AI and instructional design: what education practitioners are discovering on the ground

What does it actually look like when education practitioners in Kenya, Rwanda, Tanzania, India, and Afghanistan try to use artificial intelligence (AI) in their day-to-day instructional design work? What works, what doesn’t, and what do organisations need to move forward with confidence?

These were the questions at the heart of our first AI in Education Roundtable, a discovery session hosted by the Global Schools Forum (GSF) to bring practitioners together, surface real experiences, and begin building a shared understanding of the challenges and opportunities ahead.

The session was facilitated by Ekaterina Cooper and Habeeb Kolade and drew participants from non-state education organisations working across low- and middle-income countries. Through plenary discussion and two breakout groups, participants shared honest, grounded accounts of their AI journeys, from early wins to persistent barriers. This article captures the full range of insights that emerged.

Setting the scene 

To anchor the conversation in evidence, GSF opened with data from the Organisation for Economic Co-operation and Development (OECD) showing that nearly 40% of teachers globally are already using AI for planning work. Differentiation was highlighted as the feature teachers find most valuable, reflecting AI’s potential to help educators tailor instruction to diverse learner needs.

The session also acknowledged a significant gap: while AI adoption is growing globally, knowledge about how it is being used in low- and middle-income countries remains limited. This is precisely the context in which most GSF members work, and a key reason the roundtable was convened. The goal was to understand what is happening on the ground, in the classrooms and organisations where GSF’s community operates. In breakout rooms, participants explored two questions: what is the current state of AI use in instructional design, and what barriers and capacity gaps are limiting broader adoption?

What organisations are doing: the full range of AI applications

The breadth of AI use cases shared across both groups was striking. Participants described applications spanning content creation, assessment, teacher support, language learning, and performance monitoring. Taken together, they point to a sector that is actively experimenting, even if adoption remains uneven.

• Lesson planning and curriculum design – Generating lesson plans and teacher guides is the most common entry point for AI in instructional design. Organisations across Kenya, Rwanda, and Tanzania reported using AI to draft lesson plans. In Rwanda, one organisation has gone further by embedding AI directly into its Learning Management System (LMS) to support grading, tutoring, and reporting. This integration has freed teachers to focus more time on direct student support.

• • In Tanzania, AI is being used to revamp and localise curriculum content, adapting materials to fit the cultural and linguistic context of local learners. Participants from Tanzania also noted that AI has helped identify patterns in learner and teacher performance, enabling teams to continuously refine their modules based on real data.
  • In Kenya, one organisation is piloting AI-generated teacher lesson guides delivered directly to teachers via WhatsApp on smartphones. Uptake has been limited by rural connectivity challenges, but the approach signals creative thinking about reaching teachers in low-resource settings without requiring consistent internet access.
  • Participants in Group 2 described curriculum specialists using AI as a “substitute mind” to discuss whether questions fit lesson plans and to explore alternative instructional approaches. AI is used for ideation and drafting, not for producing final content independently. Teams use it to accelerate thinking, then apply professional judgement to refine, verify, and contextualise outputs before they reach teachers or learners.

• Higher-order thinking questions and assessment design – AI tools are being used to design higher-order thinking (HOT) questions as curriculum specialists develop teacher guides. One approach described involved using AI to create variations of assessment questions while maintaining the same structure, layering in different levels of Bloom’s Taxonomy, from recall and calculation through to synthesis and concept integration. This approach allows teams to build richer assessment banks more efficiently than developing every question from scratch. For new programmes, AI is also being used to design assessments from the ground up, generating items and automating grading to reduce the administrative burden on teachers and programme teams.

