Overview

Sofija Iyer's job title was Head of Learning Transformation, but for most of her first year at a 7,000-person manufacturing company, her real job was running a sprint - delivering AI literacy training to employees faster than the organisation's appetite was growing. She started with a two-day workshop that she delivered personally to cohorts of 25. The waiting list hit 800 people within three months. "I became a bottleneck," she told me. "The demand was real, the content was good, but I was the single point of failure. You can't upskill a company of 7,000 at that pace one workshop at a time."

Scaling AI training is one of the most underestimated execution challenges in enterprise AI programmes. Most organisations plan the content. Fewer plan the delivery infrastructure that allows that content to reach thousands of people without depending on a single expert or a single format.

Why One Format Is Never Enough

Different learners need different things. A plant floor supervisor who works rotating shifts cannot attend a Tuesday afternoon webinar. A data analyst who already knows the basics does not need a two-day introductory workshop. A senior manager who will make decisions about AI investment needs different content than the analyst who will execute those decisions.

Effective at-scale training blends four modalities, each serving different needs.

In-person instructor-led training is the highest-engagement format. It allows real-time Q&A, peer discussion, and the kind of social learning that sticks. It is also the most expensive and least scalable - you are limited by the number of qualified instructors and the physical space available. Use it for cohorts where engagement and relationship-building matter most: senior leadership, cross-functional change agents, and the first cohorts of a new programme before the content has been proven.

Virtual instructor-led training extends geographic reach without sacrificing all of the live interaction. It scales better than in-person - a skilled facilitator can run virtual sessions for 40-50 people - but requires more deliberate design to maintain engagement. Short modules (60-90 minutes maximum), frequent interaction, and breakout group work all compensate for the reduced social presence of a screen.

Self-paced online learning scales to tens of thousands of learners with no marginal cost per learner once content is built. It provides flexibility for shift workers, distributed teams, and learners who need to move at their own pace. The challenge is completion rates: self-paced programmes without accountability structures typically see 5-15% completion. Build in social accountability - cohort-based programmes where people learn together over a defined period perform significantly better than fully asynchronous open-ended programmes.

Peer-to-peer learning is the most underused modality in corporate AI training. Your organisation already has people who have learned to use AI effectively. They are the fastest and most credible teachers for their peers - they share the same context, they understand the same constraints, and their knowledge is immediately applicable. Identifying these internal experts and giving them structures to share what they know is high-leverage work.

Building a Train-the-Trainer Programme

Sofija's solution to her 800-person waiting list was not to hire more trainers. It was to build a train-the-trainer programme that identified AI-capable volunteers from each business unit and equipped them to deliver adapted versions of the core curriculum to their own teams.

A train-the-trainer programme has three components. The first is *selection* - identifying people who combine genuine AI capability (they use AI tools effectively themselves) with interpersonal credibility in their team (their peers will listen to them). These are not the same person. Someone who is technically excellent but seen as out of touch will struggle as a trainer.

The second is *training the trainers* - not just on the content, but on facilitation skills, how to handle common objections, and how to adapt the core material for their specific team's context. A trainer in finance needs to know how to translate examples into financial workflows. A trainer in operations needs to understand shift worker constraints. Generic train-the-trainer programmes that only teach the content produce trainers who deliver generic training.

The third is *ongoing support* - a community where trainers can share what is working, get help with difficult questions, and access updated materials as AI tools and organisational priorities evolve. Trainers who feel isolated and unsupported typically deliver once or twice and then stop.

Sofija's programme trained 60 trainers in eight months. Those 60 trainers collectively reached 3,400 employees - nearly half the organisation - in the twelve months that followed. Her waiting list cleared. The quality was not identical to her own workshops, but it was high enough to produce measurable behaviour change.

Building for Self-Service

Beyond formal training programmes, employees need resources they can access in the moment when they encounter an unfamiliar AI situation. A library of how-to guides, short video walkthroughs, and worked examples covering the AI tools and workflows your organisation uses gives people a way to learn just-in-time rather than just-in-case.

The difference between a resource library that gets used and one that doesn't is specificity. Generic guides to how ChatGPT works are less useful than guides showing how to use the specific AI tools in your organisation's approved stack, with examples drawn from your actual workflows. "How to use AI to draft a supplier communication" beats "Introduction to prompting" for the procurement manager who needs help right now.

Building this library is not a one-time project. It needs maintenance. AI tools update constantly. Approved tool lists change. New use cases emerge. Assign someone - ideally with a time allocation, not just a vague responsibility - to keep the library current and to add new resources based on common questions that come through support channels or trainer communities.

Measuring Training Effectiveness

Most organisations measure training at the wrong level. They count completion rates and satisfaction scores. Those numbers are easy to collect and tell you almost nothing about whether the training is working.

The four-level framework from learning evaluation gives a more useful structure. Level 1 is learner reaction (did people like it?). Level 2 is learning (did they acquire the knowledge or skills?). Level 3 is behaviour (are they doing things differently?). Level 4 is results (are those behaviour changes producing business outcomes?).

Most organisations measure only Level 1 and partially Level 2. The levels that matter for AI programmes are 3 and 4. Are employees actually using AI tools in their work? Are they using them in approved, safe ways? Has the time spent on certain tasks changed? Has error rate on AI-assisted tasks changed?

Sofija built a simple Level 3 measurement: she sent a five-question survey to managers of trained employees 60 days after training, asking whether they had observed employees using AI tools, and whether the use seemed appropriate and valuable. The response rate was 70%. The data told her which business units had strong follow-through and which needed additional support - information she could not get from completion rates alone.

Keeping Training Current

AI is one of the fastest-moving capability areas any organisation has trained. Content built in 2024 may already be partially outdated in 2025. New models emerge. Approved tool sets change. Regulatory guidance evolves. Regulatory guidance evolves. Organisational AI policies update.

A training programme without a refresh cadence becomes a liability - teaching people outdated practices while creating false confidence. Build refresh cycles into the programme from the start. Core content should be reviewed every six months. Tool-specific content should be reviewed every quarter. Major changes to the AI landscape - a significant new capability, a change to approved tools, a relevant regulatory development - should trigger immediate spot updates, not wait for a scheduled cycle.

Learner feedback is the fastest source of outdated-content signals. Trainers and community managers should have a simple mechanism to flag when something in the materials no longer matches reality. A monthly "flag log" reviewed by the programme owner costs one hour a month and prevents months of incorrect practice spreading through the organisation.

Key Takeaways

• A single training format cannot reach an entire organisation. Blend in-person, virtual, self-paced, and peer-to-peer modalities. Each serves different learners, constraints, and learning goals.
  - Train-the-trainer programmes multiply your reach without multiplying your cost. Identify credible internal experts, invest in their facilitation skills and context-adaptation, and support them with a community. Sixty good trainers can reach thousands of employees.
  - Self-service resources must be specific to your organisation's tools and workflows. Generic AI guides have low uptake. Resources that answer the exact questions people face in their actual jobs get used.
  - Measure behaviour change, not completion rates. Level 3 (behaviour) and Level 4 (results) measurement tells you whether training is working. Level 1 (satisfaction) tells you whether people liked it - useful but not sufficient.
  - AI training requires refresh cycles, not a one-time build. Review core content every six months, tool-specific content every quarter, and trigger spot updates immediately when significant changes occur.
  - Learner feedback is the fastest outdated-content detector. Build a simple flagging mechanism so trainers and community members can report when materials no longer match reality.