"We need to use AI!"
— Everyone.
"For what, exactly?"
— Almost no one.
AI is not a tech problem.
It's a capability problem.
Not "we need better models." Not "we need more tokens." The ceiling on AI value is set by your team's ability to use the tools on real work. Upskill the humans, not the stack.
Patterns We Keep Seeing
The Triangle
Leadership wants results. Employees fear for their jobs. The AI person is caught in the middle. All three perspectives are valid — ignore any one, and adoption fails.
Leadership wants outcomes they can defend to a board. Employees want to know if their job survives the transition. The person actually running AI adoption is trying to build something real without becoming either a hype machine or a layoff architect. Each perspective is correct from where they're standing. None of them is wrong.
What breaks the triangle is picking one side. Ignore the employees and your ambassadors quietly stop advocating. Ignore leadership and you ship things nobody approves. Ignore the AI person and you end up with corporate theater — a strategy deck that never turns into a workflow.
Our move is a strategic workshop before anything else. All three in the same room, same language, same ladder. Not to force alignment — to make the existing disagreement visible so you can actually talk about it.
There are seven distinct leadership archetypes. Which one is yours?The Paradox
A 10-person agency ships a real AI workflow in a Tuesday afternoon. A 1000-person company spends Q1 on an "AI strategy document" that nobody reads. It's not resources.
The familiar wrong answers: "resources", "legacy tech", "compliance". They're wrong because they're obstacles, not the bottleneck. Compliance hasn't stopped anyone from using Excel; it won't stop AI either. And the small org has less money, less tech, less everything.
The real bottleneck is coordination cost. In a 10-person org, one person sees a workflow idea, tries it Monday, shows colleagues Tuesday, it's normal by Friday. In a 1000-person org, the same cycle demands approvals, stakeholder reviews, and a governance committee. The tech isn't slower. The permission chain is.
The Ambassador model works because it compresses that chain to the team level — one person who already has buy-in, running the cycle at 10-person-org speed inside a department of 50. You don't fix coordination cost by adding more meetings. You fix it by moving the decision closer to the work.
The answer isn't what most consultants will tell you.The 80/20
80% of AI failures are not technical. They're organisational. The technology works fine. The people don't.
When an AI project stalls, the instinct is to blame the model. Wrong model, wrong prompt, not enough fine-tuning. Nine times out of ten, the technology worked fine. The organisation didn't.
The specific pattern: teams stop at Level 2 on the Capability Ladder — blank-page crusher, drafting, outlining. Everyone plays with ChatGPT for a week, says "neat", and returns to doing things the old way. No tools get built. No workflows get replaced. The AI sits there as a toy, not a tool.
Level 4 (tool-building / vibecoding) is where 80% of the durable value lives — and where 80% of organisations never arrive. That's when it stops being "using AI" and starts being "your team ships AI-powered tools that solve their own problems." The gap between Level 2 and Level 4 isn't technical. It's organisational permission to actually build.
We have frameworks for diagnosing which rung you're stuck on.The Shadow
Your employees are already using AI. They're just not telling IT. The question isn't whether to adopt — it's whether to do it intentionally.
Personal ChatGPT accounts on phones. Browser extensions for writing. Paste-and-run assistants for formulas. Whatever you pay for at the organisational level matters less than what your employees are already paying for themselves.
A real case: we've walked into organisations where an entire automation department was quietly running their daily work through personal Claude accounts. Good instincts — they picked the right model. Bad governance — customer data flowing through personal logins with no audit trail, no central policy, no way to revoke access when someone leaves.
Shadow AI is a symptom, not the disease. It points to something specific: the people closest to real work know AI is useful before anyone else in the organisation does. When they don't have a sanctioned path, they make one themselves. The fix isn't a ban — that pushes it further underground. The fix is a sanctioned tool at a realistic price point (Teams-plan, not Enterprise) that catches up with where your team already is.
Shadow AI is a symptom. We know what it's a symptom of.Questions Worth Asking
Before buying licenses. Before the all-hands meeting.
When your CEO says "we need AI," what do they actually mean? Efficiency? Innovation? Fear of being left behind?
Who in your leadership team is the skeptic? The visionary? The controller? They're all there.
What happens when an employee saves 2 hours per day? Do they get more work? Get rewarded? Lose their job?
When someone types into Word, what are they actually doing? Not "writing" — what's the real task?
24 Frameworks
Porter. Schein. Kotter. Morgan. Leavitt. Names from business school — applied to the specific challenge of getting humans to use new technology.
The frameworks aren't the point. The conversations they enable are.
Notes from building production AI systems. Click to expand.
[1] The model isn't the bottleneck
Most teams obsess over which model to use. GPT-4 vs Claude vs Gemini. In practice, for 90% of business use cases, it doesn't matter much.
What matters: how you structure the prompt, what context you provide, how you handle the output. The model is maybe 20% of the outcome.
# We have a decision tree for this [2] Prompts are programs, not sentences
A prompt isn't a question you ask. It's a program you write. It has inputs, logic, and expected outputs. Treat it like code.
Version control your prompts. Test them. Document what they do and why. "Just ask it nicely" is not a strategy.
# prompt_v3.2_customer_email_draft.md [3] RAG is overrated (for most use cases)
"We need RAG to search our documents!" Maybe. But probably not yet.
Before building a retrieval pipeline: Can you just paste the document into context? Modern models handle 100k+ tokens. That's a lot of pages.
RAG is for when you have thousands of documents and need semantic search. Most companies aren't there. They just have 50 PDFs.
# When to actually build RAG: [criteria we use] [4] The "first draft + human review" pattern
This is the pattern that works for 80% of business AI use cases:
human_input → AI generates draft
↓
human reviews/edits
↓
final output Not "AI does everything." Not "AI assists while you type." AI produces a complete first attempt. Human fixes what's wrong.
# Why this works when other patterns don't [5] Context window > fine-tuning
"We should fine-tune a model on our company data!"
No. Almost certainly no. Fine-tuning is expensive, brittle, and usually unnecessary. The context window is your friend.
Put your style guide in the prompt. Put examples in the prompt. Put your company terminology in the prompt. It works.
# Fine-tuning checklist: when it actually makes sense [6] Tool use is where it gets interesting
Chat interfaces are version 1.0. The real power is when AI can use tools — search the web, query databases, call APIs, execute code.
MCP (Model Context Protocol), function calling, agent frameworks. This is where production AI is heading. Most companies aren't ready.
# Our tool-use implementation patterns [7] The temperature misconception
"Set temperature to 0 for factual tasks, 1 for creative tasks."
This is cargo cult prompting. Temperature affects token sampling, not "creativity." A well-structured prompt matters more than tweaking temperature by 0.1.
# When temperature actually matters (rarely) [8] The real cost isn't tokens
Teams optimize for token costs while ignoring the real expenses:
API costs: ~$50-500/month (typical SMV) Employee time: ~$5,000-50,000/month Bad outputs: [unmeasured but significant]
Use the best model. Pay for quality. The cost difference is noise compared to the value difference.
# Our cost-benefit framework This page shows roughly 5% of what we've documented. The rest is in detailed playbooks, workshop materials, and diagnostic tools.
We don't share the full methodology publicly.
Curious?
If any of these patterns sound familiar, let's talk. No pitch deck — just a conversation about what you're seeing.
Start a Conversation