AI grading for teachers works best as drafting support: it compares student work to your rubric and suggests comments, while you decide the grade. It is weakest on creativity, context and unusual answers. Review every comment, sample-check the scoring against your own marking, and never let a tool assign final grades unseen.
Marking has two parts. One is applying criteria to work, and the other is judging quality in context. AI can help with the first and is unreliable at the second.
Where it helps: finding whether a response includes required elements, drafting rubric-based comments, grouping common errors across a class, and checking short answers with clear right answers. Where it struggles: original arguments, creative work, answers that are right in an unexpected way, and anything where you know context the model does not.
The safest pattern is the same every time.
Bias is the first. Models can score writing styles differently, and students who write in a second language may be marked down for expression instead of content. Check results across your own groups.
Privacy is the second. Student work often includes names and personal stories. Do not enter it into tools your school has not approved. Follow student privacy rules that apply to you, such as FERPA in the United States, and ask your administrator when unsure.
Over-trust is the third. Fluent feedback sounds authoritative even when wrong. Read it as you would a new assistant's first draft.
The saving depends on the task. For short factual quizzes and common-error feedback it can be large. For long essays, reading, checking and editing may take most of the time you would have spent anyway. Measure on one assignment before committing.
If you want a feedback workflow built into your school's process, our Teacher assistant drafts comments for teacher review using your rubric. It does not assign final grades.