# Not every AI-in-education tale is a horror story. How the world’s leading education company makes AI that’s actually useful for students > Source: > Published: 2026-08-17 09:00:00+00:00 Most of the [AI](https://www.fastcompany.com/section/artificial-intelligence) industry treats a model’s ability to answer questions as a basic measure of progress. That standard becomes less useful when the goal is learning. In education, an answer can be factually correct and [still fail the student](https://www.edweek.org/teaching-learning/opinion-whats-so-wrong-about-a-students-wrong-answer-nothing/2026/07) because it gives away too much or solves a problem without helping the learner understand it. [Pearson](https://www.pearson.com/en-us) has spent the past three years confronting that problem. Its CEO, Omar Abbosh, says the company’s AI transformation has reinforced a basic lesson: Model capability alone does not determine whether an educational product works. “Our products are grounded in learning science and academically strong by design,” Abbosh tells *Fast Company*. “Our customers know we understand learning deeply and design our products using that learning science expertise.” The nearly 180-year-old London-based learning company, which generated 3.58 billion pounds ($4.85 billion at current rates) in revenue in 2025, has [integrated artificial intelligence into products](https://www.reuters.com/business/uks-pearson-reports-5-underlying-sales-growth-q3-2024-10-29/) reaching millions of learners. Those include MyLab and Mastering, its textbook-linked courseware platforms for homework, practice, and assessment across U.S. higher education. Pearson has also deployed an AI-powered math tutor for the GED and offers AI study tools through Connections Academy, its tuition-free online K-12 public school. Pearson embedded generative AI across its portfolio in 2023, became the first major higher-education publisher to put AI study tools inside proprietary academic content in 2024, and established its AI Centre for Enablement in 2025 to coordinate governance, security, and evaluation. Its [technology stack](https://www.reuters.com/business/retail-consumer/pearson-google-team-up-bring-ai-learning-tools-classrooms-2025-06-26) spans Amazon Bedrock, Microsoft, Google Cloud, and IBM watsonx, giving its teams access to multiple models and cloud platforms. Pearson combines those technologies with proprietary content and decades of learner data to develop products for specific educational settings. More broadly, Pearson’s AI transformation offers a window into how a large company deploys AI at scale while deciding where it adds value, how to govern it, and how to measure whether it actually works. “You have to keep changing and pivoting. Over 180 years, we’ve done that many times, and we will keep showing up in the future,” Abbosh says. Rather than chase consumer products, Pearson wants to “remain the infrastructure that helps institutions and employers verify skills and knowledge,” he says. AI models continue to improve, but Pearson treats them as one component of a larger learning system. Abbosh argues that the company’s more durable assets are its expertise in learning science, academic content, and data generated by millions of learners. Pearson uses that data to identify where students struggle, which concepts they revisit, and how their performance changes over time. Product teams can then use those signals to adjust AI-powered learning experiences. Some of the most useful lessons have come from unexpected user behavior. Pearson initially expected its AI-powered Smart Lesson Generator to help teachers prepare lessons more quickly. Teachers instead began using it during class, modifying materials in real time based on students’ level, pace, and interests. Pearson responded by rebuilding the pipeline so teachers could generate smaller activities more quickly and iterate during a lesson. That kind of product knowledge can persist even as the underlying models change. “The model is only one ingredient, and the learning experience is much more than that,” Abbosh says. “What matters is how the model combines with academically sound content, learning science, and educator expertise, to produce an effective, engaging experience.” Pearson has also avoided relying on a single AI provider. Its teams use models from Amazon Web Services, Microsoft, Google, and IBM, selecting more powerful models for some tasks and smaller models or traditional machine learning for others. I asked Abbosh whether that flexibility will remain important as leading AI models become more similar in capability. “Our number one obligation is to our customers. We owe it to them to understand the different model types and how they operate,” Abbosh says. “There are certain use cases that demand frontier models and some that don’t. Our goal is to be on the front end of innovation for our customers and find the right mix for the most effective learning experience.” That raises a broader question for enterprise AI: What remains proprietary when models become easier to swap? David Brudenell, co-CEO of the Sydney-based operational AI company [Decidr](https://www.decidr.ai/us), argues that foundation models will “absorb a lot of what companies currently call their AI stack. “If your advantage is a clever RAG [retrieval-augmented generation] pipeline, you have about 18 months,” he says. “What doesn’t get absorbed is the record of how a specific business makes decisions. Decisions, not data. Most enterprise data is exhaust. Every company has terabytes of it and almost none of it explains why anyone did anything.” Kavitta Ghai, cofounder and CEO of Los Angeles-based [Nectir](https://www.nectir.io/), which builds AI assistants for universities, sees another source of durable advantage in education. “Scale and domain expertise are genuinely valuable. Proprietary content was decisive when good explanations were scarce. It isn’t scarce