• Content production and instructional materials – Beyond text-based content, participants described using AI tools integrated into platforms such as Adobe and Canva to produce voiceovers, illustrations, presentations, and other instructional materials. This expands the range of content formats that smaller teams can produce without specialist design or production skills.
• Teacher and Master Trainer support – One of the more promising use cases emerging from Group 2 involves AI’s potential to strengthen Master Trainer preparedness in cascade training models. Participants noted that AI can help trainers prepare session plans with clearer structure, generate contextual examples and case studies, and anticipate participant questions and responses. This has particular relevance in contexts where quality consistency across different levels of a cascade is a persistent challenge. Some organisations have also piloted dedicated AI tools with teachers for lesson planning, assessments, and classroom support. The potential for AI teacher coaching systems and automated instructional support tools was flagged as a significant future opportunity.
• Language learning and conversational AI – Participants described experimenting with AI to develop conversational activities for teaching English as a second language. AI is being explored to support language acquisition aligned to classroom curriculum topics, with some teams also looking at its use for formative assessment within language learning contexts.
• Monitoring learner performance and differentiated learning – AI-supported differentiated learning was highlighted as one of the most compelling future directions. Participants described the potential for AI to identify where individual students are struggling and adapt content or support accordingly. Predictive systems that flag at-risk students based on attendance or performance patterns were also discussed as an emerging application. In Rwanda, one organisation is already using AI to spot patterns in learner data, using these insights to refine modules on an ongoing basis.
• Agentic workflows – A small number of participants described experimenting with agentic AI workflows, where AI is configured to generate and review content with less direct human intervention at each step. These remain early-stage experiments, and participants were cautious about moving too far in this direction given concerns about pedagogical quality and accuracy. Human review of all final outputs remains standard practice across all the organisations present.

What is getting in the way: barriers, risks, and capability gaps

The barriers identified across both groups were strikingly consistent. They cut across infrastructure, finances, skills, ethics, language, and culture, and together they explain why AI adoption, despite genuine enthusiasm, remains fragile and uneven.

• Infrastructure, connectivity, and cost – This was the single most cited barrier. Unreliable internet access, limited device ownership, the high personal cost of mobile data, and frequent power outages all prevent teachers, particularly in rural areas, from consistently accessing AI tools. In Kenya, low uptake of a smartphone-based pilot was attributed directly to teachers lacking regular access to data. In India, solar power and computer lab infrastructure are being used as partial solutions, but these require sustained investment that many organisations cannot provide alone. Teachers often bear data costs personally, a significant disincentive in contexts where incomes are low. Power outages and network instability further disrupt any attempt at consistent use. Organisations are piloting low-data solutions like lesson guides delivered via WhatsApp to mitigate some of these challenges, but more purpose-built, low-bandwidth AI tools are needed.

  The cost of AI tools compounds these challenges. Premium AI subscriptions tend to produce higher-quality outputs but are unaffordable for most organisations in low- and middle-income countries. This pushes teams towards free tools that are less effective, less contextually relevant, and more prone to producing inaccurate or generic content. Some organisations have responded by developing their own AI-enabled platforms to reduce recurring costs, but this requires technical capacity that many do not have.

  Local data storage regulations add a further layer of cost and compliance complexity. In some countries, institutions are required to store data on local servers, which complicates or prevents the use of cloud-based AI solutions and significantly increases the cost of scaling. This has emerged as a particular bottleneck for public sector institutions. The digital divide is not simply a technical problem. It is an equity problem. Where AI tools are accessible, they tend to be more available to urban, better-resourced schools, widening existing inequalities rather than narrowing them.

