AI ASSESSMENT · HOW IT WORKS
From Rubric to Result: How Automated Assessment Works Step by Step
Rubric upload, multi-agent evaluation, reconciliation, lecturer review, and grade passback. The whole process, made visible.
By Eduface · July 2026 · 8 min read
When a lecturer asks how this actually works, a vague answer is usually the end of the conversation. Learning technologists know the moment well: scepticism about AI assessment rarely comes from a principled objection to the technology. It comes from not knowing what is happening inside the system. Once the process is visible, rubric upload through to grade passback, the question changes. It becomes: why isn’t my institution doing this already?
How does automated AI assessment actually work?
An automated assessment tool applies a lecturer-defined rubric to student submissions using multiple AI agents working in parallel. Each agent produces an independent assessment. A fourth agent reconciles the outputs into a single draft grade and feedback report. The lecturer reviews and approves the draft before anything reaches the student. Grade passback to the LMS happens automatically through the LTI 1.3 integration.
What happens when a student submits work?
From the student’s perspective, nothing changes. Eduface connects to Canvas, Brightspace, Moodle, or Blackboard through the LTI 1.3 standard. Students submit their work through the same portal they have always used. There is no separate login, no new interface, and no additional step in the submission process.
The LTI 1.3 integration handles authentication automatically. When a student submits, the work is routed to Eduface’s platform for processing. Students see no indication that anything different has happened. For them, it is a normal submission. This matters for adoption. One of the most common concerns learning technologists raise before implementation is whether students will need to change their habits. They do not.
How do AI agents evaluate a submission against a rubric?
Here is where the process diverges from what most people assume. Eduface does not use a single AI model to mark each submission. It uses three independent AI agents, each evaluating the same submission separately against the uploaded rubric.
Each agent reads the full submission alongside the rubric criteria. It assigns a score per criterion, produces written justification, and generates feedback comments. Crucially, the three agents do not see each other’s outputs at this stage. The evaluation is genuinely independent.
After the three agents complete their assessments, a fourth agent (the reconciliation agent) reads all three outputs. Where the agents agree, consolidation is straightforward. Where they diverge, the reconciliation agent works through the disagreement systematically and produces a single draft grade with written feedback and per-criterion scores.
This multi-agent architecture is not a cosmetic feature. It directly addresses one of the central objections to AI marking: the risk that a single model produces a single, unchecked output. Three independent assessments followed by structured reconciliation is, by design, closer to how moderation works in human marking.
Where the processing happens
All processing runs on Eduface’s own GPU infrastructure in the Netherlands. No third-party AI APIs are involved. Submissions are not sent to OpenAI, Google, Anthropic, or any external provider. Student data does not leave Eduface’s infrastructure.
What is the lecturer’s role in the workflow?
The lecturer is the decision-maker at every stage. Before any student submits, the lecturer configures the assessment in Eduface. The rubric (criteria, weightings, and grade descriptors) is uploaded or built in the platform. The AI works from this rubric. It does not generate its own criteria or apply generic standards. Everything it evaluates is grounded in what the lecturer has defined.
After the reconciliation stage, the draft grade and feedback are presented in Eduface’s review interface. Two modes are available. In blind mode, the AI draft is hidden until the lecturer has submitted their own independent grade; both are then revealed side by side for comparison. In AI-visible mode, the draft is shown upfront for the lecturer to review, edit, and approve.
In both modes, the grade remains a draft until the lecturer explicitly approves it. Nothing is released to students until that confirmation is given. The human decision is not advisory. It is the gate.
This structure satisfies the requirements of the EU AI Act (Regulation 2024/1689), which classifies AI systems used in student assessment as high-risk under Annex III and requires meaningful human oversight under Article 14. The design is compliant by construction, not by later addition.
How does grade passback work with the LMS?
Once the lecturer approves the grade, the LTI 1.3 connection handles the rest. The grade and feedback are sent directly to the LMS gradebook without any manual entry. The lecturer does not need to copy scores across systems, export a CSV, or log in to a separate admin panel.
