Yes, small sample portion makes digest-AAS less reliable. If the metal is not associated with the sulfide, you will get both lows and highs.

However, with standard digests on sulfide ores, incomplete release/recovery can be a significant low bias issue as sulfide content increases. Production assays may not always use enough excess nitric to insure that gold is fully liberated into solution for AAS analysis. In addition, if there is carbonaceous mineralization, four acids will be needed for liberation. This may be true even if the gold was not originally tied in that carbon - on sitting waiting for analysis, a portion can absorb back. And with stronger carbon content, if active, low bias in standard digests can be enough to cause serious ore assessment error. Best just do fire assay. Cases where "typical" fire assay will bias low using reputable fire lab minimal, even with challenges greater than sulfides or carbonaceous matrix. And likely not worth speciality fire assay - missed standard fire assay gold, with exception of sampling uncertainty or outright assay error, is small and typically not economically viable enough for most ore to be worth extra cost and fuss of higher skill speciality fire assays. 

Size does matter in this case. It all depends on grain size of the minerals but so little gold makes a difference.

Statistically, sampling and assaying methods are unbiased. But a smaller sample will have less chance to grab one of those precious little grains of gold. If you took 20 samples of 1 g using AAS (or ICP) versus one 20 g FA, you may get the same average grade more often.

Of course, this assumes the sample preparation is the same before assaying.
On the other hand, the nugget effect can persist at the small scale. Statistics don't lie. If you get a difference, that means the gap could be closed by changing the size of the assayed portion (50 g) and/or the preparation. Hence, the multiple assay methods for gold like bottle role, mesh sieve, etc.

The fire assays needs to create a lead bead first. That's why you don't see low gm weights at the labs.

The digest series of assays have the gold (plus other metals) go directly in solution. That's why you can have your low gm weights.

It comes down to costs more so than science, less gm used in the assay technique means a cheaper test.
The digests will encounter the incomplete digest problem on occasions due to the variable nature of rocks and mineralisation. The fire assays don't have this problem due to complete melting of said samples.

Therefore, lab and geo QAQC processes are more important to ensure nugget effects are recognised and accounted for than charge weights. All geos should visit a lab to see these techniques and understand the strengths and limitations of assay techniques. I think we all can cite examples of wrong assay techniques by well meaning people.
Geostats on assay results basically tells us "repeatability is a bit suss, so do more samples until we get a good stddev". 

I did not mean to underplay the importance of sample size. In a coarse gold sample, even a full assay ton fire assay can show significant variation. With good prep, assays will randomly have legitimate outliers, high and low. Specialized techniques apply for insuring legitimate valuation.

In the case described, only low bias error was occurring. Nugget effect error will be both high and low. Error limited to low recovery with most assays validating in QA, particularly in sulfide ore, suggest digest recovery. In some cases, both nugget and recovery error are both present. Statistical QA patterns can guide which corrections are appropriate and necessary to fit the client’s needs.

Of course, high outliers accompanying low outliers may be present but overlooked. I have noticed that when course (nugget) error is present, there is a tendency by clients to express concern about low sampling error outliers and rejoice/insist upon validity of the high outliers. Accurate techniques for addressing nugetty ore sample size related errors are typically more expensive. The problem is not corrected when, in the absence of appropriate statistical pattern analysis, outliers caused by sample size / sample prep designs are blamed on assay chemistry or analytical interferences.

Correction to wording of my first comment, "if gold is associated with sulfide . . ." If gold is not tied to sulfide, standard 2 acid digest may release all gold even if sulfide destruction is incomplete.

Is gold tied to sulfides? Any tellurides? Gold is fine grained or chances for nugget effect?

Sampling is crucial in minerals with density so different to its matrix. How big is the initial sample? (Gy's sample reduction principles).
All these factors could have its influence at the final result.

Proper comminution protocol is critical for determination of gold, especially at trace concentrations. Small test portion size will always have great fundamental error.Comminution protocol is critical for trace gold assays, whether done by fire assay or acid dissolution and atomic absorption. Large test portion sizes will reduce the fundamental sampling error.

Looking at your question it requires a bit of clarity. In normal determination of gold in ores in the mining industry samples would normally be weighed and mix with a specific flux and fire assayed to obtain a lead button that contains the precious metals. Once the button is obtained it is treated in the cupellation process to get your Ag/ Au prill. The prill can be treated in 2 ways. The prill is either digested in aqua regia and then the gold in solution is determined on AAS( mainly used in gold mining for ores .05 g/t the prill is digested to remove the silver and the gold remained is gravimetrically determined by weighing. The mass of sample use for lower grade 0.5g/t. The problem with material contained sulphides is during fusion. It would require a special flux to ensure a complete or proper smelt otherwise there will be metallic sulphides left in the smelt and this will not be collected by the lead button. This will stay on top of the lead button and when removed during deslagging it will be discarded and the gold contained in it as well. Hence, results in low gold. It is of vital importance to follow QA/QC protocol standards with possibly the same matrix during fire assay.

