Momentum Insights - Vol. 03 / Where AI Belongs — and Where It Doesn’t
Momentum InsightsVol. 03 / AI in Learning
A practical guide for learning leaders
Where AI Belongs — and Where It Doesn’t
Building long-term value with AI in learning: what to do today, what to expect in three to five years, and how your learning platform should grow alongside it.
For L&D leaders, educators, learning administrators, instructional designers and executives responsible for how their organisation learns.
MOMENTUM IT GROUP
Before you begin
The question isn’t whether to use AI. It’s where.
AI can give administrators back hours every week. It can also quietly do the thinking that learning depends on.
Every organisation that trains people is now being asked about AI: by boards, by vendors, by learners already using it. The pressure is to adopt fast. The risk is adopting in the wrong places — automating the struggle that builds skill, or handing judgment to a system that can’t be held accountable for it.
This guide draws a clear line. It sets out where AI creates durable value in learning, where it doesn’t belong, the lowest-hanging fruit you can pick today, and what we expect to be possible in three to five years. It closes with the governance practices that keep it safe, and how a learning platform should grow alongside AI rather than be replaced by it.
It continues the thread of this series: Beyond the Feature Checklist (Vol. 1) argued for choosing technology by long-term value; The Residual Gap (Vol. 2) argued for rewarding the people who create it. Volume 3 applies both ideas to AI.
Three horizons
Now — low-risk efficiencies you can switch on this year. Next — connected, assisted learning in one to three years. Later — adaptive, agent-connected learning in three to five.
01 The principle
Use AI to amplify human learning — never to replace the learning itself.
Learning happens through effort: retrieving, practising, getting it wrong and trying again. Researchers call it productive struggle. Anything that removes the struggle also removes the learning.
AutomateAI does it
Tagging, translation, scheduling, reminders
AssistAI drafts, human decides
Quiz drafts, marking suggestions, reports
AugmentAI coaches the learner
Hints not answers, practice, guided learning paths
Human onlyNo AI in the loop
Final grades, mentorship, wellbeing, certification
LOW STAKES · ROUTINEHIGH STAKES · HUMAN JUDGMENT
Rule 1
Give time back
Automate the predictable work around learning so people can spend more time on the learning itself.
Rule 2
Keep humans accountable
When a decision affects a person’s grade, credential, job or wellbeing, a named human makes it.
Rule 3
Protect the struggle
AI should prompt thinking, not supply the finished thought. Hints over answers; questions over summaries.
Aligned with international and Canadian frameworks
UNESCO — Guidance for Generative AI in Education and Research (2023) and the AI Competency Frameworks for Students and Teachers (2024)
Canada’s privacy regulators — federal, provincial and territorial Principles for Responsible, Trustworthy and Privacy-Protective Generative AI Technologies (2023)
OECD — AI Principles (2019, updated 2024), endorsed by Canada
Government of Canada — AI for All, the national AI strategy (2026), with a focus on trust, safety and AI literacy
Council of Europe — Framework Convention on AI and Human Rights (2024), signed by Canada in 2025
Ontario — Strengthening Cyber Security and Building Trust in the Public Sector Act (2024), requiring human oversight of AI in bodies including school boards
All call for a human-centred approach, with people kept accountable for decisions that affect learners.
02 Where AI belongs in learning
Four places AI earns its keep.
01 Learners
Personalised tutoring & adaptation
AI can track each learner’s errors and pace in real time, offering targeted practice and alternative explanations that one instructor can’t provide to a whole cohort at once.
BEST PRACTICEGround the tutor in verified content and design it to give hints, not answers.
02 Instructors
Workflow support
Brainstorming lesson plans, summarising source material, generating varied perspectives and scenarios, and drafting routine communications.
BEST PRACTICETreat every output as a first draft that a subject expert reviews.
03 Learners
A thought partner
Outlining ideas, clarifying difficult concepts, and self-quizzing before an assessment — used to prepare for thinking, not to replace it.
BEST PRACTICEMake expectations explicit: when AI is allowed, how to disclose it.
04 Administrators
Operational scale
Scheduling, enrolment, translation, reporting and spotting systemic learning gaps across programs, regions and cohorts.
BEST PRACTICEStart here. It’s the lowest-risk, highest-return place to begin.
2×
The evidence · When AI is designed to teach
In a 2023 randomised trial in Harvard’s introductory physics course (194 students), learners using an AI tutor built on teaching principles — scaffolded steps, verified solutions, timely feedback, self-pacing — achieved median learning gains more than double those of an in-class active-learning session, in less time. Kestin et al., Scientific Reports, 2025.
03 Where AI doesn’t belong
Four lines we don’t cross.
