AI For Education in Governments

Exploring AI for Education for governments and our case studies in Saudi Arabia and Qatar

With the advent of AI, about 70 countries (1) have implemented or have begun to implement some AI strategy. These strategies tend to affect most government administrations and, in particular, ministries of education. Education itself has not yet been affected and yet these new technologies offer enormous potential and the opportunity for governments to make great gains, not only by providing their students with the best possible education, but also because they could drastically reduce the workload of their administrative and teaching staff.
As shown in the image, the incorporation of AI in educational settings not only affects students and educators but also plays a crucial role in data-driven decision-making at the institutional and governmental level. This often requires integrating intelligent systems that work together to maximize outcomes.
It is important to understand the difference between an AI tool and an AI-powered system. For example, adding an AI tutor or personalized learning experience to an academic institution does not necessarily make it an AI-powered system, just as adding a streaming feature to a bookstore’s website does not transform it into Netflix.

What unlocks the value of AI-driven systems is the availability of centralized data from different types of sources, combined with AI tools and automated workflows. For example, to identify students at risk, one would need to drill down into student psychological assessments, extracurricular activities, learning speed, current teachers, school/government policies and existing support systems (mentors, advisors, AI agents, etc.), grades, etc. This process is time-consuming for educators and, aside from a few elite private schools in the world, most schools in the public system do not have the resources to make informed decisions for each student. Their student-teacher ratios are quite high and even when data is available, teachers do not have the energy or incentive to follow up with each student.
This is where AI-based systems come into play. With generative AI, these systems could incorporate all available data from entire countries and identify at-risk students in real-time. In addition, they could automatically inform key stakeholders (administration, schools, teachers, and even families) by automatically generating emails, student reports, and recommended interventions for each stakeholder. This is an example of a use case, but these AI-powered systems are capable of much more: organizing class schedules, coordinating the assignment of substitute teachers at the government level, generating customized courses for each class, or providing academic guidance and recommendations about the ideal educational and career path, to name just a few examples. These AI-powered systems can help ensure that a country’s human capital (its students at any given time) has the best possible academic careers and career options. 

Talking about the advantages and potential of these AI-based systems in a short article will not do it justice. Nevertheless, we are going to delve into some of the AI tools that make up this type of AI system. More specifically, we will analyze the characteristics of Generative AI used in a specific course.
The latest generative AI tools can be used to produce Jupyter notebooks, PowerPoint presentations, reports, assignments, quizzes, exams, videos, grade open-ended answers on an exam, provide personalized feedback, act as an AI tutor, provide recommendations, and much more. Educators no longer have to spend time generating content or grading every aspect of student presentations, which would save them an enormous amount of time daily.
These AI-driven systems do not make teachers irrelevant or obsolete elements of learning. Educators will continue to be needed for didactic support and to assess a wide range of skills that AI-powered systems cannot analyze. Teachers also remain critical to the emotional development of learners. And yet, when it comes to the enormous administrative tasks that educators have to perform, is there anything that AI cannot take on?
Let’s take a look at the learner side. Services such as personalized/adaptive learning, AI-driven tutoring, and the ongoing availability of some form of assistance are crucial for learners. Surprisingly (despite the global excitement around AI and the inflation of all the associated services), there are many relatively inexpensive strategies for implementing these technologies without resorting to large language models (LLMs).

AI in large-scale education: strategies, phases and optimization techniques.

