How to choose AI marking software for your school or trust

What AI marking software does, and what it does not

AI marking software reads scanned or uploaded student answers, compares them with a mark scheme, and proposes marks and feedback for a teacher to review. For a school mock, the teacher decides the final mark.

The software can suggest a mark and a reason. It does not take away the person who is accountable for the result a student, a parent or a head of department will later query.

In regulated qualifications the boundary is set by the regulator, not by a supplier. Ofqual's approach to regulating the use of artificial intelligence in the qualifications sector says Ofqual has made clear to awarding organisations that AI may not be used as the sole marker in any assessments that are part of regulated qualifications. That sentence is about awarding organisations and regulated assessments. It is not a review of school software, and it is not an endorsement.

Ofqual's working paper Principles of AI use in marking, published on 14 January 2026, looks at high-stakes marking. It says that using AI as the sole mechanism for determining a student's mark does not comply with Ofqual's regulations. Read that as a boundary on live, regulated exams.

Can schools use AI to mark GCSE mocks?

For a mock set inside the school, a teacher should review the marks before they are used. Live GCSE papers are marked by the awarding bodies.

Ofqual's sole-marker line is written for awarding organisations and for assessments that sit inside regulated qualifications. An internal mock is the school's own assessment. The working standard is that a teacher sees the suggested mark, can change it, and decides what is released to students.

JCQ's guidance on AI use in assessments answers a different question: how students may use AI in their own work, including non-examination assessment, and what counts as malpractice. It is not a kite mark for a marking product.

The Department for Education's generative AI product safety expectations are a useful list of questions to put to any supplier. Citing DfE, Ofqual, JCQ or an exam board is not the same as being endorsed by them. Excelas is not affiliated with any awarding body.

The 8-point checklist for choosing AI marking software

Judge any supplier on eight things: the mark scheme, teacher control, data protection, how marks leave the system, handwriting, turnaround, price, and evidence. Ask for a demonstration on your own scripts, not a slide.

1. 01

   Marking against your board's mark scheme

   The tool should mark the paper you set, against the mark scheme you use, including method marks.

   Ask. Does it use the actual paper and mark scheme, including method marks and any amended scheme? Can you upload your own? Which boards and subjects does it really cover?

   What good looks like. You can see one of your scripts marked against your scheme, not a generic rubric, and the supplier can name the boards and subjects it supports.

2. 02

   Teacher moderation and override

   A teacher should be able to see, change and reject any mark, and decide when a result is released.

   Ask. Can a teacher change or reject a mark? Can results be held back until someone has reviewed them? Is the original suggestion kept, so you can see what changed?

   What good looks like. You watch this on a real script. The teacher owns the final mark, and you know what the record keeps when they overrule the suggestion.

3. 03

   Data protection and the DPIA

   You stay responsible for your school's DPIA, so ask where pupil data goes before you sign.

   Ask. Who is the controller and who is the processor? Where is data processed and stored? Which sub-processors are used? Is pupil work used to train models? How long is data kept, and what happens when the contract ends? Will they give you what you need to complete your own DPIA? Also ask whether the tool flags a safeguarding concern in what a student has written, and what a member of staff is then expected to do.

   What good looks like. Plain answers, in the contract or the privacy notice, that your data protection lead can check. A logo on a slide is not that check.

4. 04

   MIS and data export

   Marks should leave the tool in a form your department already uses.

   Ask. Does it connect to your MIS, or take a CSV of class lists? What can you export: total marks, question-level data, a report a student can read?

   What good looks like. You see an export and a student report from a real batch, and you know which MIS systems are actually covered.

5. 05

   Handwriting and scan handling

   It should read the scripts your school already produces, on the scanner you already have.

   Ask. Does it read handwriting from your own scanner or multifunction device? What happens with crossings-out, diagrams and pages out of order? How is a page matched to the right student, and how is an uncertain page flagged?

   What good looks like. You try it on messy scripts from your own mock. The supplier can show a page it was not sure about, rather than claiming it never fails.

6. 06

   Turnaround

   Ask how long a full year group takes, what is promised in writing, and what happens in mock week.

   Ask. How long from a complete upload to marks a teacher can review? What is written into the contract? What happens when several departments upload in the same week?

   What good looks like. One written promise, with the clock defined from a clear starting point. A slogan is not a promise. Compare the wording on the supplier's own pages and keep the one they will stand behind.

