Lessons Learnt: Building AI-Powered Change Tools

When I set out to create AI-powered tools for change management, I knew I wanted something practical—tools that would genuinely help change practitioners effectively utilise AI in their work. What I got was an invaluable education in what AI can and cannot do. Here's my honest account of building both an interactive toolkit and an AI-powered video.

Part 1: Building the Change Communications Toolkit

The Journey: What I Built and Why

I built this around a specific pain point that every change and communications professional knows: creating change communications for different audiences and formats is time-consuming. Writing the same message for executives, managers, and frontline staff requires different language and emphasis. Scaling this across emails, presentations, FAQs, and video scripts is exhausting.

I envisaged a tool where change professionals would sit with stakeholders to work through the programme attributes—business drivers, impacted groups, resistance points—and then translate that business case for different audiences. AI is particularly good at taking unstructured input—rough notes and talking points—and turning it into concise, audience-tailored language aligned with the speaker's personality, tone of voice, and role.

The vision was a facilitation tool that helps you have better conversations with leaders, then assists with the mundane execution tasks so you can focus on high-value work: aligning leaders, engaging managers, co-creating the journey with employees.

The Build: The First Two Hours

Claude, which is particularly well-suited to coding tasks, proved incredible. I came in with a general idea—an "AI one-stop shop generator" for change communications—and within two hours, I had a functional HTML page with JavaScript and working features. These were skills I simply didn't possess.

The development process was refreshingly straightforward. Plain English commands: "Can you add a save function?" "Make the ADKAR selection more prominent." Claude understood and implemented. The interface was easy to use with version history—if something broke, I could roll back.

Claude went beyond my brief, extrapolating into a fully integrated product with multiple tabs, ADKAR framework integration, project save/load functionality, and a comprehensive help section. Because I had a prototype so quickly, I could see it, test it, and iterate based on reality rather than imagination. The rapid feedback loop was transformative.

The Reality Check: Days Two, Three, and Four

Here's the hard truth: the prototype took two hours. Getting to production-ready took another three days. That's roughly a 12x time multiplier.

What I discovered is that Claude produced what I call "workslop"—work that looks impressive on the surface but lacks depth once you dig in. The prototype was genuinely impressive in a demo. But proper testing revealed it was superficial: buttons that didn't work, save functions that threw errors, code that looked complete but was missing crucial pieces.

Claude hallucinated with complete confidence. It would develop functionality that didn't make sense, quote statistics that didn't exist, and—crucially—tell me it had fixed bugs when it hadn't. I'd test the code. The bugs remained. This happened five times on one particular issue before I gave up and fixed it myself.

Key Toolkit Lessons

1. Prototypes Are Not Products: What you can build in two hours is impressive but not production-ready. Budget 10-20x more time for testing, debugging, and refinement.

2. AI Produces Workslop Without Supervision: Surface-level impressiveness masks deeper issues. You need domain expertise to evaluate quality and technical knowledge to fix bugs.

3. Human Expertise Remains Essential: AI is a tool, not a replacement. For change management, the human work—understanding context, engaging stakeholders, aligning leaders—is where the value lives.

4. AI Extends Your Skills, Doesn't Replace Them: Instead of coding myself, AI helped me generate 99% of working code. It doesn't replace expertise; it extends what you can accomplish yourself when you bring domain knowledge and strategic direction.

Part 2: Creating an AI-Powered Video with Synthesia

The Promise: Democratising Video Production

Traditional video production is expensive and time-consuming. You need filming equipment, presenters comfortable on camera, editing software expertise, and significant time investment. Synthesia promises to eliminate these barriers: write a script, select an AI avatar, and generate professional video content in minutes.

For change practitioners, this could be transformative. Video is one of the most engaging communication formats, but it's often out of reach for most change programmes due to cost and complexity. If AI could genuinely democratise video production, it would unlock new possibilities for leader communications, training content, and stakeholder engagement.

The Reality: A Steep Learning Curve and Human Authorship

Creating the senior leader engagement video was easier than building the toolkit in some ways—no debugging, no broken functions. But calling it "easy" would be misleading. There was a steep learning curve with Synthesia that I didn't anticipate.

