If you sample at random times in relation to the input signal, and you take a limited number of samples, then your result will neither be accurate nor consistent, unless you just, by random luck, happens to take your samples at times when you do catch the true maxima. At that specific time, you'll manage to be accurate (in case the ADC is calibrated), but you'll still not be consistent since your "good" reading was just accidental.

For consistent results, you either have to sample at a speed way higher than the input signal, so that you always capture enough information to compute the relevant parameters of the input waveform (you can optionally manage that while capturing less than a full period of the input signal, if you know that it is a pure sine wave with a DC offset).

Or you need to have a knowledge about the input waveform, and a way control the sampling rate/time offsets, to allow you to combine the samples from multiple periods of the input data before computing the relevant parameters of the input waveform.

Or thirdly, if you sample at random times, and collects enough samples, and knows the that the input signal is a pure sine wave with a constant DC offset, you can use statistical methods to compute the DC component and the amplitude of the AC component even if never actually getting any sample at a maxima - this is possible since the dV/dt is different in different parts of the curve, so you'll get a higher percentage of samples close to the maxima/minima, and a lot fewer samples close to the crossing of the DC offset. The distribution of the amplitudes from the samples (without bothering about the time stamps when they were measured) will be enough for the computation, and you can also compute a statistical confidence.