Computers in the dugout?

By now, we're used to pretty detailed use of statistics to evaluate past contributions of players and help estimate future contributions.  A savvy GM using this information should be able to value players better than an old-school GM.

I wonder to what extent managers should be, or are, using computers and advanced analysis for making in-game tactical decisions.

Think about the most common decisions that a manager makes when on offense:
Whether to pinch hit, and with whom...
Whether to bunt...
Whether to hit and run...
Whether to flash the green light for a steal...

Many of us have seen charts showing expected runs for different situations (men on first and second, one out).  And I think some folks have done analysis of overall chances to win (home team, up by 2 runs, bottom of the 7th, 1 out, nobody on).  

Often, the use of such charts shows that certain small-ball decisions (bunting) are rather marginal - they hurt the team's overall chances of winning.

But frankly, I question such broad analysis.  Too often, these types of charts, and the analysis that stems from them, appears to be based on broad averages.  Perhaps having a 60% base stealer attempt a steal is a bad decision in most cases.  Perhaps it's even a bad decision in the current situation, analyzed superficially (using broad averages of thousands of different baseball games in the past).  But if we are playing in a big, pitcher-friendly park, down by 1, two outs, a very good opposing pitcher on the mound, but we know that the opposing catcher has a weak arm, then all those odds are shifted around rather significantly.

Of course, a manager will try to take much of this into account.  But wouldn't it be better to feed all this information (or as much as is reasonably feasible) into a computer, and have the computer feed the manager a live, quick list of options? (Computer shows - attempting a steal in this situation raises our win probability from 33.1% to 34.8%.)

I think I've read that TLR has notecards with batter on pitcher stats (batter X is 5 for 9 against pitcher Y, lifetime).  What if all this data was readily available by computer, but with a somewhat stronger analysis behind it, that takes into account things like sample size (i.e. 5 for 9 is nice, 50 for 90 is much nicer) and other tendencies (batter X is weak against fastball pitchers who can throw 92+ mph).

Again, I'd imagine that managers do take some of this stuff into account, but it's probably a little more gut/hunch driven.  

I can see why managers might not like the concept - it would feel like their jobs are being taken over, in part, by a computer.  But I think a good manager would be able to work with the guy running the system to design it well (which tendencies truly matter and which are more likely random), and would be using the system's output merely as a tool (albeit a very useful one).  If the manager knows that the guy on first has had knee issues for the last few games, the manager may choose not to flash the green light or call a hit and run even if the computer likes the odds for those.

Anyways, does anyone know if any teams are doing anything like this now?