The Script: AI as Brainstorming Partner, Human as Author

Let me be crystal clear about this: while AI tools are incredibly helpful in extending your capability, they don't replace it. I used Claude to help brainstorm the script structure and arrange ideas into a logical order. This was genuinely useful—AI is excellent at organising information and suggesting frameworks.

However, I had to be the author of every single word of the content. My 15 years of transformation experience, my understanding of what actually works with senior leaders, my knowledge of common resistance points and effective strategies—none of that came from AI. The strategic thinking, domain expertise, and practical insights were entirely human. AI assisted with structure and organisation; I provided the substance.

This is the right division of labour: AI helps with the mechanics of arranging and formatting; humans provide the expertise, experience, and quality judgement.

The Avatar Journey: Three Trials to Get It Right

Here's where things got really interesting—and frustrating. I wanted to create a custom avatar that looked and sounded like me. The process required three different trials before finding something usable:

Trial 1 - Studio-Style Recording: I recorded myself following Synthesia's requirements—right lighting, right gestures, right environment. Getting everything set up properly was incredibly time-consuming. After uploading my recording, I had to wait overnight for the avatar to be generated. This meant I couldn't iterate quickly—I only found out the next day whether my recording had worked. The result? All good technically, but something still felt off about it.

Trial 2 - Selfie Avatar (8 Versions!): I pivoted to trying Synthesia's "selfie avatar" option, which seemed more promising and quicker to generate. But even this approach had issues. The avatar would zoom in and out randomly during the video, making it distracting and unprofessional. After trying approximately 8 different versions, I finally got one that looked OK—but the voice was very unnatural sounding. The movements were acceptable, but the voice clone just didn't sound like me.

Trial 3 - iPhone Video Upload (The Winner): Frustrated with the first two approaches, I tried a third option: I uploaded a video of myself taken with my iPhone, talking through the exact script. This still required an overnight wait for processing, but the end result was not bad. The voice sounded much more like me, although the intonation wasn't perfect—there were still some artificial-sounding moments. But it was significantly better than the previous attempts. This is the version you see in the final video.

Deliberate Imperfection: Even with the best result from Trial 3, the avatar and voice aren't perfect. But I decided to keep these imperfections as a deliberate feature of the video—to demonstrate the current limitations of AI technology. It's important to be transparent about what AI can and cannot do, and the imperfect avatar serves as an honest representation of where the technology currently stands. It's good, but it's not quite human—and that's worth showing.

Cost Realities: Not as Cheap as Expected

Synthesia requires a subscription, and the costs add up quickly. To generate my 10-minute video, I had to upgrade my plan twice. Each plan has limits on video generation time, and complex videos with multiple scenes consume credits fast.

For one-off videos, this can be expensive. The democratisation of video production is real—you don't need a film crew or editing suite—but it's not free or even cheap for substantial content. For organisations producing regular video content at scale, the economics improve significantly. But for individual practitioners experimenting, budget accordingly.

Synthesia as Accelerator, Not Replacement

Synthesia is a great tool, but it's more of an accelerator rather than a replacement for good content design practices. You still need to:

• Write a compelling, well-structured script
• Understand your audience and what will resonate
• Design clear, logical flow and transitions
• Choose appropriate visual elements
• Review and refine iteratively

What Synthesia does is eliminate the filming, editing, and technical production barriers. The strategic content work remains firmly in human hands.

GenAI Integration: Powerful but Imperfect

Synthesia's integration of generative AI for creating images, videos, and animations was super helpful. I could generate background visuals, supporting graphics, and scene elements quickly without needing design skills or stock photo libraries.

However, there were natural limitations. For example, when I generated a video with the same prompt about people but a different scene, the AI would produce slightly different looking people and environments. This meant the video wasn't 100% consistent from one scene to the next—characters' appearances would shift subtly, background details would change, lighting might vary.

For some use cases, this doesn't matter. For others—particularly brand-sensitive corporate content or storytelling that requires visual continuity—it's a real constraint. You need to work within these limitations or manually ensure consistency, which adds time and complexity.

