No need for an analog demodulator. Run the ADC synchronously with the carrier at 2MHz or a multiple thereof. 4MHz will make antialiasing much easier. Get demodulation for free. There are lots of affordable 16-bit ADCs that will run at such a speed.
If there’s little noise in the signal, you can decimate directly prior to demodulation: sample at some fraction of the carrier frequency.
In either case, you’ll run a PLL synchronized to the carrier. This can be something simple. In a pinch a selected CD4046B running from 17V will do for 2MHz PLL, although I wouldn’t do that in production.
The PLL output can be divided down to get the decimated frequency you’d need. The divisor needs to be even. The phase reference has to be adjusted to get maximum amplitude. It’s better to have both 0 and 90 degree sampling clocks and do quadrature sampling, though. That’s what you’d get by sampling at 4MHz. To do IQ sampling with pre-decimation, you’ll be taking pairs of samples 1.25us apart, and a multiple of 5us periods between the start of each pair.
The timing signals can be generated with a timer fed from an external PLL. Many MCUs have enough on board peripherals to allow a PLL implementation in software, using an external phase comparator and VCO, but internal ADC and DAC to close the loop.
—--
The signal chain is usually designed so that the noise of the demodulator can be ignored when referred back to the input of the 1st stage.
Active rectifiers based on voltage-feedback op-amps aren’t great demodulators. They are adequate, but that’s all you’ll get, since their internal nodes get off-balance when the rectifier is blocked. You can use a current-steering active rectifier that converts the input voltage into a current that goes down one of two branches, selected by current polarity. Convert current in one branch to voltage and you can get ok active rectification.
For precision you can’t really beat a synchronous demodulator either done using switched capacitors or digitally, after acquiring the signal with an ADC.
For AM demodulation that doesn’t need an analog output, it’s best to do it digitally. Even then you will get better linearity dollar for dollar by going A/D, demodulator, filter, calibration, then D/A route. You can do it all analog, but it won’t be cheap in terms of either expertise needed to design one quickly, or your experimentation time getting it to work well.
You should be equipped to objectively measure the performance. Set up test equipment that lets you excite the input in time and frequency domains (eg step response vs small signal sine response) and measure the output
