Before asking for more headcount and resources, teams must demonstrate why they cannot accomplish their goals using AI, explicitly showing what their area would look like if autonomous AI agents were integrated as part of the team.
— Tobi Lütke, CEO of Shopify
Tobi’s memo resonated widely because organizations now recognize what it truly means to integrate AI – not merely as an add-on but fundamentally into their operating structures.
At Exponential View, we’d embraced precisely this mindset to reimagine how we work. We consider ourselves AI-native, which means:
-
We use AI reflexively; it’s a core skill for everyone.
-
We’re tool-agnostic, continually evaluating and updating our stack.
-
We build new workflows from scratch rather than just automating old ones.
-
Our team is becoming increasingly technical – the baseline expectation for coding skills has risen significantly.
-
At the same time, I’ve encouraged my portfolio companies to scale through synthetic intelligence during this time of transformation.
In recent weeks, we sprinted to prototype new workflows using LLMs, automation tools and structured systems thinking. Today, I’ll share our most valuable lessons.
At the end, you’ll get access to something unique: our internal stack of 40+ tools – everything we’re actively using, testing or intend to test – to help you decide what tools might work for you.
Let’s jump into our seven lessons.
We’ve started applying a simple heuristic: if a task, or even a question, comes up five times a month, it’s a candidate for automation. This “5x rule” helps spot patterns hiding in plain sight and forces you to think in systems rather than routines. This habit sets the expectation that workflows should evolve constantly, not calcify.
Of course, we now ask the question “what do we do five times a month” more than five times a month, making that a candidate for automation.
One of my (Azeem’s) favourites is a simple workflow which does my expenses. I have to contend with dozens of invoices a month and my automations, relying on Dext and Gmail filters, are good but not great. Expense reconciliation has involved a lot of time in Gmail. My new expenses agent eliminates that repetition: it pulls out invoices, bills and receipts in my emails and puts them into a correctly structured spreadsheet. It also makes a PDF copy of the bill and sticks it into Google Drive. This saves me my least favourite hour every month.
If the bill is a plane, train or hotel booking it also dumps them into a different document. A separate agent reviews that document and turns it into a chronological, structured travel briefing which I use. With fifty travel days across ten trips to the end of June this is an enormous time saver – perhaps a dozen back-and-forths with Gmail has been replaced by the occasional check of this summary document.
One early and useful lesson in building with AI was to break down our workflows into smaller, autonomous components, rather than trying to automate an entire process in one go. This modular approach makes it easier to test individual pieces, troubleshoot in isolation and evolve parts of the system without destabilizing the whole.
This approach draws inspiration from classic software architecture: encapsulation and separation of concerns. But it also reflects how AI-native workflows behave. When you have an LLM doing part of the work, you want its task to be as narrow and unambiguous as possible. Broad instructions like “write a summary of the latest AI developments” often result in generic or unusable output. In contrast, narrower prompts like “list three key recent breakthroughs in battery technology and explain their relevance to electric vehicle adoption” yield precise answers and clearer points of failure, making them easier to improve iteratively.
One modular workflow we’ve found particularly valuable automates the discovery and initial research for potential partnerships. The first module scans our broader ecosystem, discovering companies actively engaging in areas aligned with our priorities. The second module enriches these initial leads, pinpointing key decision-makers and compiling relevant context from publicly available information. Finally, a third module– acting as our digital comms specialist – helps create outreach to get across as clearly as possible. The result is a process that frees our team to focus on building relationships rather than hunting down details.
A modular system also helps teams think like system designers. If Module A breaks, we know not to debug Module C. That clarity saves time. It also supports scale: when each unit functions independently, it’s easier to assign ownership, train interns, or plug in new AI tools.
Some of us use Claude + Mermaid chart to sketch out projects and their modules for a visual work plan.
Treat the LLM as the foreman, not the worker. That is, use the model to structure the task, but don’t ask it to do everything. Once the model has identified what tasks need to be carried out, you can decide whether a given task is deterministic (in which case you may need to farm it out to traditional software code) or requires more judgement (in which case an LLM might be able to handle it.)
