AI essay feedback that scores like the teacher, quotes every sentence it rests on, and reaches students only when the teacher releases it
Students get feedback in days, not weeks, with every comment signed by their teacher.
See one case, screen by screen ↓AI Essay Feedback is AI essay feedback for teachers: it drafts rubric-based feedback on student essays for the teacher to edit and release. Six agents read the class's essays, check first that their scoring agrees with the teacher's own grading, draft one comment per rubric criterion in the teacher's voice with the sentence and rubric line it rests on, and check every draft for fairness. Nothing reaches a student until the teacher releases it.
AI in education: a buyer's guide to agentic AI for colleges, universities and school districts →
Updated
AI essay feedback is only worth using if it scores like the teacher and the teacher signs every comment. An English teacher with five sections and 152 students who spends 10 minutes on each essay needs about 25 hours to return one assignment, so feedback on a Monday essay arrives a week or two later, after the class has moved to the next unit. Teachers using the same rubric score Evidence differently, and a tool that has never been calibrated to the teacher adds a second inconsistency.
General AI writing feedback tools draft quickly, but they over-correct English learners' grammar, can treat dialect features as errors of argument, and turn writing-process signals into verdicts about who wrote the essay. Essay feedback comments typed one teacher at a time leave no district record of what was drafted, changed and released, and nothing to show for a grade appeal or a records request. AI Essay Feedback puts six agents in front of the reading and drafting, and leaves every decision with the teacher.
Estimated hours for one argument essay across five sections, by hand and with the agents: 152 essays at about 10 minutes each is about 25 hours today. The 'after' column is our arithmetic, kept modest: the teacher still reads every essay with open items and a spot check of the rest.
Six agents read, calibrate, draft, check fairness and ask; the teacher keeps, edits or drops every comment and releases.
Marlow Bay High School, Tuesday, October 13. Desmond Akers teaches English 9, and Period 3's 34 students wrote an argument essay, "Should Marlow Bay Unified move to a four-day school week?", from three sources. Essays were due Monday at 8:00 a.m. Here is how AI essay feedback reaches them, screen by screen, in the working solution.
By 5:46 p.m. Monday the Feedback agent had drafted 155 comments on 31 of Period 3's 34 essays. Three wait: Marcus Bell's arrived late, Priya Raman had extended time, and page 2 of Mateus Oliveira's handwritten essay was too faint to read. The Calibration agent agrees with Mr. Akers on only 24 of 30 sample scores, so the briefing's next step is Calibrate first.
Fairness agent: changes on 3 essays. Writing process agent: one question on Owen Pratt's essay.
Linh Truong's essay, "Four Days Is Not Enough", sits in the center. The rubric rail shows Grade 9–10 Argument Writing v3.2 with a proposed level for each of five criteria, 12 of 20 in all, which reaches the gradebook only when he confirms it at release. He selects Evidence and the three sentences its comment rests on light up; every comment quotes its sentence, names its rubric line and offers Keep, Edit, Drop and Ask.
Period 3 argument essay: 31 of 34 drafted, 0 released
Linh arrived in February and is at English language development level 2; she outlined her essay in Vietnamese. The Fairness agent shows each of the five lines beside an English translation checked by Mai Pham, English learner specialist. Line 4 says parents work on Friday, so younger brothers and sisters stay home alone, and the Reasoning comment credits that thinking and asks her to bring it into the essay.
Drafting never sees student names: 31 of 31 drafts were written name-blind. On Linh's essay four grammar comments became one pattern; on Jaylen Brooks's, two comments that treated dialect features as reasoning problems were rewritten; André Baptiste's Haitian Creole margin notes were translated. With names swapped, no score changed, and every change shows before and after with Accept and Restore.
Mr. Akers had graded six sample essays on Monday. The Calibration agent scored them without seeing his scores and agreed exactly on 24 of 30 criterion scores, 80%, under the 85% gate that locks release. It rates Evidence one level higher than he does on 4 of 6, crediting cited sources even when nothing explains them, and suggests anchor wording from his own scores.
Suggested anchor, level 2: "Evidence is listed or quoted without explaining how it supports the claim, however many sources."
