Sanoma launches Sanna teacher assistant after a 1,500-teacher pilot

A classroom product begins with curriculum context

Sanoma Learning has launched Sanna, an AI teacher assistant built on Amazon Bedrock and grounded in the publisher’s own learning materials. The company supports 25 million students across 11 European countries and argues that generic assistants often produce an answer that is technically correct but uses a method or difficulty level unsuitable for the class. Curriculum alignment is therefore a core requirement rather than an optional retrieval feature.

The product launched in selected markets including Finland, Sweden, the Netherlands, Belgium, Poland, Italy and Spain. Sanoma can choose among models through Bedrock as the product evolves without rebuilding the entire application. That flexibility is useful, but every model change still needs local validation because a new model can alter pedagogy, language quality and safety behaviour even when the interface remains unchanged.

Teachers shaped the development cycle

The project began with an AWS-facilitated workshop in Amsterdam in March 2025, bringing together staff from four countries, pedagogical experts and the AWS Generative AI Innovation Center. Sanoma then moved through prototypes, trials and repeated teacher research. Early testing involved dozens of teachers, followed by wider work in late 2025 and a scaled classroom pilot with 1,500 teachers across seven countries in April 2026.

Researchers met teachers as often as every two weeks, using interviews, surveys and longitudinal diaries. That process exposed differences between markets and users: some teachers preferred open prompts while others wanted guided workflows. Treating those preferences as design evidence is more credible than assuming one conversational interface fits every subject, grade and education system.

Grounding supports trust but does not prove correctness

Sanoma says Sanna draws on textbooks and workbooks that teachers already use, backed by hundreds of pedagogical experts. Grounding can reduce unsupported answers and keep material aligned with a national curriculum. Visible sources also help a teacher check why the assistant proposed an exercise or explanation rather than accepting it because the prose sounds authoritative.

The content base still requires versioning and provenance. A curriculum changes, a textbook receives a correction and schools may use different editions. Each generated artefact should identify the material and version it relied on. Teachers need a quick path to report a poor explanation, and confirmed issues should become regression tests across languages and subjects.

The teacher keeps the consequential decision

Sanoma’s stated model is that AI suggests, organises and assists while the teacher decides. A planned grading and feedback module would scan paper tests, group similar answers and propose scores against criteria set by the teacher. The teacher would calibrate the results and make the final grading decision. That division is especially important because assessment affects a student’s record and opportunities.

Human oversight must be practical rather than ceremonial. If reviewing a suggestion takes longer than grading directly, teachers may either ignore the system or approve recommendations too quickly. Interfaces should highlight uncertainty and unusual answers, preserve the original student work and make correction easy. Schools should monitor whether suggestions vary unfairly across language, disability or response style.

Education data demands durable safeguards

Sanoma points to its experience with the General Data Protection Regulation and the EU AI Act. A classroom assistant can encounter student names, work, learning needs and behavioural context, much of it involving children. Data minimisation, regional storage, age-appropriate notice and clear retention limits should be built into the product and contracts, not left to individual teachers.

Sanna’s development offers a useful pattern: begin with an educator-defined problem, ground outputs in approved materials, pilot in real classrooms and keep the professional responsible for the decision. Reported time savings are encouraging, but long-term evidence should include learning quality, teacher workload and equity. The product succeeds if it gives teachers more time and better tools without weakening their agency or turning students into unexamined training data.

What to watch next

Schools should involve students and families as the deployment grows. Their concerns may differ from those of teachers or publishers, especially around recording, profiling and the use of work for improvement. Feedback channels need to be accessible and should not require a child to challenge a teacher directly. Publishing age-appropriate explanations of what the assistant does, what it cannot do and how data is handled would make the human-centred design claim visible to the people most affected.

Implementation detail

Teacher professional development will determine whether the product is used thoughtfully. Training should cover when to rely on approved source material, how to verify an output, how to identify bias and how to avoid entering unnecessary student information. Schools should give staff enough time to learn rather than measuring adoption by prompt volume. A smaller number of well-reviewed uses can create more educational value than constant generation. Support teams also need a clear escalation route when a teacher finds unsafe, inaccurate or culturally inappropriate material.