia tribunal held the airline liable after its website chatbot told a customer he could claim a bereavement discount retroactively, which the airline's actual policy did not allow; the tribunal rejected the argument that the chatbot was responsible for its own actions. Clear refusal rules ("never describe a refund policy that is not in the knowledge base") are cheap insurance.
How Entagl handles instructions
A few design choices in Entagl follow directly from the research above:
- Facts and behavior live apart. Services, products, FAQs, and hours sit in your business data and knowledge base, so the instructions can stay short and focused on behavior.
- Specialists instead of one giant prompt. The Receptionist runs a team of specialist agents (booking, products, FAQs, sales, and more), each with its own narrower job, so no single prompt has to hold every rule.
- A contradiction check on edits. When you ask the AI Manager to change your agent's instructions in plain language, it checks the full prompt for existing rules the change would contradict or duplicate and reconciles them before the change goes in, asking you which behavior to keep when two rules genuinely clash. That is step 6, done for you.
- Handover rules you set. You decide when a conversation goes to a person, and your team picks it up in the shared inbox with the full history.
- A test panel that shows its work. The Test Agent panel traces what the agent did on each reply, so when a test answer looks wrong you can see why instead of guessing.
None of this replaces your judgment about what the rules should be. It makes the rules you write easier to keep short and current. If you are planning a launch, our guide to launching an AI agent without losing customers covers the rollout side.
FAQ
What should I put in my AI agent's instructions?
The agent's job in one sentence, sample replies that show your tone, the handful of hard rules that protect money and trust (with the reason for each), what to collect before booking, and exactly when to hand over to a person. Leave prices, hours, and policies to the knowledge base.
Is a longer prompt better for an AI customer service agent?
Usually not. Research such as IFScale and Chroma's Context Rot report shows instruction-following and accuracy drop as prompts get longer and denser. A short prompt with clear priorities beats a long one where the important rule sits in the middle.
How often should I update AI agent instructions?
Whenever your process changes, and after reviewing real conversations that went wrong. Our guide to keeping an AI agent accurate after it goes live covers that review loop. Each time, reread the whole prompt for rules the new one contradicts, then re-run a fixed set of test messages.
Can I write AI agent instructions in my own language?
Yes. Write them in whatever language you think in. Entagl's agent replies to customers in the language they write in, across 100+ languages, regardless of the language your instructions use.
Do I need prompt engineering skills to set up an AI agent?
No. The skills that matter are the ones you already use to train a new employee: say what the job is, show examples, explain the reasons behind the rules, and check the work.
Start with the ten messages
Before you rewrite anything, pull ten real customer messages and read how your agent answers them today. Most instruction problems show up within those ten. If you want to see how Entagl checks new instructions against old ones before they go live on WhatsApp, Instagram, Messenger, and your website, book a 30-minute demo.
Sources: Jaroslawicz et al., "How Many Instructions Can LLMs Follow at Once?" (IFScale, arXiv 2507.11538, July 2025); Hong, Troynikov and Huber, "Context Rot: How Increasing Input Tokens Impacts LLM Performance" (Chroma, July 2025); Liu et al., "Lost in the Middle: How Language Models Use Long Contexts" (TACL, 2024); Anthropic, "Prompting best practices" (Claude Platform Docs); Moffatt v. Air Canada, 2024 BCCRT 149, as summarized by the American Bar Association.