01
Replacing reading, writing and thinking
Drafting, deep reading and original problem-solving are the work of learning. AI that does them for the learner removes the very effort that builds lasting capability.
02
Bypassing foundational assessment
Using AI to produce the essay or solve the equation short-circuits the iterative process, leaving knowledge that looks complete but doesn’t hold up without the tool.
03
Over-reliance in early learning
For young learners, excessive screen and AI use can crowd out emotional, social and hands-on development. UNESCO recommends a minimum age of 13 for classroom use of generative AI tools.
04
Replacing human relationships
Empathy, mentorship and social-emotional guidance come from people. AI can connect a learner to the right person; it shouldn’t pretend to be one.
The evidence · When AI is designed to answer
+48%
Practice performance with an unrestricted chatbot
−17%
Exam performance for the same students once AI was removed
≈0
Exam harm when the tutor gave hints and teacher guidance instead
Field experiment with nearly 1,000 high-school maths students. Bastani et al., “Generative AI without guardrails can harm learning,” PNAS, 2025.
04 AI inside the LMS
The platform is where principle becomes practice.
An LMS sits at the centre of content, learners, records and reporting. That makes it the natural place to apply AI — and the place where boundaries matter most.
| Where AI fits | Where it doesn’t — or needs a human |
|---|---|
| Content tagging. Tag courses, videos and quizzes with skills, metadata and standards as they’re uploaded. | Final grading of nuanced work. Essays, discussions and creative projects need human interpretation for the final mark. |
| Smart content delivery. Recommend remedial or advanced modules based on assessment results. | Credential and certification decisions. Pass/fail on a licence or qualification stays with a qualified person. |
| Early-warning analytics. Flag learners at risk from logins, submissions and performance trends. | Automated consequences. A risk flag should start a human conversation, never trigger an automatic penalty. |
| Drafting assistance. Help authors format modules, write discussion prompts and seed question banks. | Unreviewed publishing. AI-drafted content goes live only after subject-expert review. |
| Translation & accessibility. Translate courses and forums; draft captions and alt text. | Emotion or attention monitoring. Inferring learners’ emotions is prohibited in EU education settings and erodes trust everywhere. |
| Marking assistance. Suggest scores and feedback on short and long answers against a rubric. | Ungoverned data access. AI should only reach approved sources, under the same permissions as the user. |
The working rule
AI drafts, suggests and flags. People decide, publish and grade. Every AI-assisted decision should be visible, explainable and reversible by the person accountable for it.
05 The AI fit test · A tool for any proposed use
Two questions decide how AI should be used.
Is the task part of the learning, or work around it? And how high are the stakes for the person affected? Plot any proposed use and the right level of AI involvement follows.
HIGH STAKES · AROUND THE LEARNING
Assist, with review
AI drafts · A named person signs off- Marking suggestions on written answers
- Compliance and audit reports
- At-risk learner flags
- Published course content
HIGH STAKES · PART OF THE LEARNING
Keep it human
AI may inform · Never decides- Final grades on nuanced work
- Certification and licensing decisions
- Mentorship and wellbeing support
- Discipline and academic integrity rulings
LOW STAKES · AROUND THE LEARNING
Automate
AI does it · Humans spot-check- Content tagging and metadata
- Translation of courses and forums
- Enrolment, reminders, scheduling
- Report summaries and dashboards
LOW STAKES · PART OF THE LEARNING
Guide with guardrails
AI coaches · Hints, not answers- Practice quizzes and self-testing
- AI tutors grounded in course content
- Suggested resources and learning paths
- Discussion prompts for cohorts
How to use it
Run every new AI feature or request through these two questions before it’s switched on. If you can’t agree which quadrant it belongs in, treat it as the higher-stakes one.
06 Horizon 1 · Now
The lowest-hanging fruit: value you can unlock this year.
These uses sit mostly in the “automate” and “assist” quadrants. They’re low risk, quick to adopt, and return time immediately — the right place to build confidence before moving further.
| Use | What it gives back | Effort | Risk | Human role |
|---|---|---|---|---|
| Translation of courses & forums | Reach learners in their own language without separate builds | Low | Low | Spot-check key terms |
| Content tagging & search | Findable content; skills mapped automatically on upload | Low | Low | Approve the taxonomy |
| Quiz & question-bank drafts | Hours saved per course; more practice items | Low | Low | SME review before use |
| Marking assistance | Faster feedback on short and long answers | Medium | Medium | Instructor confirms every mark |
| Accessibility drafts | Captions, transcripts and alt text at scale | Low | Low | Review for accuracy |
| Reminders & nudges | Better completion with less admin chasing | Low | Low | Set tone and frequency |
| Report summaries | Plain-language insight from dashboards for leaders | Low | Low | Validate before sharing |
| Survey & evaluation drafts | Better-designed evaluations tied to outcomes | Low | Low | Align to your measures |
← Swipe to see all columns →
Start small
Pick two uses, one team, one quarter. Measure time saved before and after.