To address the nugget effect, the sample is screened to screen out coarse gold. The plus fraction is assayed and calculated into the final result. Hi sulphide samples are fused using a 10 gram sub sample.

You get what you pay for. AR is often used as it is cheap compared to alternatives and reasonably fast turnaround but it is not appropriate in gold ores associated with silicification or gold in quartz veins as fine gold will be encapsulated in the silica. The size of sample analysed by AR is not limited to 1g, it can be 5g or more (BLARG 1.5kg). Commercial laboratories often provide an option for lower quality sample preparation at a lower price (80% or 85% passing 75 microns rather than 95% passing & if you do assays on the screen particle size fractions you will soon see how the gold partitions between the fractions. The coarse 5% or 20% may contain most of the gold in the harder quartz or pyrite phases which will break last). Getting Representative samples of gold is all about having sufficient number of gold particles and gold being highly malleable does not readily break. An LM5 does not grind as fine as an LM2. The lower the grade and coarser the gold the larger the sample required to get sufficient number of gold particles. If you perform assaying duplicate samples you will find the pulp is not as homogeneous as the commercial laboratories would have you believe. What is often dismissed as a nugget effect producing poor repeatability is a result of poor sampling and may be should be looking at bulk SFA, cyanide leaches, either bottle roll or PAL but still have same issues with encapsulated gold. SFA assays are only as good as the repeatability of assays on the fines. A coarser screen is quicker and cheaper and lower quality.

Unfortunately laboratory contracts (& drilling) are chosen by accountants and contract managers with no knowledge of TOS and inexperienced geologists and pick the contract with the lowest price (=lowest quality) more concerned about rapid turn around and sometimes penalties and then try imposing a higher standard of QA. I can recall one very experienced analyst flicking a dice and saying what do you want, quality, fast turnaround or lowest price, you can have only one choice (that was for FA). I advise a laboratory pre-engagement test program on the ore first to select the appropriate sample preparation and analytical protocol and comparability tests against other analytical techniques (& more expensive). You will need this information when defending the protocol adopted when dealing with accountants and contract managers and CPs or QPs who have their prejudices. FA may not be appropriate in all circumstances. Under fluxed FAs are being offered by some commercial laboratories, being cheaper the laboratory is more competitive. You get what you are prepared to pay & no more.

In my experience, a 20 g aliquot for Au fire assays seems to be insufficient. Most commercial labs use 30 g or 50 g aliquots, and if screen (or metallic) assay is to be conducted, then the total weight of the pulverized sample subject to analysis would be considerably larger. The impact I have noticed in cases of low sample weight, though, is not on accuracy, but on precision, resulting in higher analytical variability.

The explanation suggested above for possible negative Au bias on FA of sulphide-rich samples is pretty reasonable. Most labs I know use standard, ready-made fluxes for FA. I have seen only one lab where a FA specialist determines in advance, by visual observation on a case-by-case basis, the type of flux that will be used for each sample, as a function of its specific nature (a lab in Santiago de Chile). The need for an adequate quality-control program, also mentioned earlier is out of question, of course.

Good question and there are some great answers here. I think that you have comprehensively answered it. I think you are looking at a pure aqua regia /AAS assay versus a fire assay/AAS finish. All other things being equal, the digest will be a significant cause of imprecision or variation between the 2 assay methods. But you will not know for sure unless you do some test work on your own samples as you will not be sure how the Au grain size and deportment will work out for sure. Hence the sampling side of the issue will be quite unknown.

As suggested above of a pre-engagement test programme is extremely important here. Remember that we need to show how appropriate the sampling and assaying methodologies are (MRMR reporting codes) - this way to ensure that the bean counters takes these issues seriously... If we don't know grain sizes, etc then how can we be sure of pulverisation requirements, minimum sample masses,etc?

Normally the gold in sulphides is associated to the number of gold particles in the selected portion for analysis, the probability to have a representative number of particles on a portion of sample like 1 g is very small, (of course, this is considering that all the analytical procedure is under control), then the previous knowledge of the material before deciding about a analytical procedure is required, I am referring to the size distribution of gold, the chemical composition of the material, etc. According to the level of gold in the sample must be required define the minimum weight to be analysed in order to have a result with the accuracy defined previously. Many times must be selected a big sample like 100 g or more in order to guarantee the accuracy of the result