When implementing AI in education at the government level, it is necessary to establish strategies and phases to familiarize the ecosystem with such a disruptive technology. Depending on the government and the entity, it is advisable to start by delimiting the scope of the project as follows: 

1. In-depth analysis of the education system: a thorough analysis of the current education system is needed to identify areas where AI can bring immediate benefits (“low-hanging fruits”).
2. Strategic plan for AI integration: at this stage, a strategy would be developed for AI integration at multiple levels of the education system, from the ministry to local schools and universities.
3. Stakeholder involvement: all stakeholders, including educators, administrators, parents, and students, need to be involved to ensure that their needs are met and to facilitate smooth adoption.
4. AI transformation timetable and budget allocation: a realistic transformation timetable must be drawn up for the entire education sector, accompanied by budget planning and a precise allocation of the resources that would be mobilized.
5. Implementation plan: in this phase a clear implementation plan is outlined, specifying roles, responsibilities, and milestones to ensure successful implementation and adoption.
6. Executive capacity building and training: it is essential to conduct specific training sessions for the executive levels of public education administrations and school principals to develop leadership competencies related to AI implementation. The ultimate goal is to ensure that they are empowered to drive AI-related initiatives within their institutions. 
7. Monitoring and evaluation framework: metrics and benchmarks are established to periodically assess the impact of AI integration and allow for necessary corrections. 

The results of scoping an educational AI implementation project translate into: 

• A concrete AI strategy for education adapted to the national level. 
• Detailed timetable for AI transformation with budget allocation appropriate to the objectives.
• Clear implementation plan with roles and functions defined and stakeholders identified.
• Leadership trained in AI for education, capable of driving initiatives.
• Accountability metrics to ensure implementation and adoption.

Educational objectives and the identification of public needs are one of the main tasks of administrations. However, these are not enough to conceive an AI implementation plan. There are aspects such as training and tuning of AI-based systems that have such an impact on results that it is advisable to entrust the elaboration of the AI implementation plan, as well as its technical direction, to experts. 

Executive Upskilling in educational administrations

One of the critical ingredients for the successful implementation of AI in education is training and capacity building at the highest levels of educational administration. 

These are the key components of an AI training program for education:

1. Introduction to AI in education.
2. Data management literacy.
3. Digitization of educational legislation and preparation of formats that allow the AI to extract information automatically.
4. Personalized and adaptive learning.
5. Unit plan, lesson plan, and task sheet generation.
6. Assignments, exams, and test generation.
7. Personalized evaluation and automatic evaluation processes.
8. AI-assisted monitoring of public education.
9. Returns on investment in AI for education (this point must be customized for each country or region).
10.  Deployment strategy and execution plan (also customized for each country).
11.  Infrastructure requirements.
12.  Impact evaluation and assessment.
13.  Budgeting and resource allocation.
14.  Technical assistance and troubleshooting.
15.  Maintenance and supervision of AI systems.
16.  Ethics, privacy, and data security.
17.  Change management and continuing professional development for educators.
18.  Case studies.

Training is not identical for all parties involved. Depending on the role and levels of responsibility, each party will tend to focus on different modules in the list. 

Case study of Saudi Arabia

Saudi Arabia has started to implement AI in its education system and over the last year LiveTech AI has been working very closely through KAUST with their National Center of AI to assist in a national deployment of AI in public sector education. One of the first challenges it faced was parsing Arabic textbooks. Arabic is considered a low-resource language in terms of AI. A low-resource language is defined as a language for which little data is available to train AI models. Many OCR models have difficulty parsing textbooks. Transforming Arabic curriculum into formats that can be assimilated and processed by AI is one of the most critical steps when it comes to AI-driven education. After several iterations, LiveTech managed to create a system that automatically parses PDF books, extracts learning objectives, practical problems (with their pictures), lessons, glossaries, and examples, and presents them in a structured format.
The results were then used to create personalized and adaptive learning systems, enhanced with generative AI to produce virtually unlimited repositories of exercises. These activities provide each learner with a vast bank of activities to practice and meet any needs they may have.
Generative AI was also used to create a STEM tutor, which was complemented with tools such as Wolfram Alpha to produce a guide that showed students how to solve problems step-by-step, rather than just giving them the answers.
Other tools requested by the Saudi administrations, and developed under this project, included the generation of PPTX and customized lesson plans based on classroom performance and curriculum. This customization takes time, but it is important. When preparing a pilot project, it should be kept in mind that whatever is built should be scalable nationally. This often means that proposals such as those based on OpenAI or other APIs may not be sustainable because of their high cost per pupil.