AI Doesn't Behave Consistently: The Capitals Nightmare

Here's a perfect example of AI's unpredictability that cost me hours of frustration: For 3 days, every time I regenerated the video with a new avatar, everything sounded fine. The majority of the video was delivered by Synthesia's "professional" voices, and I'd included many words in CAPITALS for emphasis (I'd also opted for subtitles, so the visual emphasis mattered).

Then, on day 4, something changed. Suddenly, the voices decided to spell out letter by letter any words in capitals. Instead of emphasising "CREATE A SENSE OF URGENCY", the voice would say "C-R-E-A-T-E A S-E-N-S-E O-F U-R-G-E-N-C-Y". Completely unusable.

The platform hadn't changed. I hadn't changed anything. But the AI's behavior had shifted inexplicably. This forced me to go back and search-and-replace every capitalised word in the entire script—tedious work that shouldn't have been necessary.

The lesson? AI tools can behave inconsistently even when you're doing everything "right". What works on day 3 might break on day 4 for no apparent reason. Budget time for these unexpected setbacks, because they will happen. This unpredictability is a current limitation of AI systems that you need to account for in project planning.

The Learning Curve Payoff

After going through this learning journey with Synthesia—the avatar frustrations, the plan upgrades, the version iterations, the consistency challenges—I now estimate my next video production would be around 4 times quicker. I know what works, what doesn't, how to structure content for the platform, and how to avoid the pitfalls I encountered.

I would absolutely be happy to use Synthesia again. The learning curve was steep, but now that I've climbed it, the tool becomes significantly more valuable. This is true of most AI tools—the first project is painful; subsequent projects benefit from that hard-won knowledge.

Best Use Cases for AI-Powered Video

Based on my experience, I believe Synthesia and similar tools are most useful where a project needs to create:

• Quick instructional videos: Step-by-step guides, how-tos, process explanations
• Microlearning content: Short, focused learning modules that can be produced at scale
• Multilingual communications: The same content delivered in multiple languages with appropriate avatars—this is genuinely powerful for global programmes
• Localised content: Adjusting videos for different regions (UK vs US, for example) where cultural context matters

The UK-US Divide: A Real-World Challenge

One thing I struggled with in my past company was the UK-US divide in communications. Language, humour, cultural references, even business terminology differ significantly. Synthesia could address this quickly—generate one video for UK audiences, adapt it for US audiences.

However, the script itself would need a rewrite. You can't just change the avatar and call it localised. You need to adjust vocabulary, examples, even jokes to ensure they land with the target audience. I deliberately made my video British—for instance, referencing senior leader calendars being "guarded like the Crown Jewels"—which works in the UK but might fall flat or confuse US audiences.

AI tools like Claude can help with this rewrite, identifying culturally-specific elements and suggesting alternatives. But again, you need human judgement to evaluate whether the adjustments actually work. AI extends your capability; it doesn't replace your cultural understanding and communication expertise.

Comparing the Two Experiences

Building the toolkit and creating the video taught me different lessons about AI's current capabilities:

Complexity vs. Simplicity: Coding with AI required deep involvement, debugging, and technical knowledge. Video generation was conceptually simpler but had a steep learning curve around avatars, consistency, and platform-specific quirks.

Refinement Needs: The toolkit needed days of debugging; the video needed hours of avatar iterations. Both required significant human involvement beyond the initial "generate" step.

Cost Structure: Toolkit development consumed time and Claude subscription usage. Video production consumed direct costs (Synthesia plan upgrades) plus time for the learning curve.

Learning Curve Impact: First toolkit project: 12x longer than prototype. First video project: steep learning curve, but subsequent videos would be 4x quicker. Both demonstrate that initial AI-assisted projects take longer than you expect, but the learning compounds.

Human Contribution: Both tools proved that AI extends rather than replaces human expertise. The toolkit needed change management knowledge; the video needed transformation experience and communication skills. AI handled execution; humans handled strategy, content, and quality judgement.