• Teacher readiness and digital literacy gaps – Many teachers, especially those in low-fee private schools and rural settings, have minimal prior experience with technology. Before AI integration is even possible, foundational digital literacy training is needed. Some teachers would need to start with basic computer skills, including email and internet use, before they could critically evaluate AI-generated content. Even in organisations where staff are already using AI, structured skills in prompt design, output validation, and instructional alignment are largely absent. A common pattern emerged: staff use AI for drafting, but lack the frameworks to assess whether what it produces is accurate, pedagogically sound, culturally appropriate, or suitable for their specific learners. This creates real risk, particularly in contexts where teachers may pass AI-generated content to students without adequate review.
• Data privacy and ethical risks – Data protection concerns featured prominently in both groups. Existing AI systems may embed biases and reflect limited representation of African or South Asian contexts, producing outputs that are culturally generic or simply inaccurate. Participants raised concerns about teachers uncritically trusting AI outputs and passing unverified content to students, with the potential for misinformation to spread. One participant noted that UNICEF’s ethical AI framework is being introduced in some contexts to guide responsible use, but awareness of such frameworks remains limited across the sector. Organisations called for clearer guidance on data privacy, appropriate use, and the specific risks associated with AI tools in educational settings.
• Academic integrity and over-reliance – Concerns about student use of AI ran through both group discussions. Participants described situations where students were submitting AI-generated work without citation or authentic personal voice. Academic integrity and plagiarism have become real issues that organisations are actively grappling with. Beyond academic integrity, participants observed students becoming dependent on AI as a form of cognitive offloading, completing tasks without the underlying thinking and learning those tasks were designed to develop. Organisations expressed hesitancy about deploying student-facing AI tools without strong guardrails, supervision, and clear pedagogical intent. There was also a candid acknowledgement that AI does not always save time. In some cases, producing a high-quality, contextually appropriate output requires significant human effort to review, correct, and adapt. When that effort is not invested, the output may be worse than what the teacher would have produced independently.
• Localisation and language barriers – AI tools are largely trained on datasets from high-income countries, producing outputs that often do not reflect local curricula, languages, cultural norms, or pedagogical traditions. Participants from multilingual contexts described particular difficulties: AI struggles with languages such as Dari, Pashto, and Urdu-English classroom settings, and translation can be unreliable when tools do not recognise the specific terminology used in a country’s official curriculum materials. Adapting AI outputs to local contexts requires human intervention at multiple points. Participants emphasised that this is not a marginal issue; it goes to the heart of whether AI tools can be genuinely useful, or whether they produce content that looks plausible but is actually misaligned with what teachers need. Feeding AI systems with local data and developing region-specific models was proposed as a longer-term solution. Increased participation from educators in Africa and South Asia in AI training datasets is expected to improve contextual accuracy over time, but this will require deliberate effort from developers, policymakers, and educators, not passive inclusion.
• Pedagogical quality and AI limitations – Participants raised specific technical limitations that affect AI’s reliability in instructional design contexts. AI can produce hallucinations, logical errors, and weak pedagogical alignment, particularly when conversations become long or contexts become complex. Outputs are often too lengthy or too resource-intensive for the classrooms they are supposed to serve. AI tools are not yet considered pedagogically mature, and human review of every output remains essential.

Strategic adaptations: how organisations are responding

Despite the challenges, participants shared a range of practical approaches that are helping them integrate AI more effectively within their constraints.

• Keeping humans in the loop – Every organisation described maintaining human oversight as non-negotiable. Outputs are reviewed by curriculum specialists, managers, and teachers before being finalised. Teachers check whether lesson plans and activities are realistic and executable in their specific classroom contexts. This human-in-the-loop model is both a quality control mechanism and a safeguard against AI’s known limitations.
• Using AI for ideation, not final outputs – Rather than treating AI as a content generator, teams are using it as a thinking partner. Curriculum specialists described using AI to explore whether questions fit lesson plans, test alternative approaches, and generate starting points that are then substantially reworked. This positions AI as a tool that accelerates and enriches human thinking, rather than replacing it.
• Delivering AI content through low-data channels – Organisations in Kenya are experimenting with delivering AI-generated materials via WhatsApp, reducing the data burden on teachers and meeting them on platforms they already use regularly.
• Embedding AI within existing systems – Rather than introducing standalone AI tools, some organisations have embedded AI functionality within their existing LMS platforms. This reduces the friction of adoption, keeps data within familiar systems, and allows AI features such as tutoring, grading support, and performance monitoring to be integrated into workflows teachers already use.
• Building AI literacy incrementally – Organisations working in low-resource contexts described building AI capability in stages, starting with foundational digital literacy and introducing AI tools only once teachers have the basic skills to evaluate and adapt what the tools produce. This incremental approach reduces risk and builds confidence before expanding use.
• Introducing ethical frameworks – Some organisations are drawing on established ethical AI principles to structure how they introduce AI to educators, helping teams think through questions of data privacy, bias, and appropriate use before they encounter them in practice.

