To add
Before hiring a person or buying a tool, we have to answer the question first: can AI solve this? Hiring or buying should stop being the default answer to a capacity problem. It becomes the fallback only after AI has been genuinely ruled out.
Source: Shopify. Tobi Lütke’s internal memo told teams they had to show why they could not get something done with AI before asking for more headcount or resources. Where I got the idea: https://public.hey.com/p/kwe5Vzv4cpBWDp3fs4zbpQNX
IT policies - Raise the ceiling
Make it as easy as possible to use AI. AI models are useless without eyes (data/context) and hands (tools.) You unlock AI gains by making data and tools easy to integrate and iterate with.
Cut out the layers of red tape and bureaucracy, and optimize for giving access to data connectors and approved budgets for tokens, whenever you can.
Someone who’s bought in is going to accomplish five to 10 times more work than someone who hasn’t seen the magic yet.
Promote, hire, and retain AI doers
Show that people who use AI aggressively (at or approaching tier 3) get promoted first or interface the most with senior management.
When their colleagues see that person advancing in their career, that’s a more effective motivator than any mandate.
Nominate people who are already AI-forward as internal cheerleaders. It gets other people to come out of the woodwork rather than hiding their AI usage by making it clear that using AI is encouraged.
Teach through building
One lesson we learned from doing Marq AI trainings is that the most impactful way to spend time is on building.
Workshops teaching people how to use AI in a hands-on way are an effective way to teach your team but they need to be heavy on building tools.
You learn AI by doing. It’s a mix of:
- Getting a feel for what AI models do well
- Where AI models have shortcomings
- Picking up management skills (more on that below)
Train everyone to be managers of agents
Advancing beyond prompt engineering requires applied judgement (i.e. management skills.)
Individuals must know where agentic flows fit in their workflow and how to know when the output is trustworthy.
This requires much of what good managers already know. They must know how to:
- Decompose projects into tasks.
- Direct what to start on first.
- Specify what good looks like.
- Give constructive, iterative feedback.
Most individuals struggle because they don’t have management training. They’re not used to context switching, setting up systems and rules, or evaluating whether something that they haven’t worked on themselves is any good.
The script is flipping: Managers are also becoming individual contributors, because managing a team of agents is often easier than managing human teams. Sometimes it’s easier as a manager to vibe code a task using Claude Code than it is to brief a human, wait for them to send it to their own Claude instance, and get a response in a couple of days.
You need to help people understand context switching, how to do evals, develop good taste for deciding what to work on, and train AI in specific skills.
How do you systematically write a good PowerPoint skill or a good daily update report skill? That’s the work now.
OKRs - Assign impossible tasks
Boris Cherny, a creator of Claude Code has said that you should slightly under-resource most teams, which makes employees think, “The only way I can do this is if I use AI.”
Strategically choose the tasks so that they can’t possibly be done without AI but give the runway and focus to accomplish it.
For example, if your goal is to write one blog post a week, you can likely do that manually.
Don’t set the goal as, “Starting today, you have to produce one piece a day.” Instead, say: “Our goal is to work up to producing one piece a day. What needs to happen for you to make progress toward that goal?”
It might take time, but if they know that’s where they’re heading rather than where they’re starting, they’ll start thinking strategically about how to use AI to save time, and start experimenting.
Buy the model direct, not third-party tools
When you evaluate AI-powered tools, you’re also—whether you realize it or not—evaluating the tool vendor’s choices and constraints, rather than what the underlying model provider (like Anthropic, Google, or OpenAI) is capable of. It’s often faster to build your own Claude/Gemini/Codex skill with your own rules and preferences already built in.
Companies are increasingly building, not buying, AI software on top of models, because it gives you flexibility.
It also teaches the fundamental skills of how to use AI.
It’s far more cost effective too because it’s difficult to compete with Anthropic offering 200 subscription.
Third-party tools tend to be less flexible, less cutting-edge, and more expensive. That’s not always the case, but as a general rule, it holds.