If the volume of tweets you have to handle is rather high you might want to automate the handling of it. This is actually the thing Bank of America did and what went completely wrong. As you can read in this article you’ll see that the automation of Bank of America was quite poor and there was not much intelligence in it (or at least it wasn’t a showcase of smart automation).
However automating a certain workflow is rather common, it has become part of our life to automate repeating tasks. Therefore the archetype of the bot might be more common than that we think since not every bot leads to instant failure.
It would rather short-sighted to say on a sample size of one that bots are absolute failures in handling messages on social media. We have just seen one big failure and probably if you really want to you can find some more examples of companies that over-automate and under-test their social media automation. Since that was what happened to Bank of America: too much automation and too little testing. If it had been thoroughly tested these obvious glitches would not have happened.
How to make it better
Automation in social media and for Web care in particular has been always been a sensitive topic. Some might say it removes any for authenticity and that it is a sin to even think about it. However how many authenticity would you really want to have if you are asking for a simple fact based answer (e.g. When does the train to city X leave), probably not too much, you just want to have that answer as soon as possible.
Basically that is what you can now automate really well: factual questions. First all a factual question can only be asked in a limited number of ways, second there is no wrong answer on the factual questions since it is just informing one who asked the question about the facts he is looking for. However before going completely bezerk on automating every factual piece you need to make sure it actually works. Therefore you should start with semi automate it. Let the system come up with suggestions and let the human pushing the button validate this suggesting and post it. This way you can train the system and validate its outcomes. As certain as the system comes to a certain success level you can push it live without any human being the intermediate between the bot and the person who asked the question.