If you where able to automate anything in your business,
what would it be?
Very good question,
Depending on the business, the response may vary considerably.
In our case what we want to automate is the administration (invoices, payments, bureaucracy, and mechanical tasks) since they take away a lot of time and are tasks that we believe could be automated
I would love some advanced banner auto creation tool - I have something coded, but the outcome is quite simple... so I still have to use a designer 
My white whale of optimisation for a while has been a Machine Learning tool to identify "pockets of profit" within existing campaigns.
Current ML algorithms are good at prediction, but it's hard to get them to spit out recommendations. I have some approaches that have had initial success, but nothing at the level I'd like to see.
(You can always just iterate all possible variables through the predictor, of course - but the results there tend to be too granular to be useful. )
^ What type of algo are you using for that? I've played with the AWS Machine Learning thing.
Can you get anything significant from normal amounts of conversion data for machine learning, or is it just statistics?
I feed banners, landers, and offers into the machine. It tests each combination to statistical significance and spits out the best result.
communication i.e. messenger chit-chat
would love some bot that could do this for me and send me the result/outcome of the back and forth
@mindfume - check out ManyChat on FB. Impressive stuff.
@jessejames - CNNs, Random Forest, all the usual stuff. As I say, the issue isn't with prediction - you can get some great predictive results from the data we've got as it's clean by ML standards - but with persuading the ML program to say "aha, I notice that Carrier A and Lander B go together very well" rather than just predicting a win whenever Carrier A and Lander B happen to come up in conjunction.
You can force the effect by artificially throwing your trained ML program all possible combinations of variable and seeing which ones it predicts as successes, but that's still not very useful as the data you get out tends to be noisy as hell.
@priest - sounds interesting! Can you tell us more?