7. 07

   Pricing transparency

   A published price, and a clear list of what is extra, is easier to take to a budget holder than a quote you cannot show.

   Ask. Is the price published? What does one licence cover: a subject, a school, a volume of scripts? Which add-ons cost more, and is VAT included?

   What good looks like. You can put the cost on a budget line without a sales call, and you can see the add-ons before you sign.

8. 08

   Evidence

   Ask what is claimed, how it was measured, against what, on which papers, and whether anyone independent has checked it.

   Ask. What were the marks compared with: one teacher, a senior examiner, or the moderated mark? Which papers and question types? Where did the tool disagree? Has anyone outside the company reviewed it?

   What good looks like. A method you can read, and a pilot on your own scripts. A single accuracy percentage, with no method, is not evidence.

What evidence should a supplier show you?

Ask what the marks were compared with, on which papers and question types, and where the tool disagreed. Agreement with a human mark is not, by itself, enough to show that a tool is valid.

Ofqual's Principles of AI use in marking is about high-stakes marking of regulated qualifications. In that setting it says that research and industry guidance both emphasise that agreement with human marks alone is insufficient for assuring validity. A supplier should still be able to say whether the comparison was one teacher, a senior examiner or a moderated mark, and which question types were in the set.

The Department for Education has published findings from the EdTech Evidence Board pilot. We have written up what that pilot means when you are buying, in Before you buy AI marking. It is not a ranking of products.

The strongest evidence you will get is a pilot on scripts your own department has already marked. A sample marked report shows the shape of the output. It does not replace that pilot.

Where does the teacher sit in the process?

A good tool lets a teacher see the suggested mark and the reasoning, change any mark, and decide when results are released to students.

Ask to watch that step on one of your own scripts. Ask who owns the final mark, and what is stored when a teacher changes the suggestion. If the supplier cannot show the screen, treat that as a gap.

There is a longer account of what that review should look like in keeping humans in the loop. Buying for several schools is covered in AI marking for multi-academy trusts.

Data protection: what to ask before you sign

Ask who is the controller and who is the processor, where pupil data is processed and stored, which sub-processors are used, whether pupil work is used to train models, how long data is kept, and whether you are given what you need for your own DPIA. You remain responsible for that DPIA.

Put the same questions to every supplier, including Excelas. The documents to read alongside the contract are the Department for Education's data protection guidance for generative AI in schools, its product safety expectations, and the ICO's guidance on AI and data protection.

There is no separate Excelas DPIA page to download. The privacy policy is the published notice, and a DPIA information pack is available on request from info@excelas.ai. The school still completes its own DPIA. A wider set of questions for a school policy is in an AI policy for schools, and the practical worries heads raise are covered in is it safe to use AI to mark student exams?

Ask, as well, whether the product flags a safeguarding concern in a script and what happens next. That is a fair question. This guide does not claim that Excelas does it.

How to run an AI marking pilot

Mark a sample of your own already-marked scripts, compare the results with your marks, record the time, and ask the teachers what changed. Decide what would count as good enough before you start.

Keep the sample small and honest: about 20 to 30 scripts, more than one question type, including messy handwriting and extended answers. Include pages that are crossed out or out of order. A tidy demo set will not tell you how the tool behaves in mock week.

This page describes a pilot you run on work you already have. Our AI mock exam marking service is where the product itself is described, and a sample marked report shows the output before you upload anything of your own.

1. Write the pass criteria first: what difference from your own marks you will accept, and what would make you stop.
2. Mark the sample and compare question by question, not only the total.
3. Write down how long the upload, the review and the corrections took.
4. Ask the teachers who reviewed the scripts what they still had to do themselves.

Red flags

Be wary of a supplier that will not show how marks were checked, offers to remove the teacher from the process, cannot say where pupil data goes, or quotes an accuracy percentage with no method behind it.

• No account of what the marks were compared with, or on which papers.
• A promise that teachers need not review the marks.
• No answer on where pupil data is processed, or who else handles it.
• An accuracy percentage with no method, no paper and no date.
• A claim that DfE, Ofqual, JCQ or an exam board has endorsed the product.

How Excelas answers this checklist

These are Excelas's answers where we can point to a published page. They are not a score, and they do not mean every box is ticked. Where a point is missing, ask us on a demo. The product is described on our AI mock exam marking service, and the home page is the short account of the platform.

Frequently asked questions