Edit: Or maybe one of these bad boys? http://physicsworld.com/cws/article/news/2013/dec/09/new-gen...
But they tend to have low efficiency, thermally-dependent efficiency, and the need for a temperature differential. They operate in response to a temperature gradient, not just heat.
Using Wikipedia for a quick-and-dirty figure, you can expect efficiencies to peak around 5-8%. And you can't just stack infinite thermocouples until system efficiency reaches 99.99% because of the thermal gradient issue.
Ultimately, you don't even need thermocouples, you just need more heat engines with ever decreasing working fluid temperatures, and then it all comes down to your capital costs.
You can even make several layers of that, with thermal machines between the layers. As the number of layers increase, the efficiency of each one reduces, but the total work goes always approaching a constant assintote.
If there's a lot of it and you're not making it go elsewhere, assume that the system and it its environs are going to get a bit toasty.
Sometimes you also get things like radiation and vibration. These mostly show up in meaningful quantities in predictable places - nuclear systems, mechanical systems, high-power A/C or RF systems, high-power transducers.
You'll usually do okay if you guess that the system uses energy at around ~60% efficiency, which is enough to give you some idea of the heat to expect if you know the input power.
Or more generally, assume all input power goes to heat and plan to handle operation at that limit - while being aware that your expectations of power consumed by the system may well be low.