For learning technologists managing assessment at scale, this is operationally significant. Grade passback errors and manual data-entry mistakes are a known source of disputes and appeals. Automating the connection removes that risk. The LMS gradebook reflects exactly what the lecturer approved, transmitted by the integration. The same LTI 1.3 connection that handles student submission also handles grade passback. There is one integration to set up and maintain, not two.
What happens to student data?
Eduface processes all submissions on its own infrastructure, hosted in the Netherlands. Data does not pass through external AI APIs. This is a deliberate architectural decision, not a policy position added after the fact.
For institutions with GDPR obligations (which includes every UK and EU university), the location and processing of student data is a procurement requirement, not a preference. Eduface is Jisc/CHEST approved and EU AI Act compliant. Data processing agreements are available at institutional level. Student submissions are used to produce the assessment output. They are not used to train external models or shared with third parties. The boundary around student data is the boundary of Eduface’s own infrastructure.
The automated assessment workflow, step by step
Step
Who acts
What happens
1. Rubric setup
Lecturer
Uploads or configures criteria, weightings, and grade descriptors in Eduface
2. Student submission
Student
Submits via LMS as normal. No new portal or separate login required
3. Multi-agent evaluation
Eduface AI (x3)
Three agents evaluate the submission independently against the rubric
4. Reconciliation
Eduface AI (x1)
Fourth agent consolidates three outputs into a single draft grade and feedback report
5. Lecturer review
Lecturer
Reviews draft in blind or AI-visible mode, edits if needed, approves
6. Grade release
System
Grade and feedback sent automatically to LMS gradebook via LTI 1.3
Frequently asked questions
Does automated assessment require students to use a new submission system?
No. Eduface connects to the institution’s LMS through LTI 1.3. Students submit through the same system they already use: Canvas, Brightspace, Moodle, or Blackboard. There is no new portal, no separate login, and no change to the submission experience. The integration is invisible to students. This is one of the most common adoption concerns raised during procurement conversations, and it is a non-issue in practice.
How many AI agents does Eduface use to mark each submission?
Four in total. Three agents evaluate each submission independently against the rubric, producing separate scores, justifications, and feedback. They do not see each other’s outputs at this stage. A fourth reconciliation agent then reads all three assessments and produces a single consolidated draft grade and feedback report. Where the three agents agree, the reconciliation is straightforward. Where they diverge, the fourth agent resolves the difference systematically before presenting the draft to the lecturer.
Can a lecturer override or edit the AI’s draft grade?
Yes, in full. The AI output is always a draft. In Eduface’s review interface, lecturers can edit individual criterion scores, amend the written feedback, and adjust the overall grade before approval. Nothing is released to students until the lecturer explicitly confirms it. In UK pilots, Eduface has achieved 95% alignment with lecturer assessments, meaning edits are typically minor. But the ability to override is always present and always final.
How does Eduface integrate with Canvas, Moodle, Brightspace, and Blackboard?
Eduface uses the LTI 1.3 standard, which is supported by all four platforms. LTI 1.3 handles authentication, submission routing, and grade passback within a single integration. From an institutional IT perspective, this follows the same integration model used by other learning tools in the VLE. There is no custom connector to build for each platform. If the LMS supports LTI 1.3, Eduface connects to it.
Is student data processed outside the EU when using Eduface?
No. Eduface runs on its own GPU infrastructure in the Netherlands. Submissions are not sent to external AI providers such as OpenAI, Google, or Anthropic. Student data remains within Eduface’s own infrastructure throughout the entire assessment process. This supports GDPR compliance for both UK and EU institutions and is a requirement built into Eduface’s architecture, not an optional configuration.
Conclusion
The process is not a black box. Rubric upload, submission via LMS, three independent AI evaluations, structured reconciliation, lecturer review in blind or AI-visible mode, grade passback through LTI 1.3. Each step is visible, each decision point is documented, and the lecturer confirms every grade before it reaches a student. For institutions still weighing whether AI assessment is ready for deployment, the more useful question may be whether they can afford the inconsistency, delays, and workload that come from not deploying it.
Pilot partners: Bath Spa University, De Haagse Hogeschool, Tilburg University, Hogeschool Rotterdam, UMCG.
See the whole process on your own rubric
Book a demo or create a free lecturer account, and run one assignment end to end.