"bots" lack any standardized discovery interface, which eventually causes frustration for users. Two popular bots - Siri and Google Now, won't tell you everything they can do, and give such useless answers to most requests that people use them as a gimmick once or twice, maybe memorize a handful of key phrases, but otherwise won't use them.
They're also typically not extensible - I haven't come across bots that allow you to string multiple bots together to accomplish complex tasks. It's all the bad parts of a CLI with none of the good parts. No one wants to memorize a bot specific query language along with it's capabilities for something they interact with very infrequently.
Instead of sinking money and effort into natural language processing, services should be building responsive, discoverable websites with all the capabilities of a customer service agent. Work on better on-site search, and expose available actions to Google.
I am on the other side than you - I am betting on the NLP in my current interface.
The biggest reason for me is logging. I can see what are yet-unfulfilled needs of my customers and prioritise building new features. In responsive website you hardcore your assumptions about needs of your customers. What's more, you can fall-back to real customer support.
One of reasons why bots exploded recently are advances in NLP including deep learning, other new methods, cheaper hardware, etc.
NLP and logging failed queries doesn't excuse the current lack of any customization and interop. This is the same thing as with IoT all over again - everyone wants to take over the entire market for themselves, and thus everyone forgoes the only feature that can make it actually useful - free interoperability.
I agree with your conclusion (services should be focused on getting better and getting things done), but the rest of your arguments are needlessly negative.
If it lacks standardized discovery interface, the next rational thing should be to think about how to come up with one, instead of saying bots suck because of that.
That's like trying out a raw coffee bean and saying "coffee sucks, I tried chewing on these coffee beans but they all taste like shit". Someone else out there is figuring out how to add hot water to coffee beans to make it much better while you're complaining about how coffee itself sucks and moving on.
Same thing with extensibility, if you haven't seen any bot platform that lets you string multiple bots together, then great! It means there's an opportunity there. That doesn't mean bots suck.
I hope the answer to discoverability is that bots (within a single app) will be able to fulfil 95℅ of your needs, such that you're more willing to just try it since your expectation is that it will succeed.
I also only really see it taking off on mobile if voice input becomes near ubiquitous, since typing on mobile sucks.
They're also typically not extensible - I haven't come across bots that allow you to string multiple bots together to accomplish complex tasks. It's all the bad parts of a CLI with none of the good parts. No one wants to memorize a bot specific query language along with it's capabilities for something they interact with very infrequently.
Instead of sinking money and effort into natural language processing, services should be building responsive, discoverable websites with all the capabilities of a customer service agent. Work on better on-site search, and expose available actions to Google.