Qatar case study 

Qatar, on the other hand, adopted a different strategy and proposed a top-down rather than bottom-up approach. LiveTech AI has been collaborating with their Qatar Academy since 2023 to harness the potential of AI and help scale it nationally within the country. In this project, it has been developing an intelligent tool that acts as a layer that centralizes the many educational platforms being used.
This allows educators to make informed decisions quickly, as well as to consult on the educational status of their students, identify at-risk students, and generate reports.
This method allows the use of retrieval-augmented generation (RAG), a technique that helps to produce more robust generative AI models by using data extracted from external sources (in this case, any document related to or produced in the educational system). Retrieval-enhanced generation makes it possible to perform intelligent searches for information in restricted-access databases.
This technique is complemented by tools for the automatic generation of graphs, reports and tables. This frees educators, administrators and potentially even entire governments from performing the more tedious tasks associated with managing the education system.
While LiveTech AI provides the engineering know-how, Qatar Academy brings experienced educators and administrators who help design the product to solve the schools’ main weaknesses. Their expertise comes directly from first-hand experience in teaching and learning processes from different perspectives. This allows LiveTech AI to customize and create features that from day one are tailored to the schools’ teachers and administrators, leading to improved operations and results.
We believe that this collaborative effort between AI industry specialists and Qatar Academy streamlined decision making at critical moments in the project and is a good model for future projects implementing these technologies. 

ROI of AI in education in the public sector

In terms of operational efficiency, AI can easily save educators 25-30% of their daily tasks, and even more in the case of managing the administration of the ministries of education. However, the question that all administrations are probably asking themselves is: how much can it cost to implement AI in education on a national scale?

As a general rule, we have found that an investment of less than 0.5% of the budget of a ministry of education can lead to an increase in operational efficiency of more than 25% (a very conservative estimate).
A ChatGPT subscription costs $20/user/month, which makes it prohibitively expensive for public education, despite the tireless effort put in recent years to reduce this cost through extreme optimization procedures and proprietary models/systems. The goal to be pursued is to make AI deployments on a national scale affordable for less wealthy countries. Currently, we are close to reducing the cost to a few dollars (single digit) per user/year. Of course, this price will vary depending on the features enabled and the existing infrastructure. This significant reduction will only be possible through a unique combination of hardware and software optimization, to create a fully integrated AI solution at all levels of the education system.
At this point, we will try to answer the question: why is it so expensive to deploy these AI systems?

Implementation of AI in education

To run many of the AI tools, large language models (LLMs) running on GPUs are required. This section shows how to calculate the number of GPUs required to run the AI tools needed for a national deployment. 

Why are GPUs important?

GPUs are Graphics Processing Units (GPUs), electronic circuits dedicated to performing a very high volume of mathematical operations at high speed. GPUs execute many requests in parallel, enabling faster data processing and increasing the capacity of tasks, which is crucial for training and inference in deep learning models. Its multiple cores perform computations simultaneously, ensuring low latency and high energy efficiency. This efficiency is vital for large-scale deployments, as it reduces costs and environmental impact. According to NVIDIA, GPUs are 42 times more energy efficient than CPUs (2) in AI inference tasks. AI models require large datasets and high computational power, which GPUs efficiently provide for both training and real-time inference.
The capacity of the GPU’s RAM, VRAM (Video Random Access Memory), is crucial in determining the size of models and data sets that can be processed directly on the GPU without relying on the system’s main memory. Greater memory capacity allows larger models to be handled more efficiently, reducing data transfers and improving performance.
Advanced memory technologies such as HBM (High Bandwidth Memory) provide superior performance per watt, delivering up to three times the bandwidth compared to conventional memory types, which is vital in power-constrained environments to reduce operating costs and promote sustainability. Maximizing GPU utilization by optimizing batch sizes and leveraging advanced memory technologies improves the overall efficiency and performance of GPU-accelerated systems, contributing to improved throughput and energy efficiency.