Stay inside the platform
Use AI built into governed systems, not copy-pasting learner data into public chatbots.
Tell people
Let learners and staff know where AI is used and how to raise concerns.
07 Horizons 2 & 3 · Next and later
What we expect to become possible in three to five years.
Forecasts in AI age quickly, so we focus on direction rather than dates. Based on what’s already working in pilots today, these are the capabilities we expect organisations to be using routinely.
- Now · 0–12 monthsAssisted
Work faster
Translation, tagging, drafting, marking assistance, summaries. AI as a capable assistant inside existing workflows.
- Next · 1–3 yearsConnected
Work together
AI that understands your content and data, grounded in approved sources. Guided learning paths, cohort prompts, SME matching and richer analytics.
- Later · 3–5 yearsAdaptive
Work ahead
Agents that coordinate across systems, anticipate needs and keep learning aligned to the work, under human oversight.
Six capabilities we expect within five years
Agent-to-agent integration. The LMS, HR, CRM and credentialing systems exchange context through open protocols, ending re-keying and data silos.
Lifecycle automation. From first onboarding to lifelong continuing education, enrolments and renewals follow role and regulation automatically.
Grounded AI mentors. Learners get guidance drawn only from your organisation’s approved knowledge, with a path to a human expert.
Skills inference. Capability profiles built from assessments and activity, driving personalised paths and workforce planning.
Continuous assessment. Short, frequent checks and workplace “chalk talks” that keep critical knowledge current, not annual tests.
Measured transfer. Evidence linking learning to on-the-job performance, rather than completions alone.
08 Guardrails & governance
Trust is the real infrastructure.
AI in learning touches personal data, assessment and credentials. The organisations that get long-term value treat governance as a foundation, not an afterthought. Four reference points are worth knowing.
Framework
NIST AI Risk Management Framework
A voluntary framework organised around four functions: Govern, Map, Measure and Manage AI risk.
Standard
ISO/IEC 42001
The international standard for an AI management system — policies, roles, risk assessment and continual improvement.
Regulation
EU AI Act
Classes AI used for admission, assessment and exam proctoring in education as high-risk (obligations apply from 2 December 2027) and already bans emotion recognition in education.
Guidance
UNESCO & national guidance
Human-centred use, age-appropriate access, data protection and teacher capacity-building. Privacy law (e.g. PIPEDA, GDPR) applies throughout.
Eight guardrails every AI-enabled learning platform should have
09 Your AI-in-learning policy, in ten parts · A starter checklist
A short policy beats a long silence.
Most learners and staff are already using AI. A clear, plain-language policy turns guesswork into good practice. Use these ten parts as a starting outline.
General guidance only. Align your policy with the laws, regulators and collective agreements that apply to your organisation.
10 Measuring long-term value
Measure what AI changes, not how much AI you use.
Usage counts tell you AI is switched on. They don’t tell you it’s creating value. Track outcomes at each level of the organisation — the same five dimensions from Volume 1.
| Level | What to measure | Example indicator |
|---|---|---|
| Learners | Engagement, mastery and time to competency | Assessment gains; days to proficiency |
| Instructors | Time returned and feedback speed | Hours saved per course; marking turnaround |
| Administrators | Manual work removed | Tickets and manual enrolments per month |
| Leaders | Decision quality and early warning | At-risk learners reached before failure |
| Organisation | Transfer to performance and compliance | On-the-job behaviour change; audit readiness |
| Trust | Accuracy, fairness and confidence | Override rate on AI suggestions; learner sentiment |
1 · Reaction
2 · Learning
3 · Behaviour
4 · Results
11 Momentum’s AI roadmap
Administrators first. Learners next. Humans always.
Our priority is the work that surrounds learning: giving time back to the people who run programs, then using that foundation to guide learners — always under the guardrails set out in this guide.
Where we are today
Unbound · LMS
AI translation, built in
Full integration with the DeepL API translates course content and discussion forums, so learners can participate in their own language.
ExamDeck · Assessment
AI-assisted marking
Full integration with OpenAI suggests marks and feedback on short- and long-answer questions, with the instructor reviewing and confirming results.
Where we’re going
Priority 1 · Administrators
- 1
Lifecycle efficiency
Manage learners from first onboarding through lifelong continuing education with far less manual effort.
