2025 AI Game-Plan For Engineering Managers
In the chaos of the AI flood, what do you do?
Can you do me a favor and share it on HN if you think it is worth a read and to discuss?
AI can be overwhelming for Engineering Managers like #CTOs on one hand, and developers and engineers on the other. When a lot of things happen, when you’re in the eye of a hurricane, everything flying around, it’s not clear what is happening and what direction you should move to.
I see three phases of AI adoption in software development:
AI Autocomplete: Do now
AI-as-Compiler: Prepare for
AI is not software: Think about
Let’s take a deeper look into these.
AI Autocomplete
DO NOW
When: Now, start immediately
Goal NOW: Get everybody on board, get everyone in technology aligned around AI and AI usage, learn how best to use AI for your product and your company
What: Using auto complete in Cursor and other IDEs and using agents like Claude Code to implement features, fix bugs, explain code.
How:
Hackathons - use Hackathons to work together on your code with AI. Goal is to work together and see how AI can be used cooperatively. Get the team behind AI.
AI Friday - Every (second) Friday developers can do with AI whatever they want. Play around with new tools, try crazy things. Goal is to get better at using AI and find technical innovation around AI.
1 Prototype / Day - Think bigger with AI. In the past prototypes have been expensive, it often took a developer a sprint to create a prototype. The goal of a prototype is to get something into the hand of others, the C-suite, product managers to play around and decide if the feature/product makes sense and has potential. Here you can experiment with agents and create a prototype in half a day. Every day another developer is creating a prototype, you’ll end up with 5 / week. Plenty of opportunity for the company to grow.
1 MVP / Week - With all those prototypes (5) over the week, one of these might be a potential MVP. Create the MVP with AI from the prototype (or new) and test it with potential users and customers. The idea of an MVP is to find something that users use and customers buy. Again, possible with AI and was too costly before.
AI (Bi-)Weekly - 30min exchange of what is going on in AI and what people have seen with AI. Show around AI use-cases people found on the net. Developers decide on what is important based on what time is spent on. Instead of talking about AI, invest time on AI topics so everyone see this is not something to wait out.
Minimum: Everyone uses Cursor/Windsurf
Think about what has been a bottleneck before and with AI no longer is. There is where you need to change to get the most out of AI and a competitive advantage.
AI-as-Compiler
PREPARE FOR
When: ~2-5 years
AI will be used as compilers. It is easier to one-shot modules and parts of apps than cooperate on that code with different agents and different people. The problem of vibe coding is coordination and understanding what others have done.
Vibe coding still needs “a coder” - AI-as-compiler doesn’t
Goal NOW: Understand what you need to change in your team, process, architecture and code to make AI-as-compiler work.
What: AI as compiler works on requirements like text use cases, data descriptions, DB schemas, designs, (REST) API descriptions and fake screenshots as input. In one go an application or module is generated.
How: Change your code base to modules that can be one-shotted from requirements.
Reorganizing everything for this phase prepares you for “AI is not Software”
AI is not Software
THINK ABOUT
When: 5-10 years
Goal NOW: Think now how you can replace your software with AI and adopt your vision and strategy
What: AI is not software - in the sense of being planned and build. AI is a trained blob of weights. Instead of CRM software, there is an AI which you tell “Send an email to all customers”. There is no CRM code in the AI, so there is no CRM software. Same for reporting, sales, ads, tickets, holiday planner, and everything else. See, no software, just a trained AI.
When you’re software is split up, from AI-as-compiler, it is easy to replace parts with a trained AI.
We’re quite some time away from this, I recently wanted to play Tic-Tak-Toe with ChatGPT and it failed. But I could probably train it do play this very simple game. That’s where we are.