He applies the tightened anchor to his sections. The agents re-score Periods 3 and 5: 9 Evidence scores in Period 3 and 7 in Period 5 go down one level, and 13 Period 3 Evidence comments are rewritten to ask for explanation. Agreement rises to 28 of 30 and the gate is met; the two remaining differences, on Organization and Conventions, are his scores and they stand.
Agreement rose from 80% to 93%.
138 of Owen Pratt's 184 words arrived in one paste at 11:41 p.m. Sunday, and his essay cites a 31 percent figure that is not in Sources A–C. The agent marks it "Not a finding" and offers three questions to ask him; nothing here reaches the student or his family. Mr. Akers chooses to release the feedback and hold the score until they talk Wednesday at 7:45 a.m.
The release check lists six gates. Marcus Bell's late essay is accepted, Priya Raman's extended time is never counted as late, and Mateus Oliveira's page 2 has been rescanned; the agents draft all three and he keeps the five comments on each. Essays with open items are opened one by one, and a random spot check of five essays covers the rest of the batch.
He releases the class in one step and confirms the scores as the checker. The Release agent posts the comments to the learning management system, writes 33 scores to the gradebook with Owen's held, and sends 4 family summaries. Of 170 drafted comments, 121 were kept as drafted, 38 edited and 11 dropped with a reason, and Recall stays open for 10 minutes.
In the learning management system Linh sees feedback from Mr. Akers on "Four Days Is Not Enough", 12 of 20, each comment under the sentence it is about, with English and Vietnamese side by side. Her family receives a short summary in Vietnamese, checked by Mai Pham at 3:31 p.m. and sent with the release.
The same essay assignment, seen by the four people who carry it: what their day looked like, and what it looks like now.
Six agents read, calibrate, draft, check fairness and ask; the teacher keeps, edits or drops every comment and releases.
Pulls the assignment, rubric, sources and submissions from the learning management system, reads handwriting and notebook photos, and removes names, demographics and English learner status before drafting.
Scores essays the teacher already graded without seeing the scores, reports exact and within-one agreement and the lean per criterion, suggests anchor wording and re-scores after a change.
Drafts one comment per rubric criterion in the teacher's voice, learned from 120 of the teacher's released comments, quoting the sentence it is about and naming the rubric line; proposes a level with its confidence.
Checks drafts on English learners' essays and writing with dialect features, merges grammar notes into one pattern at English language development levels 1–2, translates first-language notes beside the English, and runs a name-swap test and a comment balance check.
Reads an essay's version history and checks its facts against the assignment's sources, then gives the teacher questions to ask the student.
Checks every gate, posts the teacher's kept and edited comments, writes confirmed scores to the gradebook, sends checked family summaries and keeps the receipts.
Releases every comment. The agents propose; a named person decides.
Ask in plain words, or press Ask on any comment. Every answer points to the sentence, the rubric line or the note it rests on, and opens the essay.
Why did Linh Truong get a 2 for Reasoning?
Linh answers the other side, "parents work on Friday, so who take care the young children?", but does not connect it back to her claim, which is level 2 on your rubric. Her Vietnamese notes make the link: younger children home alone on Fridays, so the comment credits that thinking and asks for one sentence. The Fairness agent kept the comments on her argument and put her grammar into one pattern.
Where do the agents disagree with my grading?
On your 6 sample essays we agree exactly on 24 of 30 scores (80%). The gap is Evidence: I scored it one level higher on 4 of 6, because I credited two sources even when nothing explained them, as in Ethan Cole's essay. Tightening the level 2 and 3 anchors closes it; your other two differences are Organization for Amara and Conventions for Kenji.
What still stops me releasing Period 3?
Three things. Calibration is 80%, under the 85% gate. Linh Truong's essay still needs you one by one. And 3 essays are not drafted: Marcus Bell's arrived late, Priya Raman had extended time, and Mateus Oliveira's page 2 was rescanned. Owen Pratt's feedback can go while his score waits.
Draft a note to Linh's family in Vietnamese
Done: a three-line summary in Vietnamese beside the English, with one strength and one next step from your comments. Her family's preferred language is Vietnamese. Mai Pham, English learner specialist, checked it at 3:31 p.m.; it goes out only with your release.