anymore,” Ghai says. “What compounds is harder to see: relationships with institutions, permission to touch student data, and knowing from experience how a deployment fails. None of that can be bought quickly.” Ghai also argues that access to a capable model does not eliminate the need for specialized education software. “A campus AI license is not a campus AI deployment,” she says. She has seen institutions with free, system-wide access to a leading AI model still purchase separate education platforms because the model license does not include learning management system integration, instructor training, or institutional controls. Pearson has also had to decide which AI functions should be standardized across the company and which should remain with individual product teams. It created the [AI Centre for Enablement](https://plc.pearson.com/en-GB/news-and-insights/news/pearson-announces-newly-expanded-role-chief-technology-officer-dave-treat) to establish common approaches to governance, security, and evaluation. Abbosh says the goal was consistency without requiring a centralized approval process for every AI decision. “For a long time, Pearson operated as a set of separate businesses. That was part of our history of being a holding company. Everyone came with separate technology and processes,” Abbosh says. “Now we are moving toward common functional and common technical architectures that we can repurpose across the company.” Pearson can use that structure to standardize quality measures, security requirements, and responsible-use principles while allowing product teams to adapt them to particular learners and uses. The company has also built a quality-standards agent that helps teams apply those requirements while developing AI-enabled products. More autonomous AI systems complicate that approach. A system approved for one purpose can change as it gains access to new data, tools, or software. An initial review may therefore say little about how the system eventually operates in production. Brudenell at Decidr sees the same problem across enterprise deployments. In his view, companies often scrutinize the model while giving less attention to the credentials and permissions attached to it. An agent might receive broad access through a service account during a pilot, for example, and retain those permissions when it moves into production. Pearson is also trying to determine whether its AI products improve learning rather than simply increasing engagement. In one internal study, Pearson’s learning scientists examined nearly [80 million](https://plc.pearson.com/en-GB/news-and-insights/news/new-data-shows-ai-study-tools-turn-passive-reading-active-learning-college) learning interactions from almost 400,000 students in higher education, and found that users of its AI Study Tool were significantly more likely to exhibit active reading behaviors. Another analysis of more than 62,000 students found that AI-powered adaptive practice made students 90% more likely to reach initial mastery without additional study time. A third analysis, covering about 128,000 AI prompts, suggested that 97% of students used the tool as intended. Those findings provide substantial data on how students interact with the products. They offer less evidence about whether students ultimately learn more. Mutlu Cukurova, a professor of learning and artificial intelligence at University College London, says measures such as active reading and initial mastery capture behavior within the same platform providing the AI. Establishing a causal effect on learning requires a different standard. “More data is good but it often helps with precision, not validity,” he says. “Tens of millions of interactions give you good narrow confidence intervals around quantities that may not be the ones that matter.” Abbosh, for his part, says Pearson is still building that evidence base and is cautious about claims it cannot substantiate. “We have an ongoing research agenda focused on assessing how AI affects learning across different educational contexts and settings,” he tells *Fast Company*. “We can really take it to the next level by marrying the data we have with how students perform on exams and what course grades they receive.” He says Pearson is also pursuing research partnerships with instructors and universities. The limits of the available evidence also affect how Pearson assigns responsibility when an AI system gets something wrong. Human review of every interaction is impractical for products serving millions of learners. Abbosh says Pearson nevertheless considers itself responsible for what those systems do and does not regard them as autonomous decision-makers. The company starts with the desired learner outcome and works backward, considering whether AI is appropriate, what safeguards are required, and how performance should be measured. Its responsible-AI controls also align with the EU’s [AI Act](https://artificialintelligenceact.eu/) and the National Institute of Standards and Technology’s [AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework). Cukurova argues that an AI tutor’s effectiveness ultimately depends on what happens when the tool is no longer available. Students should “be able to solve problems independently, retain what they learned, and apply that knowledge to new situations without relying on the system,” he says. Abbosh acknowledges the challenge. “As these technologies become more capable, our approach is to become more deliberate, not less,” he says. “Internally, we’re experimenting aggressively and learning quickly, but we also want to move responsibly.” Pearson’s strategy depends on whether its learning science, data, [product design](https://www.fastcompany.com/section/product-design), governance, and institutional experience can translate increasingly capable AI models into measurable improvements in learning. The company is working to establish the evidence needed to answer that question.