AI techniques are widespread in banks since things like optimisation and search have a lot of background in machine intelligence research.
To my certain knowledge all the main firms use AI in their prop trading, as do a lot of the smaller ones.
joel_b talks of 'trust'and there are several ways of looking at this.
Today each of the firms in the original will put billions through code directly descended from algorithms to be found in AI textbooks.
That's not quite the same as "trusting" them. Banks don't do 'trust', at least not if they want to stay being a bank long.
Independent algorithms (hereafter called KSs) are treated much like trader, with limits and oversight. If a KS seems to be losing money with no prospect, it will be curtailed and just like a trader bad losses can be fatal, but at least a trader is rarely actually deleted.
Like an organic trader a KS will often be executing flow trades, and market impact models often don't work all that well without some AI-derived techniques.
Some techniques are as skinnymonkey says really slow. I recall looking at one technique at university and saying something like "you'd need a processor running at a gighertz and literally tens of megabytes of RAM to make that work"....
Actually that's why so much of advanced programming look like AI, trying to make things that have exponential time complexity go faster and it is no coincidence that the AI community got into parallel processing so early. Functional language like Lisp or F# are inherently easier to spread over multiple CPUs.
Also there are code generation techniques like genetic programming. To an extent it doesn't matter too much how long they take because you are executing the code they build for you not the host itself.
Similar things apply to FPGAs where you can code a neural net that is pre-calibrated and runs really quite quickly, and if you are Real Programmer you'd already be looking at overcoming the issues of GPUs to do NNs on them.