I've Now Run AI Workshops for the Same Company Three Times. The Third Batch Is Nothing Like the First.

This week I'm preparing for Batch 3 at Garden — an operations department, 30 people, three Wednesday afternoons starting April 15. When I opened my Batch 1 folder to see what I could reuse, I found 19 slides. I used zero of them.

I've been training at Garden since January. HR first, six sessions, 19 participants. Then Sales & Marketing, three sessions, around 30 people. Now Operations. There's also an executive AI Agents workshop booked for April 20 — the Business Leaders cohort coming together for the second time.

Four separate engagements. One company. My assumption when I started was that the second batch would be easier than the first, and the third easier than the second. That is not what happened.

What the First Batch Taught Me About the Company

HR teams bring a specific kind of fear to AI training. They worry about confidentiality — employee records, performance data, anything that names a specific person. They are also usually the department most aware that AI might affect hiring in ways that touch their own jobs. The two concerns pull in opposite directions: use AI cautiously, but don't stick your head in the sand about where this is heading.

Garden's HR cohort started at Stage 2 on the AI Maturity model — people had heard of AI tools and tried a few, but nobody had a consistent workflow. Six sessions to build foundations: what Copilot can actually do, what the security tiers mean in practice, how to save two hours a week on tasks you already dislike doing.

The security architecture was fixed before I arrived. Nigel, the IT lead, had set the boundaries. Copilot as the primary tool. Tier 1 work — research, summarization — is fine. Tier 2 — formatting, drafting — is fine. Tier 3, anything touching internal customer data, P&L figures, or employee records, stays offline. I agreed with that call. I still do. Working within an institution's actual risk appetite is the job, not a constraint on the job.

By Batch 2, I Had Rebuilt Roughly 60% of the Content

Sales & Marketing showed up with completely different concerns. Confidentiality wasn't their primary worry. They were thinking about whether AI would make their outputs generic — write a proposal, then worry that a competitor's AI generated the same language. They wanted to move faster, not safer. They wanted AI for prospecting, for follow-up emails, for meeting prep that actually saved time rather than adding a new step.

The security boundary that the HR team had understood almost instinctively had to be re-explained from scratch, because Sales' Tier 3 is different from HR's: it's client pricing and deal terms, not employee data.

Three sessions instead of six. A completely different pace. The HR team needed time to get comfortable with the tools before they trusted them. The Sales team disengaged if they didn't see a practical application in session one. I moved the hands-on exercises earlier, shortened the conceptual framing, rewrote the prompts for commercial contexts.

What I kept from Batch 1: the AI Maturity framework, the traffic light protocol for data safety. Everything else was rebuilt.

The Operations Batch Is the One I Am Most Uncertain About

The MO department runs production workflows. They track output by line, manage supplier escalations, handle quality deviation reports. When I started building Batch 3 materials, I realized quickly that a hypothetical exercise — "imagine you're drafting a memo" — would fall flat the moment I walked into the room. These are people who spend their days with actual production data. A fictional scenario built around office tasks would lose them in the first ten minutes.

So I built a 120-row fake production dataset from scratch. Four product lines, patterns embedded in the numbers, designed for Session 3 when we use Copilot to analyze Excel data. I also wrote 26 prompts specifically for operations contexts: shift handover summaries, supplier escalation emails, quality deviation write-ups. Prompts that wouldn't transfer to HR or Sales without significant reworking.

I will be honest about this: I don't know if it will land. I have trained HR teams at this company. I have trained Sales. I have not trained an operations team using this specific approach, with this specific content. Every third batch at a new department inside an organization you think you know feels like walking into a room you've never been in before. The map updates, but the territory keeps expanding.

The Executive Layer Is a Separate Problem Entirely

The April 20 Business Leaders session is technically the same cohort from Session I — Garden's senior management. But the content is completely different. This time it's about AI Agents: Copilot's hidden functions, prompt optimization, building a simple automation. The question at the executive level is not whether to use AI — they've already made that call. The question is what to do once the team below them is using it, and the gap between departments is more visible.

Bess, the MD, made one specific request: no Claude Code. She wants the session built around what is already deployed across the organization, not what's theoretically possible. I think that's the right call. When executives see a capability that the rest of their team can't access, either they get excited about something unimplementable, or they tune out because it doesn't apply to their context. Working inside the constraint is the actual job.

What AI Maturity Actually Looks Like Across One Organization

We talk about AI maturity as if it's a company-wide score. Garden doesn't have one. HR is somewhere around Stage 3 now — people have built consistent workflows and are saving real time weekly. Sales & Marketing is between Stage 2 and 3. Operations hasn't started yet. The executive group is Stage 2 in practice, Stage 4 in intention.

That patchwork is probably the normal state of any organization doing serious AI training. Pockets of genuine adoption surrounded by larger areas of inertia. The training programs don't close that gap on their own — they create more pockets. Whether those pockets eventually connect is a question about organizational culture and management, not about the training content.

I'll know more after April 29.

I write about what I actually observe in training rooms across Hong Kong and beyond. If that's useful to you, connect with me on LinkedIn.