13 screens from the working solution, on its sample data. Pick one to see it large.
The thresholds every draft is held to are settings, set by the district and changed only with a second approver.
| Setting | Default | Choose from |
|---|---|---|
| Calibration gate | 85% exact · 95% within one level | Calibrate again for each new assignment type and each term |
| Leave the score blank below | 0.80 confidence | District-locked; the teacher sets the level |
| Feedback released within | 5 school days from the due date | Reminder to the teacher on day 3; the department chair told on day 6, with statistics only |
| Spot check before a batch release | 5 random essays per section | Essays with open items are always opened one by one |
| Reason codes | 8 codes for dropping a comment or changing a score | Not accurate, Not in my voice, I will say it in person, IEP or 504 accommodation, Other with a note, and more |
| Keep released feedback and scores | 5 years, then deleted with a log entry | Drafts never released kept 2 years |
| Essay access for anyone else | Stated need, principal's approval, 48 hours | Logged; names replaced by IDs in exports unless the privacy officer approves |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
Sources: Steiss et al., Learning and Instruction (2024): feedback from a general AI model on 200 grade 6–12 essays averaged 3.6 of 5, against 4.0 for trained expert teachers · RAND survey of teachers, 2023–24, reported by eSchool News (March 28, 2025): about a quarter of US ELA, math and science teachers used AI for planning or teaching · "demo" = seen in the working solution, on its sample data · "target" = design goal, measured in the live solution · "typical" = published figures · "estimated" = our arithmetic (152 essays at about 10 minutes each). Marlow Bay Unified School District, its schools and the people named on this page are fictional — characters and sample data in the working solution.
AI Essay Feedback is AI essay feedback for teachers: it drafts rubric-based feedback on student essays for the teacher to edit and release. Six agents read the class's essays, check first that their scoring agrees with the teacher's own grading, draft one comment per rubric criterion in the teacher's voice with the sentence and rubric line it rests on, and check every draft for fairness. Nothing reaches a student until the teacher releases it.
The Feedback agent proposes a level for each criterion with its confidence, and leaves the score blank below 0.80 for the teacher to set. A score reaches the gradebook only after the teacher confirms it at release, and a change to a posted score needs the assistant principal. It works as an AI grading assistant for teachers, not a grader.
In a published study of 200 grade 6–12 essays, feedback from a general AI model averaged 3.6 of 5 against 4.0 for trained expert teachers. That is why the Calibration agent first scores essays the teacher already graded, without seeing the scores, and shows where it leans. Release stays locked until it agrees with the teacher on 85% of criterion scores; in the working solution one anchor change took agreement from 80% to 93%.
Drafting never sees student names. The Fairness agent checks every draft on English learners' essays and writing with dialect features, merges grammar notes into one pattern at English language development levels 1–2, and shows first-language notes beside the English so comments credit the thinking. It runs a name-swap test and a comment balance check, and every change it makes shows before and after with Restore.
No. The Writing process agent reads version history and checks facts against the assignment's sources, then gives the teacher questions to ask. It never scores, never says who wrote the essay, and its questions are hidden from students and families.
Teachers see only their own students and English learner specialists only the students they support; a department chair sees statistics, never essays. Anyone else needs a stated need and the principal's approval, for 48 hours, logged. Released feedback and scores are kept 5 years and unreleased drafts 2 years, exports replace names with IDs, and student work is never used to train models.
Yes. The teacher keeps, edits or drops every comment, with a reason code when dropping, and releases the class. Essays with open items are opened one by one, a random spot check covers the rest, and a release can be scheduled or recalled within 10 minutes.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your rubrics, a set of teacher-graded essays and your district's rules for AI — and it goes live once every quality gate has passed. We will walk you through it on one of your own assignments first.
Plans drafted in minutes, and no student judged on sessions that never happened.
1-min demoAI Lesson PlanningEvery lesson drafted from the adopted curriculum, every activity cited to a standard.
1-min demoLiteracy & Dyslexia ScreeningEvery flagged reader's family told in its own language, inside the state deadline.
1-min demoWe'll run AI Essay Feedback on one of your rubrics and a set of teacher-graded essays.