For this you might consider looking at the academic work of OLI and their new spinoff, Acrobatiq:

http://acrobatiq.com/ 

http://oli.cmu.edu/ 

I would consider it essential to include confidence intervals based on the conditional standard error of measurement from the maximum likelihood theta estimate.  But I assume that most places just provide the point estimate of theta, overall and on subscores.  And some places even less than that, if the purpose of the test is only categorical classification.

The PROMIS Initiative of the NIH has extensive work on CATs for a range of domains. You could do a demonstration from their website and see what information they report for an individual. They are not a company but should provide a good example of best practice. (You can also use their CAT systems free of charge)

http://www.nihpromis.org/default