What would help: what practitioners are calling for

When participants discussed what would enable more confident AI adoption, several consistent themes emerged.

• Practical, role-specific training – Training on AI tools was the most frequently requested form of support, particularly training focused on AI for instructional design rather than general AI literacy. Participants called for practical guidance on prompt engineering, output validation, and instructional alignment, as well as accessible, jargon-free introductory resources for teams just getting started. For organisations in low-resource settings, simple guides on how to use AI at a very basic level were specifically requested.
• Frameworks for evaluating AI tools – Organisations need practical frameworks for assessing AI tools before adopting them: how to evaluate pedagogical quality, cultural relevance, data privacy implications, and alignment with local curriculum standards. Without these frameworks, tool selection is often driven by availability or cost rather than fitness for purpose.
• Clearer approaches to impact measurement – Participants questioned the metrics being used to assess AI’s value in education. Saving teacher time is a useful indicator, but it is not sufficient on its own. The field needs clearer approaches for measuring AI’s impact on actual learning outcomes, and for distinguishing between AI adoption that genuinely improves education quality and adoption that creates an appearance of innovation without changing what students learn. Understanding how AI changes teacher roles and school relationships was also identified as an important area for further exploration.
• Support bridging policy and classroom practice – There is a gap between AI policy discussions at the sector level and what teachers and curriculum designers need to do in practice. Participants called for support that bridges this gap: practical guidance, case studies from comparable contexts, and expertise that speaks directly to the realities of low-resource instructional design.
• Peer learning and shared resources – Several participants expressed strong interest in learning from peers navigating similar challenges. The roundtable itself was valued precisely because it created space for this kind of exchange. Future sessions that bring in practitioners and experts to share specific solutions and case studies were seen as a priority. Cross-organisational collaboration was expected to accelerate progress on shared challenges more effectively than organisations working in isolation.
• Purpose-built, low-resource tools – For organisations operating in bandwidth-constrained environments, the current generation of AI tools is largely inaccessible. Participants called for purpose-built, low-data AI tools, and for investment in developing offline-capable AI solutions that can function reliably in rural and low-connectivity settings.

What this tells us

Across both groups, a coherent and revealing picture emerged of a sector that is genuinely curious about AI, actively experimenting, and running into real, structural, and predictable barriers. A few things stand out.

First, AI adoption in low- and middle-income country education contexts is happening, but unevenly. Some organisations have embedded AI into their systems and are beginning to see real benefits. Others are in early pilots or have tried and found limited success. Many are watching from the sidelines, held back by cost, connectivity, or capacity. The range of applications being explored is broader than many might expect.

Second, the most significant barriers are structural, not attitudinal. Practitioners are not resistant to AI. Many are enthusiastic. But infrastructure gaps, affordability challenges, and limited training opportunities prevent them from using AI well. Closing these gaps requires investment and systemic support, not just motivation.

Third, localisation is not a secondary concern, it is the central one. An AI tool that generates lesson plans in English for a generic global curriculum has limited value for a teacher working with learners in rural Tanzania, Karnataka, or Kabul. The sector needs AI that is built with, not just for, educators in low- and middle-income countries.

Fourth, human agency is non-negotiable. Across both groups, participants were clear that AI should augment, not replace, teacher and curriculum specialist judgement. The risk of uncritical AI reliance, whether among teachers or students, is real and requires explicit attention in any AI integration strategy.

Fifth, the evaluation gap is urgent. The field is investing in AI without yet having the tools to know whether it is working. Developing robust, practical approaches for measuring AI’s impact on learning outcomes, not just efficiency, is one of the most pressing needs the sector currently faces.