Now that we know that GPU memory influences model size and data sets used, how do we measure GPU performance? 

GPU performance is measured in floating point operations per second (FLOP). Precision, which varies with the number of bits used, refers to the level of numerical accuracy of the calculations. The number of bits used determines the degree of precision with which a value can be represented: double precision (64 bits) provides high numerical accuracy and a wide range of values, single precision (32 bits) balances precision and memory usage, and half precision (16 bits) provides lower precision with significant memory savings. Multiply-Accumulate (MAC) operations, common in matrix multiplications, are equivalent to 2 FLOPs each. Machine learning models are composed of layers that perform operations such as matrix multiplications and convolutions. They rely on FLOPs to measure computational needs and efficiency, which is critical when optimizing performance and planning hardware deployments.

Here is a formula to calculate the FLOP/s of the GPU:

GPU FLOP/s = clock_frequency x cores x FLOPS_per_clock_cycle x float_point_accuracy

Important considerations

It is important to take the following aspects into account when planning computing requirements: 

• Theoretical versus actual performance: actual FLOP/s may be lower than calculated due to variations in clock speed (number of cycles the CPU executes per second) and operating efficiency.
• Variability in operations: the number of floating point operations per clock cycle may vary according to the type of operation.

Inference time

Framework and library support: Tools such as NVIDIA’s CUDA Toolkit and cuDNN offer various APIs to help with profiling and GPU FLOP/s estimation.
Calculation of the inference time: the inference time can be determined by the formula: 

Inference time = (FLOPS of the model) / (FLOPS/s of the GPU)

Calculation of the number of GPUs required

So, now that we know how to obtain the processing time as a function of the model size (i.e., the FLOPS of the model), we can quickly calculate the number of GPUs required for AI deployment in a national education system. This depends on the following 6 criteria:

1. Model size and complexity: larger and more complex models require more computing power.
2. User concurrency: the number of users interacting with the AI simultaneously.
3. Frequency of requests: the number of requests made by each user per minute.
4. Latency requirements: the acceptable delay between a user request and the AI response.
5. Batch processing vs. real-time processing: whether AI tasks are batch processed or need real-time processing.
6. GPU specifications: the performance characteristics of the GPUs used.

The total number of GPUs is calculated as follows: 

Concurrent_users×Petitions_per_user_per_Minute×Average_processing_time_per_request (in seconds)

---

GPU capacity (requests/second)

The numerator indicates the total processing demand in requests per second. It can then be divided by the GPU capacity (requests per second) to determine the number of GPUs required.

The average processing time per request can be calculated from the FLOPs of the model divided by the FLOPs per second (FLOP/s) of the GPU, as shown above.

With this information, any educational administration will be able to estimate its own needs based on the size of the model it plans to use and the requirements of its country. 

Conclusion

Implementing AI in education at the national level requires a strategy that encompasses skills enhancement, software and hardware infrastructure. We have addressed the four essential categories and summarized the requirements that affect all education administrations. Implementing AI-powered systems has the potential for tremendous returns on investment and can significantly improve not only operational efficiency, but also outcomes for administrators, educators and learners.
The possibilities of the responsible use of AI in education should not be available only to students with more resources. In this article we have tried to show what variables need to be taken into account to enable public systems to carry out a profound transformation of education.
The challenge is considerable but manageable. In a context such as the current one, in which many narratives around the use and possibilities of AI promote exclusionary visions that magnify digital and economic divides, this article advocates an urgent but, we believe, realistic improvement of educational systems.
We believe that the risks of missing these opportunities are greater than the gains in an Ibero-America that needs at all costs to recover lessons learned and become a benchmark of prosperity and productivity. 

References

1. OECD (2024). National AI policies and strategies. Retrieved from https://oecd.ai/en/dashboards/overview.
2. Harris, D. (2023). “What’s Up? Watts Down – More Science, Less Energy”.
3. Retrieved from GPUs Lead in Energy Efficiency, DoE Center Says | NVIDIA Blogs