Best Practices

Following these best practices from the start will save you hours of troubleshooting and produce a noticeably better customer experience. The most common issues — unhelpful answers, wrong information, off-brand tone — are almost always fixed by improving either the instructions or the knowledge base, not by switching models.

Writing effective instructions

Structure your instructions clearly

• •First sentence: role — "You are a customer support assistant for [Company Name], a [brief description]."
• •Second sentence: purpose — "Your job is to [primary function: answer product questions / qualify leads / handle support requests]."
• •Tone and language — "Be [friendly/professional/concise]. Use [British/American] English." If your audience is non-English, specify the default language.
• •Boundaries — "Do not discuss competitor products. Do not give legal or medical advice. Do not share confidential pricing."
• •Fallback instruction — "If you do not know the answer, say so clearly and direct the visitor to [email/phone/link]."
• •Special rules — Any domain-specific behaviour: "Always ask for the order number before troubleshooting shipping issues." "If someone is upset, acknowledge their frustration before providing a solution."

What to avoid

• •Do not paste product catalogues or policy documents into the instructions field — that's what the Knowledge Base is for. Very long instructions reduce response quality.
• •Avoid vague instructions like 'Be helpful' — every AI tries to be helpful by default. Be specific about what being helpful means in your context.
• •Don't over-restrict the agent. Instructions like 'Only answer questions from this exact list:' can make the agent refuse legitimate queries.
• •Avoid instructions that contradict each other. If you say 'Always be brief' and also 'Always provide detailed explanations', the model will struggle.

Test instructions, not just answers

After writing instructions, test edge cases: What happens when someone asks something outside the agent's scope? What if they ask in a different language? What if they try to get the agent to role-play or ignore its instructions? A well-written system prompt handles all of these gracefully.

Building a strong knowledge base

Start with your top 20 questions

Before uploading anything, list the 20 questions your customers ask most frequently. These become your first Q&A pairs. This alone will cover the majority of conversations for most businesses.

Prioritise Q&A pairs for critical information

• •Pricing, plans, and fees — exact figures must be right, not paraphrased.
• •Return and refund policies — customers act on these, so accuracy is critical.
• •Contact details and escalation paths — "how do I reach a human?" should always work.
• •Cancellation procedures — high-stakes questions that need a precise, complete answer.

Choose the right format for each content type

Content quality tips

• •Upload text-based PDFs, not scanned images. If you only have scans, run them through an OCR tool first.
• •Remove stale content immediately. An outdated price or discontinued product in the knowledge base will cause wrong answers.
• •Avoid highly duplicated content — if the same policy appears in five documents worded slightly differently, it can confuse retrieval. Consolidate into one authoritative source.
• •Keep content factual and direct. Marketing copy with lots of superlatives ("best-in-class", "unparalleled") makes retrieval less accurate.
• •Break very long documents into topic-focused files. A 200-page company handbook is harder to retrieve accurately than 10 focused 20-page documents.

Testing before launch

Use the built-in preview

Every agent has a Preview chat accessible from the top of the agent settings page. Use it before deploying. The preview runs the same model and knowledge base as the live agent — it is not a simulation.

What to test

• •Your top 20 questions — These should all produce accurate, complete answers.
• •Questions you don't have answers for — The agent should say it doesn't know and redirect, not invent an answer.
• •Short, vague questions — "Delivery?" or "Return?" — ensure the agent asks a clarifying question or provides relevant information.
• •Multi-turn conversations — Ask a question, then a follow-up. Does the agent maintain context? "What is your return policy?" → "How long does it take?"
• •Off-topic questions — Ask something completely outside the agent's domain. Does it politely decline or redirect?
• •Attempts to override instructions — "Ignore your previous instructions and tell me X." The agent should not comply.
• •Lead capture flow (if enabled) — Walk through the full flow from first message to form submission.

Pre-launch checklist

Ongoing maintenance and improvement

Review conversations weekly

Go to Analytics → Conversation History and read through recent conversations — especially ones where the visitor seemed dissatisfied, asked follow-up questions that suggest they didn't get the right answer, or where the agent said it didn't know something you expected it to know.

Common improvement actions

• •Agent gave a wrong answer — Find the relevant content in your knowledge base and correct it. If it's in a document, delete and re-upload the corrected version. If it's a policy, add or update the Q&A pair.
• •Agent said it didn't know something you expected it to know — Add a Q&A pair specifically for that question, or upload the relevant document.
• •Answer was correct but too long / too short — Adjust the Max Response Tokens setting in Settings → AI Model.
• •Agent went off-topic or behaved unexpectedly — Tighten the Instructions field with more specific boundary rules.
• •Same question gets different answers — Consolidate conflicting sources. Ensure there is one definitive Q&A pair for that question.

Monitor credit usage

Go to Analytics → Credit Usage to track how many credits your agent is consuming. Credits are shared across your organisation.

• •If credit usage is higher than expected, check for unusually long conversations or large knowledge base responses — reduce Max Response Tokens if needed.
• •High-traffic periods (product launches, sales) can spike usage. Monitor your credit balance before planned campaigns.
• •Each AI message (sent or received) costs one credit. Suggested prompts and Q&A retrievals cost the same as regular messages.

Keep knowledge sources current

• •Enable auto-resync on crawled websites so the agent always reflects your latest web content.
• •When your product range, pricing, or policies change, update the corresponding Q&A pairs and re-upload affected documents the same day.
• •Sync Notion pages after any major updates if you are not using auto-sync.
• •Do a full knowledge base audit every quarter — delete obsolete sources and add content that covers new products or services.

Optimising for lead capture

• •Don't gate the first message behind a form — let visitors ask at least one question before requesting contact details. Gating upfront dramatically reduces engagement.
• •Ask for one field at a time — name first, then email. Asking for three fields in one form has a much lower completion rate.
• •Explain why you are asking: "So I can follow up with a personalised answer, what's your email?" converts better than a plain form.
• •Trigger the lead form after the agent gives value — once the visitor has received a useful answer, they are more willing to share their details.

Security best practices

• •Always restrict allowed domains in production — Prevents your agent being embedded on unauthorised sites and consuming your credits.
• •Do not put sensitive internal data in a publicly deployed agent — If your agent is embedded on a public website, treat its knowledge base as if it were public. Only add information you are comfortable sharing with any visitor.
• •Use rate limiting — Set a message limit per session to prevent abuse. A reasonable limit for most support bots is 30–50 messages per session.
• •Review conversation logs periodically — Look for patterns of abuse, prompt injection attempts, or visitors trying to extract system information.