Why AI CPD Fails: Three Mistakes Schools Keep Making

Schools are spending more on AI professional development than ever. Sessions are being booked. Speakers are being hired. Training days are being scheduled. And yet, in most schools I work with, teacher AI competency is barely shifting. The investment is going in. The outcomes are not coming out.

Across school AI-readiness work, three recurring design mistakes explain why otherwise well-intended CPD often fails to change practice. They are systemic: they reflect how professional learning is selected, delivered and followed up, and they are amplified by the specific characteristics of AI as a topic.

Mistake 1: One-Off Sessions Instead of Sustained Programmes

This is the most common and most damaging mistake. A school organises a twilight session or a training day on AI. An external speaker comes in, shows some impressive demonstrations, walks teachers through a few prompts, and leaves. Teachers leave the session feeling informed and maybe even inspired. Six weeks later, nothing has changed.

The research on professional development, not AI-specific research, but decades of work on what makes CPD effective, is unambiguous on this point. One-off sessions do not change practice. They change awareness, temporarily, and that awareness decays rapidly without follow-up, practice, and reinforcement.

Effective CPD is sustained over time. It involves repeated engagement with the same concepts, progressive deepening of understanding, practical application between sessions, and structured reflection on what worked and what did not. Joyce and Showers' research on coaching and transfer of training showed decades ago that single-input training has near-zero transfer to classroom practice. Adding coaching, practice, and peer support increases transfer dramatically.

AI CPD is no different. A teacher who attends a one-hour session on ChatGPT will, at best, try a few prompts in the following week. Without follow-up, a structured task to complete, a colleague to discuss it with, a facilitator to troubleshoot problems, that experimentation fizzles out within days.

The fix: Replace one-off sessions with term-long programmes. Structure them in cycles: input session, practical task, peer discussion, reflective review. A minimum of six touchpoints over a term, with each building on the last. This is more resource-intensive than a single training day, but the return on investment is orders of magnitude higher.

Mistake 2: Teaching Tools Instead of Building Understanding

The second mistake is focusing CPD on specific AI tools rather than on the underlying understanding teachers need to use any AI tool well.

I see this constantly. A school buys a licence for an AI platform, a marking tool, a lesson planning assistant, a student feedback generator, and then runs a training session on how to use it. Teachers learn the interface, the features, the workflow. They become proficient in that specific tool.

Then the tool changes. Or a better one comes along. Or the school's licence expires. And the training becomes obsolete, because it was attached to a product rather than to a transferable understanding.

Tool-specific training has its place; teachers do need to know how to use the tools they are expected to work with. But if tool training is all you do, you are building dependency rather than competency. Teachers need to understand the principles that underpin all AI tools; when they encounter a new one (which they will, repeatedly, for the rest of their careers), they can evaluate it, learn it, and use it effectively without needing to be retrained from scratch.

What does this underlying understanding include? It includes how language models work at a conceptual level: not the mathematics, but the principle of pattern recognition and probabilistic text generation. It includes understanding why AI "hallucinates" and what that means for trust and verification. It includes data protection principles: what happens to the data you put into an AI tool, and why that matters. It includes understanding bias: how training data shapes outputs, and why AI-generated content is not neutral. And it includes pedagogical reasoning about when AI adds value to learning and when it detracts from it.

A teacher who understands these principles can evaluate any AI tool they encounter. A teacher who has only been trained on a specific platform is helpless when that platform changes.

The fix: Build your CPD programme around principles first, tools second. Dedicate the first phase of any AI CPD programme to foundational understanding: what AI is, how it works, its limitations, its ethical dimensions. Then layer tool-specific training on top of that foundation. The foundation endures; the tools are interchangeable.

Mistake 3: No Measurement of Impact

The third mistake is the absence of any meaningful measurement of whether CPD is actually working. Schools track inputs, how many sessions were held, how many teachers attended, what topics were covered. They do not track outcomes, whether teacher AI competency has improved, whether classroom practice has changed, and whether student learning has benefited.

This is not unique to AI CPD. It is a systemic weakness in how schools approach professional development generally. But it is particularly damaging in the AI space because the gap between "attended training" and "competent practitioner" is so wide.

In my audit work, I ask schools to demonstrate evidence of teacher AI competency development. Almost all of them can show me a calendar of training events and a list of attendees. Almost none of them can show me competency data: evidence that teachers have actually progressed from one level of AI literacy to another.

Without this data, schools are flying blind. They do not know whether their CPD investment is producing returns. They cannot identify where additional support is needed. They cannot demonstrate progress to their governing body, their accreditation body, or their parent community. And they cannot justify continued investment, because they have no evidence that previous investment has worked.

The fix: Adopt a competency framework, I use a five-level model drawn from UNESCO and other international frameworks, and assess teachers against it at the start and end of each academic year. Use a combination of self-assessment and evidence-based evaluation. Define what evidence demonstrates each competency level (a reflective log, a lesson plan, a tool evaluation, a CPD session delivered). Track the data over time and report on it to your governance structures.

An AI literacy audit can examine the evidence around teacher capability, professional learning and classroom practice against versioned framework expectations. Repeating that review with new evidence can show what has changed, while avoiding causal claims that the documents alone cannot support.

The Root Cause

These three mistakes share a root cause: schools treat AI CPD as an event rather than a system. An event is something that happens once and is then complete. A system is an ongoing process with inputs, activities, outputs, feedback loops, and continuous improvement.

Effective AI CPD is a system. It has a baseline assessment at the start. It has a structured programme of sustained, principle-based learning. It has practical application and peer support built in. It has outcome measurement at the end. And it has a feedback loop that uses outcome data to inform the next cycle of CPD design.

Schools that build this system will develop genuinely AI-competent teaching workforces. Schools that continue to treat AI CPD as an event, however well-designed the individual events are, will spend money, feel productive, and change nothing.

I know which model I would invest in. The question is whether your school is ready to shift from one to the other.