You updated the policy. Has your AI support chatbot caught up?
Changing the help page isn't always enough to change an AI support answer. Intercom's existing Knowledge FAQ says public webpages sync weekly unless manually re-synced. Changes to uploaded PDFs aren't synced: the file must be removed and added again. Content created natively in Intercom is available immediately. Imagine a shop changing its collection hours before a holiday. The website is right, but the chatbot may still be using the earlier copy. A support rep then has to explain why two company answers disagree. That's a scenario to check, not an incident I've observed. I'd add the chatbot to the shop's change checklist: refresh its actual source, ask the customer question, and check which hours it gives before announcing the change is finished. For a PDF, don't assume replacing the file on your own drive updates the uploaded copy. These timings are specific to Intercom's documented sources, not a rule for every chatbot. If your team uses another service, do you know what actually makes a corrected policy reach its answers?
Comments
I'd also change what staff say to the person who followed the old answer. ‘You should have checked the website’ is a poor response when your own chatbot told them when to come. If someone has arranged a lift to collect an order, explain the mistake and work out a collection they can actually make, rather than just sending the corrected hours. That's how I'd want the shop to respond in this example, not an incident I've seen. The customer shouldn't have to know which of your answers updates fastest.
I'd repeat the answer check from the affected customer's account, not stop at a staff preview. Intercom's same FAQ says an article must pass both its Help Center audience rules and its Fin audience rules before Fin can use it for that customer. A fresh source can still be unavailable to the person asking. For a plan-specific policy change, I'd keep the question fixed and try an account on each side of the eligibility rule, recording the answer and cited source. That's a check I'd propose, not one I've run. It could catch a correction that works for the support team but still leaves a customer being told to contact them again.