The Bigger Picture: AI's Role in Change Management

These experiences crystallised my thinking on how AI should and shouldn't be used in change management:

Where AI Excels

• Execution of defined tasks: Creating variations of communications, generating video from scripts, formatting documents, producing visualisations
• Organizing and structuring: Taking rough notes and creating logical frameworks, arranging ideas coherently
• Speed and scale: Rapid prototyping, multiple versions for different audiences, multilingual content
• Breaking down barriers: Making video production accessible, enabling coding without programming skills, democratising professional outputs
• Brainstorming and ideation: Suggesting approaches, alternatives, and frameworks to consider

Where Humans Remain Essential

• Strategic thinking: Understanding organisational context, identifying resistance points, designing engagement strategies
• Content authorship: Writing every word based on real experience and domain expertise
• Stakeholder engagement: Building relationships, reading the room, navigating politics, creating genuine buy-in
• Emotional intelligence: Demonstrating empathy, sensing concerns, building trust, managing anxiety
• Judgement and quality control: Evaluating whether AI outputs serve the strategic purpose, spotting subtle issues, ensuring appropriateness
• Cultural understanding: Knowing what will resonate with specific audiences, adapting content appropriately
• Learning and adaptation: Working through steep learning curves, iterating based on feedback, building expertise over time

The Right Division of Labour

The most effective approach is treating AI as a highly capable assistant rather than an autonomous agent:

Humans do: Strategy, content creation, stakeholder engagement, relationship building, quality assurance, decision-making, cultural adaptation, expertise application

AI assists with: Execution, formatting, variations, rapid prototyping, organising information, generating supporting elements, scaling outputs

This division lets change practitioners focus on what adds unique value—understanding organisational dynamics, building coalitions, navigating complexity, creating compelling content—while AI handles time-consuming execution tasks.

Practical Advice for Change Practitioners

Based on building both tools, here's what I recommend:

1. Expect a Steep Learning Curve: Even "easy" AI tools require significant learning. Budget time for experimentation, failures, and iterations. Your second project will be much faster than your first.

2. Budget Realistic Time and Cost: A 2-hour prototype needs 20+ hours to become production-ready. A video project might require plan upgrades and multiple avatar iterations. Plan accordingly and don't promise quick turnarounds on your first attempt.

3. Maintain Strategic Control: Use AI for brainstorming and execution, but keep content authorship and strategic decisions in human hands. Every word should reflect your expertise, not AI's suggestions.

4. Be Transparent About Limitations: If your AI-powered content has imperfections (like my unnatural avatar), consider being transparent about it. This manages expectations and demonstrates honesty about technology's current state.

5. Think Long-Term Value: The first project is the learning investment. If you plan to create multiple videos or tools, the learning curve pays off significantly on subsequent projects.

6. Consider Ideal Use Cases: AI-powered video shines for instructional content, microlearning, and multilingual needs. It's less ideal for highly nuanced emotional content or situations requiring perfect human authenticity.

7. Invest in Quality Scripts: Whether it's video, presentations, or emails, the content quality drives everything. AI can help organize, but you must author based on real expertise.

8. Test Everything: Don't assume AI outputs are correct, consistent, or complete. Test thoroughly, gather feedback, iterate.

The Bottom Line

AI is an incredible assistant for change management work, but it's not autonomous and it's not a replacement for human expertise. It extends your capabilities when you bring strategic thinking, domain knowledge, and quality judgement.

Building the toolkit taught me that AI accelerates work when you know what you're doing—it's a force multiplier, not a replacement. Creating the video reinforced that AI tools have steep learning curves, require human authorship of every word, and work best when you're transparent about their limitations.

For change practitioners: Use AI to assist with execution, organisation, and scaling. Keep strategic thinking, content authorship, stakeholder engagement, and change leadership firmly in human hands. That's where you add unique value. That's where AI falls short.

For anyone building with AI: Expect the learning curve. Budget more time than you think. Be the author, not just the prompter. Test everything. Maintain quality control. Be transparent about limitations. And remember: the second project will be much easier than the first.

Would I do it again? Absolutely. Both tools work, solve real problems, and demonstrate AI's potential. Now that I've climbed the learning curve, I'd create my next video 4x faster and build my next toolkit with far fewer struggles. But I'd still be the strategic thinker, the content author, and the quality controller. AI would still be my assistant, not my replacement.

Have you built something with AI? What lessons did you learn? I'd love to hear about your learning curve experiences—particularly if you've experimented with tools like Synthesia or other AI platforms for change management. Connect with me on LinkedIn to continue the conversation.