I built a super simple system a while ago that simply kept up the conversation as long as it was 'on track', the first time it would deviate it would engage a human on the back-end and record the answer extending the track just a little bit further.

After a couple of months of training it it could keep up a conversation for a few minutes along most of the tracks that it had learned. To make it look more life-like I added random delays in answering and sometimes splitting an answer across two messages. Spelling errors were retained as well instead of being corrected.

Of course that's not a 'bot', merely a recorder and replay but it was at least as effective as long as the conversation was kept relatively short.