HB (under the right conditions) can be used with very small sample sizes with CBC, ACBC, and MaxDiff questionnaires. The right conditions are relatively few parameters to estimate (attributes & levels) and a relatively large number of choice tasks per respondent. For example, if those 20 people represented nearly a census of the population (such that sampling error was held at a minimum), I wouldn't hesitate to use HB if the CBC study had 5 attributes each with 2 to 4 levels, and I had 15 to 20 choice tasks per respondent.
In the early years of the introduction of HB, Greg Allenby was at our conference and the audience pushed him regarding how low one could go with sample size and HB and obtain equally good or better estimates as with traditional methods given the same sample size (which at the time meant aggregate logit). He said as few as 10 respondents, or maybe even lower, if I remember correctly. And, he felt HB would be better than aggregate logit in those situations (given the same sample size restrictions).
On my own, I have tested ACBC with subsamples as small as n=9 and gotten good results. I wouldn't expect "biased" results given small sample sizes. There is less certainty on the upper-level model in that case, so the lower-level model (individual) tends to carry a large weight. So, bias? I'm not seeing it. Noise, yes, as there always is for such models.