- 2
No more data silos
Agent-to-agent API connections between Unbound and the systems you already use.
- 3
Guardrails by design
Control what information AI can access, from which sources, with continued data security.
Priority 2 · Learners
- 4
Guided learning
Suggested resources and learning paths; AI mentorship; connecting learners with an internal SME.
- 5
Engagement that lasts
Drafted cohort discussion questions, quizzes and workplace “chalk talks” that keep key topics top of mind.
- 6
Evidence of impact
Meaningful reports and survey questions tied to recognised measures of performance and knowledge transfer.
Roadmap items describe direction, not committed delivery dates, and may change as technology and customer needs evolve.
12 Growing alongside AI
Your LMS shouldn’t be replaced by AI. It should grow with it.
AI will change faster than any organisation can re-platform. The durable advantage is a learning system that holds your content, people, records and guardrails in one place — and adds AI capability as you’re ready for it.
1 · FoundationClean data, one platform
2 · AssistedQuick wins switched on with human review
3 · ConnectedSystems talk to each other; AI grounded in your own knowledge
4 · AdaptivePersonalised paths, predictive support, agents under oversight
| Stage | What the organisation needs | How Momentum’s LMS supports it |
|---|---|---|
| Foundation | Content, learners and records in one governed place; an AI policy | Unbound as the single system of record, with roles and permissions |
| Assisted | Low-risk efficiencies that save time this year | Built-in translation and assisted marking; drafting tools as they arrive |
| Connected | Integration with HR, CRM and credentialing; trusted data | API and agent-to-agent connections with source and access guardrails |
| Adaptive | Personalised, measured learning across the employee lifecycle | Suggested paths, AI mentorship, SME matching and impact reporting |
Long-term value
You don’t have to choose between today’s platform and tomorrow’s AI. Start at your stage, prove value, then climb — with a partner who stays accountable for the result, as we argued in Volume 1.
The best use of AI in learning is the one that makes more room for people.
- Let AI handle the routine.
- Let it coach, not answer.
- Keep judgment with people.
- Grow one stage at a time.
AI will keep changing.Your learning should keep getting better.
Next step
Talk to us about your AI starting point
We’ll walk through the fit test with your team, identify two quick wins, and show you where Unbound and ExamDeck fit today.
momentumitgroup.comSources & notes
Sources
- Kestin, G., Miller, K., et al. “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting.” Scientific Reports (2025). 194 students, Harvard introductory physics, Fall 2023.
- Bastani, H., et al. “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences 122 (2025). Nearly 1,000 students; GPT Base +48% practice, −17% exam; GPT Tutor no significant exam effect.
- UNESCO. Guidance for Generative AI in Education and Research (September 2023). Recommends a human-centred approach and a minimum age of 13 for classroom use of AI tools.
- UNESCO. AI Competency Framework for Students and AI Competency Framework for Teachers (September 2024).
- Office of the Privacy Commissioner of Canada and provincial/territorial privacy authorities. Principles for Responsible, Trustworthy and Privacy-Protective Generative AI Technologies (7 December 2023).
- OECD. Recommendation of the Council on Artificial Intelligence (OECD AI Principles), adopted 2019, updated May 2024.
- Council of Europe. Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225, 2024). Signed by Canada, February 2025.
- Government of Canada. AI for All: Canada’s National Artificial Intelligence Strategy (June 2026).
- Ontario. Strengthening Cyber Security and Building Trust in the Public Sector Act, 2024, S.O. 2024, c. 24 (AI provisions apply to prescribed public-sector entities as regulations are made).
- National Institute of Standards and Technology. AI Risk Management Framework (AI RMF 1.0) (January 2023). Functions: Govern, Map, Measure, Manage.
- ISO/IEC 42001:2023, Information technology — Artificial intelligence — Management system.
- Regulation (EU) 2024/1689 (AI Act), Article 5(1)(f) and Annex III; as amended by Regulation (EU) 2026/1744 (Digital Omnibus on AI), published 24 July 2026, deferring Annex III high-risk obligations to 2 December 2027.
- Kirkpatrick, D. L. & Kirkpatrick, J. D. Evaluating Training Programs: The Four Levels (3rd ed., 2006).
Notes
This guide is general information, not legal or compliance advice. Regulation of AI is changing quickly; check current requirements in your jurisdiction. Momentum roadmap items describe direction and are not commitments to specific features or dates. Third-party product names (DeepL, OpenAI) are trademarks of their respective owners.
Momentum Insights · Vol. 03
DOWNLOAD AS PDF
Where AI Belongs — and Where It Doesn’t
© 2026 MOMENTUM IT GROUP
