After years of investing heavily in ed tech, school districts are asking an important question: What actually works?
The era of adopting every tool that promises to help students or teachers is giving way to a more disciplined approach. In Illinois, one school district is reviewing tool usage data alongside feedback from its technology committee to identify tools that see little to no use. In California, another district is piloting an outcomes-based contract where payment depends on the tool meeting student achievement goals. Districts are prioritizing solutions that address specific student needs, integrate with existing systems, align with instructional goals, and most importantly, demonstrate measurable impact. In this environment, trust has become the priority.
The Tools Competition, a program of Renaissance Philanthropy and organized by The Learning Agency, offers a unique view into how AI in education is moving from hype to implementation. Drawing on six years of submissions, a new report by Meg Benner, Kristyn Manoukian, Joon Suh Choi, and Kumar Garg, shows that while AI adoption has accelerated rapidly, the strongest tools share three characteristics: they integrate AI into authentic learning experiences, evaluate their effectiveness, and are intentionally designed to improve outcomes for learners and teachers who have too often been left behind.
From AI Adoption to Implementation
The newest report from the Tools Competition tells a story that is now playing out in school systems across the country. Looking across six years of submissions, particularly the competition’s 171 winners, it traces how AI adoption evolved rapidly. But while the technology changed quickly, the quality of implementation did not always keep pace.
In the first two competition cycles, launched in 2021 and 2022, few winning proposals mentioned or utilized AI. At the time, foundation models were not easily accessible without machine learning expertise, limiting their incorporation into ed tech tools.
That changed quickly. Within a year, ChatGPT made foundation models widely accessible. By the third competition cycle in 2023, foundation models had become the base layer for roughly three-quarters of winning proposals. By cycles four through six (2024–2026), as multimodal and agentic AI tools became more accessible, competitors shifted from simply adding AI features to deliberately designing around these new capabilities.
What changed more slowly was the depth of technical implementation: whether developers were truly engineering around these capabilities rather than simply naming them.
That distinction ultimately separated the strongest submissions from the rest. The winners weren’t defined by whether they used AI, but by how intentionally they applied it to solve real educational challenges.
What Leading Competitors Are Building
The strongest competitors are not using AI simply because it is available. They are using it where it solves a clear learning problem, addresses a specific user need, and fits within a realistic implementation context.
These tools are moving beyond text-based interactions to support speech, images, voice, and other forms of engagement. They are also exploring how AI can reason through multi-step tasks, support more adaptive learning experiences, and function in low-resource environments.
Multimodal and voice interfaces have become a leading attribute of competition proposals. These tools reflect the needs of young children, early readers, multilingual learners, neurodivergent learners, and students with disabilities who may not be best served by text-only interfaces.
The strongest competitors are not using AI simply because it is available. They are using it where it solves a clear learning problem, addresses a specific user need, and fits within a realistic implementation context.
In 2026, about half of winning tools incorporated voice interaction, animated avatars, or speech recognition. KIVA, a recent winner, uses an animated AI avatar to help young children learn. The avatar engages students in conversations about the texts they read or listen to, using speech recognition in both English and Spanish.
Other tools are moving beyond one-off interactions to build systems that can reason through problems, plan, and take action. These approaches are focused on reducing teacher workload while supporting more adaptive learning experiences. PAL for Early Math Learning, a 2026 winner, uses agentic systems to support adults guiding children’s math learning. The tool incorporates what adults observe, asks targeted follow-up questions, and generates tailored guidance based on a child’s knowledge and understanding.
A small but growing number of tools are also being designed for environments where connectivity is limited. Developers are exploring ways to run AI on classroom hardware or basic smartphones, allowing students and teachers to continue using these tools in rural schools, low-bandwidth environments, and the Global South. These approaches can improve access while also strengthening privacy and reliability.
SmartCoach Assess, another 2026 Tools Competition winner, is an offline app designed to support reading assessment in low-connectivity classrooms. The tool helps teachers administer EGRA-aligned assessments for grades 1–3 and has been piloted in schools across Ghana and Uganda with strong randomized controlled trial results.
