Your AI replied something wrong on WhatsApp: how to correct it and prevent it
Correct the customer first, by replying in the same chat from the WhatsApp Business app; on a Coexistence number that reply also makes the automation step back on that chat. Then open the run to see what the AI read and wrote, add the missing fact or rule to its instructions, test the same question again, and switch the automation off meanwhile if the mistake could reach other customers.
easyCoexistence logs each AI reply on your WhatsApp number for 30 days with the text sent and the tokens used, pauses the automation on a chat as soon as you answer from the WhatsApp Business app, and lets you test a corrected prompt without sending anything.
Default pause on a chat after you answer it from the phone, adjustable from 0 to 168 hours.
Run log kept with the customer's message, the exact reply sent and the AI tokens used.
Default chat text retention (pick 30, 90 or 365): an uncorrected wrong reply can reappear among up to 20 earlier messages the AI sees.
Default Longest reply (tokens) for an AI reply; an answer that reaches it is cut off. Range 50 to 2,000.
Default cap on AI replies in one chat, from 1 to 60; replies over it are skipped.
What should you do first when the AI got it wrong?
Answer the customer yourself, in the same chat, from the WhatsApp Business app: say what is right, briefly, and apologize if the mistake cost them something. The reply goes out from the same number, and the automation then stays quiet on that chat for the hours in "Pause on a chat after I answer it (hours)", 4 by default. Your reply is also stored as the business's side of the chat, so if the AI answers that customer again it reads your correction, as long as it is among the earlier messages it reads. If the same wrong answer could reach other customers, such as an old price, switch the automation Off in the Automations list until it is fixed.
How do you find out why the AI said it?
Open the automation, then Runs, and find the run by the customer's message and time. The Reply with AI step shows the exact text that was sent and the tokens it used, in and out. Read the customer's message as the model saw it, including placeholders for media, then compare the answer with your Instructions: either the fact is missing, it is wrong, or the model ignored it. Most wrong answers trace back to one of a few causes, and each has a direct fix:
| Likely cause | Fix | |
|---|---|---|
| Wrong price, hours or address | The Instructions are out of date, or say it in two places | Correct the fact once, or link a live price list |
| Invented a policy or a promise | The topic is missing, so the model filled the gap | Add the policy, or a line that hands the topic to a person |
| Answered something off-topic | No scope in the Instructions | Tell it to stay on your business and decline the rest |
| Cut off mid-sentence | The reply reached Longest reply (tokens) | Ask for shorter answers, then raise the cap if needed |
| Misread an audio note or photo | Without an OpenAI or OpenRouter key it sees only "[voice message]"; it sees a photo only as "[image] caption" | Tell it to ask the customer to type the question |
| Repeated an earlier mistake | Its own wrong reply is still in the chat history | Correct it in the chat, or lower Earlier messages it reads |
How do you change the instructions so it does not happen again?
Fix the fact, then add the rule and an example of the right answer. A model follows a concrete example better than a general warning, so a short block of corrections at the end of the Instructions works well: the question as the customer wrote it, the answer you want, and what to do when unsure. Edit it in the editor and press Save, or ask ChatGPT or Claude to make the change through the MCP connector; either way it applies from the next message. For an AI agent, add the fact to its Instructions or its knowledge.
CORRECTIONS (keep these exact)
- "Is delivery free?" -> No. Delivery costs R$ 8 within 5 km. Pickup is free.
- "Can I pay later?" -> No. Payment by Pix or card when the order is placed.
- "Do you have it in stock?" -> You cannot see stock. Say: "Let me check
with the team, we will reply here shortly."
- Never offer discounts. If asked, say Ana decides discounts herself.Which settings keep a wrong reply small?
Four numbers on the automation decide how far one mistake can travel, and a fifth choice decides which answers the AI gives at all:
- "Longest reply (tokens)", 600 by default: shorter answers carry fewer claims that can be wrong.
