The ten mistakes that make a WhatsApp AI agent fail, and the fix for each
When an AI agent disappoints, the software is rarely the reason. The setup is. The same ten mistakes account for most of the cases we see, and each has a fix that takes minutes. Read the list before you go live, or use it to diagnose an agent that is already live and not earning its keep.
Quick answer
The big ones: automating your main number instead of a dedicated one, no handover rule, feeding the agent marketing copy instead of facts, prices in images, never testing the hard messages, ignoring the unanswered-questions list, instructions that contradict each other, no reminder or follow-up, expecting it to sell rather than serve, and treating it as finished after setup. Each fix is below.
The ten mistakes
- Automating the main number. A personal WhatsApp line cannot be connected to the platform, and the number in the WhatsApp Business app should be connected as itself rather than moved. Fix: a dedicated number.
- No handover rule. The agent keeps trying to help an angry customer. Fix: name the triggers (complaints, medical, money, "talk to a person") and the exact handover sentence in the instructions.
- Marketing copy as knowledge. "We are passionate about excellence" teaches the agent to make claims. Fix: facts only, prices, hours, policies, in plain lines.
- Prices in images. A photo of a menu has no text. Fix: a text price list, one line per item, per this format.
- Testing only the easy questions. Fix: the twenty-message test, including a discount request, a complaint and a jailbreak.
- Ignoring the unanswered list. The agent tells you what it does not know and nobody looks. Fix: ten minutes every Friday.
- Contradictory instructions. "Always offer a booking" and "do not book new patients" together. Fix: read the instructions as a new hire would, and cut to twenty lines.
- No reminders or follow-up. Bookings made and forgotten. Fix: reminders on, and the one-time nudge on Growth or Pro.
- Expecting it to sell. An agent told to be persuasive pushes people away and drifts into claims. Fix: tell it to be useful and offer the next step; the booking is the sale.
- Treating setup as finished. Prices change, staff leave, a new question appears. Fix: replace documents when things change and re-run the test after any instruction edit.
Diagnosing a live agent in five minutes
Read the last twenty conversations. Count how many the agent handled fully, how many it handed over, and how many it answered wrongly or awkwardly. Handovers above about a fifth usually mean mistakes 3, 4 or 6. Wrong answers mean 3, 4 or 7. Awkward tone means 9. Zero bookings from interested customers means 8 or that services and hours were never set. ReplyKit's inbox filters and the Monday digest give you the counts without scrolling.
If you fix only one
Fix the handover. An agent that hands over correctly can be wrong about nothing that matters, because every hard case reaches a person. Everything else improves the agent; the handover rule protects the business while you improve it. The handover guide has the triggers and the wording.
Frequently asked questions
What is the most common AI agent mistake?
Automating the main business number instead of a dedicated one, closely followed by having no clear handover rule.
Why does my agent hand over so often?
Usually missing or badly formatted knowledge, or an ignored unanswered-questions list. Feed it facts and teach weekly.
Why does my agent sound like a brochure?
Marketing copy in the knowledge base or 'be persuasive' in the instructions. Replace with facts and 'be useful'.
How often should I revisit the setup?
Weekly for the unanswered list, and whenever prices, staff or policies change. Re-run the test after instruction edits.
Which mistake should I fix first?
The handover rule. It protects the business while you fix everything else.
Set it up right the first time
Dedicated number, handover rule, text price list, twenty-message test. 7-day free trial.
Sources and further reading
- ReplyKit, Will WhatsApp ban my number for automation? · how Meta's messaging limits and quality rating work
- ReplyKit Acceptable Use Policy · what we refuse to automate
- Harvard Business Review, The Short Life of Online Sales Leads (Oldroyd, McElheran, Elkington, 2011) · firms that responded within an hour were about seven times more likely to qualify the lead
ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. This article is general guidance for running a small business, not legal, financial or regulatory advice.