Importantly, this evolution does not represent a replacement of foundational models. Instead, leading developers are combining new models and existing foundation models in thoughtful ways to create tools that better meet the needs of learners and educators.
A small but growing number of tools are also being designed for environments where connectivity is limited. Developers are exploring ways to run AI on classroom hardware or basic smartphones, allowing students and teachers to continue using these tools in rural schools, low-bandwidth environments, and the Global South.
What Sets Strong Proposals Apart
The strongest proposals do more than explore new AI capabilities, they thoughtfully apply them.
Leading submissions do not simply select the highest-quality AI model available. Instead, they take a strategic approach: comparing multiple model options and selecting the one that performs best for the specific task and context. For these teams, model choice matters less than how well the model is evaluated and adapted to solve a real problem.
In some cases, this requires customizing the models already available. Off-the-shelf models can fall short when applied to specific learners or educational contexts. For example, speech models such as Whisper, Azure Speech, and similar services have been found to underperform when recognizing child speech, atypical speech, and accented speech. Without customization, these limitations could make tools inaccessible to the very learners they are designed to support.
The strongest proposals also build on what has already been shown to work. Winners are extending existing platforms, products, research bases, and infrastructures while leveraging new open-source tools, capabilities, models, and datasets that can serve as public goods for the broader field.
This approach allows developers to move faster from a tested foundation while focusing their innovation where it matters most. The advantage is not simply having access to powerful AI tools, it is knowing how to adapt them to solve a specific educational challenge.
Leading submissions do not simply select the highest-quality AI model available. Instead, they take a strategic approach: comparing multiple model options and selecting the one that performs best for the specific task and context.
Designing for Trust, Responsibility, and Equity
As AI continues to transform education, innovation alone is not enough. New technologies must also earn trust by demonstrating responsibility, protecting learners, and expanding access. The strongest proposals are beginning to treat these considerations as core design requirements rather than afterthoughts.
Leading competitors are approaching responsible AI as a necessity, building systems that consider privacy, bias mitigation, hallucination guardrails, human oversight, and data governance. They are also increasingly accounting for important policies and regulations such as FERPA, COPPA, and HIPAA, considerations that appeared far less frequently in early competition proposals.
Competitors are not simply discussing these issues; they are building systems with them in mind. In 2026, Tools Competition submissions included deeper consideration of approaches such as constitutional AI, explainability, and federated learning. This reflects both the competition’s emphasis on clarity in these areas and a broader recognition that responsible AI requires intentional design choices.
Equity is another area where leading proposals are showing meaningful progress. Rather than treating access as a separate consideration, strong submissions are aligning product and technical choices with the needs of the learners they aim to serve. In 2026, many proposals focused on supporting neurodivergent learners while others addressed challenges in low-bandwidth, offline, and low-connectivity environments.
The field is raising the bar, but important gaps remain. Many proposals still need to better explain how student data will be managed, retained, and deleted to ensure stronger privacy protections.
As AI continues to transform education, innovation alone is not enough. New technologies must also earn trust by demonstrating responsibility, protecting learners, and expanding access. The strongest proposals are beginning to treat these considerations as core design requirements rather than afterthoughts.
The Future of Ed Tech
Schools are looking for tools that can meet their needs in action, not simply in words. Now more than ever, the opportunity is to create edtech tools that deliver the measurable gains districts want to see for their students and teachers.
Insights from past submissions to the Tools Competition show that many developers are moving in that direction. They are incorporating AI and new techniques through thoughtful approaches designed to meet the needs of the students they serve, while also ensuring that students are protected and that important policies and responsibilities are considered.
The results of this report show that ed tech continues to evolve, and that evolution should be expected. Like AI itself, ed tech will continue to become more capable, more explainable, and more responsible. The goal is not to move away from these tools, but to continue preparing for what comes next by building technologies that earn trust and create meaningful outcomes for learners and educators.
This column first appeared in Forbes.