- "Earlier messages it reads", 0 to 20: fewer means an old mistake falls out of context sooner.
- "AI replies per chat per hour, at most", 10 by default: a confused exchange stops at the cap.
- "Pause on a chat after I answer it (hours)", up to 168: a longer pause keeps the AI off chats you took over.
- A keyword automation with fixed text for facts that must be exact, such as a payment key: keyword automations answer before the AI.
How do you test the fix before customers see it?
Run the exact message that went wrong. Open the automation, press Test, paste the customer's words into "Run test", and compare the new answer with the one in the old run. The AI really answers on your key, so the result is what a customer would get, but nothing is sent to anyone. A test reads only the sample, so add any earlier context to it; the editor's Try it as a customer keeps the turns of its own chat. Try three or four variants of the same question, including a rude one, then switch the automation On again and read its next live runs. Your AI can run the same check with test_automation.
When should a fixed reply replace the AI?
When the answer must be identical every time and a wrong word costs money or trust: a payment key, bank details, an address, a price list link, legal or medical disclaimers. Put those in a keyword automation with "Reply with text", which you check once and which never paraphrases, and leave open questions to the AI. Because keyword automations are matched first and only one automation answers a message, a customer who writes "pix" gets your exact text, and one who asks something else gets the AI.
How to set it up
- Correct the customer from the phoneReply in the same chat in the WhatsApp Business app; the automation pauses on that chat, 4 hours by default.
- Stop the spread if neededIf other customers could get the same wrong answer, switch the automation Off in the Automations list.
- Read the runOpen the automation, then Runs, find the message, and read the AI step's output and token counts.
- Fix the instructions or settingsCorrect the fact, add the rule and an example answer, or change the reply length or history, then press Save.
- Test and turn it back onUse "Run test" with the same message, then switch the automation On and read the next live runs.
Common mistakes
- Fixing the instructions and not the chat. The customer still holds the wrong answer, and the AI can read its own reply as context for as long as it stays among the earlier messages it reads.
- Raising Longest reply (tokens) to stop answers being cut off when the prompt asks for long answers. Ask for two or three sentences first; the cap is a safety net.
- Letting the AI quote payment details from its instructions. One paraphrased digit sends money to the wrong place; answer payment questions with a keyword reply you check once.
- Switching the automation Off and forgetting it. Messages that arrive while it is Off start no run and are not queued for later.
Your AI can read the runs and rewrite the instructions through the MCP connector. US$ 9 down to US$ 0.75 per number per month, falling as you add numbers, with the first 3 days free.
easyCoexistence is a multichannel API and MCP server that connects AI agents such as Claude and ChatGPT to WhatsApp, Instagram, TikTok, Facebook and more. On WhatsApp it uses Coexistence: the business keeps using its number in the WhatsApp Business app while the AI works through Meta's official API.
Frequently asked questions
Does the AI learn from my correction?
Not as training. It reads your correction as part of that chat for as long as it stays among the earlier messages it reads, up to 20, and other chats never see it. To change how it answers everyone, add the correction to the Instructions.
Do I get an email when the AI says something wrong?
No. A wrong but delivered answer is a Done run, and emails go out only for failed runs. Read the Runs tab, or ask your AI to list today's runs with list_automation_runs and flag the answers that look off.
Why did the AI answer the customer again after I corrected it?
The pause after your reply ended, 4 hours by default, and the customer wrote again. Raise "Pause on a chat after I answer it (hours)" under Advanced, up to 168, if you want to keep chats you took over for longer.
Can wrong AI answers hurt my number?
They can, indirectly. Customers who feel misled block and report, and Meta's quality rating for the number is built from blocks, reports and similar feedback over the last 7 days. Fast corrections and a scoped AI keep that risk low.
Keep reading
Ready to get started?
Set up WhatsApp Coexistence in minutes, not months. The app keeps working on the phone.
Start Free TrialVerified on