Google Document AI recorded 23.3% WER on our handwriting sample, the highest of the three big cloud providers. The individual character recognition was not dramatically worse than Azure or AWS, but it scrambled the reading order, interleaving lines from non-adjacent parts of the page.
That reading-order failure makes the raw error number understate how much repair the output needs, because fixing scrambled sequence is slower than fixing isolated word errors. It is a known pattern for engines built around structured forms when they meet free-flowing handwritten prose.
On printed documents inside Google Cloud it is a perfectly reasonable choice. On handwriting it is hard to recommend over the other two cloud options, let alone a specialist.