Rolling Out AI That Actually Sticks
Great AI starts with a single prompt… and sometimes ends with 2,500 out-of-office replies.
A days ago, my inbox looked like a North Korean military parade: row after row of identical “I’m on vacation until…” emails. A newsletter had gone out with my address as the reply-to (oops), and now I had 2,500 auto-responses doing cartwheels in Gmail.
Fifteen minutes, one Google Apps Script, and, obvisiouly, a lot of ChatGPT later, the mess was sorted. A combination of Google App Script + Google Sheets + OpenAI APIs for sorting it all, categorizing them, and even killing emails of people who left their companies. A microscopic win, sure, but it hides a larger truth:
Rolling out AI is a change-management challenge, not a tech install.This exercise would have taken me hours (even days) to code myself. But knowing the right tools and capabiltiies, while building on top of past experiments, had it take about ten minutes.
And who cares? This is, all in all, a fairly useless example, But it reflects on how much can be done when you know how to wield your tools. That has become, increasingly, a larger part of my job - making sure that people can learn how to make magic with AI.
So I figured it would be a good opportunity to share some stuff I’ve learned along the way while doing just that.
Below are lessons I keep bumping into while helping teams adopt AI. I definitely don’t have all the answers. Think of this as a field notebook, scribbled in the trenches and lightly dusted with coffee stains.
1. The Pain Points (a.k.a. Why Everyone Keeps Clicking “Remind Me Tomorrow”)
- Capability Whiplash – People either assume AI is magic or they think it’s Clippy on steroids. Both break momentum. Knowing what you can and can’t do is a great starting point.
- Secret Cyborg Syndrome – Employees tinker in the shadows because compliance talks sound like HR horror films. There’s a whole lot more going on under the scene but people are afraid to talk about it. We need to normalize using AI for everything.
- ROI Bermuda Triangle – Experiments pop up, look clever, then vanish before they reach a KPI anyone cares about. Instead, focus on experiments that actually matter. Ideally, they make money. Close second? They save it. Distand third…saving time.
2. Two Lenses To See The Board
a. Bottom-Up, Top-Down & The Lab
Ethan Mollick’s Leadership / Lab / Crowd triangle maps neatly to:
- Bottom-Up (The Crowd) – grassroots tinkering, hashtag-filled Slack channels, Friday demo-offs.
- Top-Down (Leadership) – mandates, guard-rails, and budget lines that say “yes, this matters”.
- The Lab – a sandbox where rule-breakers can test wild ideas without blowing up production.
All three of these are key for really successful usage. Leadership doesn’t know what inefficiencies lurk in the corner. The rank and file don’t know what the goals are or what they have a mandate to do. And the real aspirational ideas need staffinf.
b. Augment → Automate → Agentic
Anthropic’s AI Fluency course slices capability into three stages:
- Augmentation – human steers, AI boosts.
- Automation – repeatable tasks handed off to scripts or low-code glue.
- Agentic – goal-driven systems that operate (mostly) on autopilot.
Progression here is a staircase but every stair is critical. Build the right best practices on augmentation with vanilla prompting, context engineering and feedback loops and the rest fall in place.
3. Where The Two Grids Collide
Imagine a 3×3 board: columns are Bottom-Up, Top-Down, Lab; rows are Augment, Automate, Agentic. Each cell asks a different question:
| Bottom-Up | Top-Down | The Lab | |
|---|---|---|---|
| Augment | “How can any employee shave 30 mins today?” | “Which task must use AI by Q3?” | “Can we prototype a custom prompt library in a week?” |
| Automate | “Who’ll turn that prompt into a Zap?” | “What SLA can scripts own safely?” | “Let’s break something nightly and learn.” |
| Agentic | “What’s the smallest agent worth trusting?” | “What’s the risk budget for autonomy?” | “Spin up a proof-of-concept, measure hallucinations.” |
Plot current experiments, then look for empty squares—that’s usually where value hides.
4. Lessons From The Field
Quantity → Quality. Mess around. Always.
- Start Painfully Small
One high-ROI annoyance beats a grand “AI strategy deck” every time. The dopamine of a solved headache fuels evangelists. - Goldilocks Guidance
Too much freedom and folks freeze; too many rules and they hide. Offer just enough starter prompts, then get out of the way. - KPIs Need Owners, Not Observers
If a metric matters, someone’s bonus should twitch when it moves. Leadership can’t outsource accountability to “the AI committee”. - Integrate Into Existing Tools
Salesforce, Hubspot, Zapier, Intercom, Google Sheets—wherever people already live. New tabs == new friction. - Evangelists Trump Task Forces
Your loudest experimenter is worth three slide decks. Give them budget, stickers, and a stage. - Labs Should Ship, Not Just Shine
Prototypes that never graduate become museum pieces. Force a quarterly “leave the nest or get deleted” review. - Normalize Sharing The Bloopers
Every mis-labelled email, every hallucinated quote—show it. Transparency beats perfection and kills Secret Cyborg fear.
For more geeky war stories: check out Building Personal GPT or my rant on surviving the Post-AI Content tidal wave.
5. A (Humbly Offered) Roll-Out Checklist
- Pick one annoying use case—no moonshots.
- Map it on the 3×3 grid: where does it live today, where could it live next?
- Assign a KPI and an actual human owner.
- Document the prompt → script → agent evolution path.
- Schedule a five-minute demo; recorded is fine, memes encouraged.
- Rinse, share, repeat.
Will this guarantee success? Ha. If only. But it nudges momentum in the right direction—and momentum, not perfection, is what sticks.