How much AI should my team use? A framework for managers

Marco

Jun 03, 2026

In 2026 most knowledge work teams have settled into one of two failure modes. There are the Slop Cannons who paste ChatGPT walls into every Slack thread, every architectural decision, every customer call. And there are the holdouts who still write strategy memos and code completely by hand because they think the alternative makes you soft.

Both are wrong. The interesting question was never “how much AI?” It is where and when.

Is AI good or bad for my team?

It’s clear that AI can help you move fast: At Google, 75% of Google’s new code is now LLM generated before being reviewed by an engineer. The ICIJ’s Pandora Papers surfaced thousands of passports from huge amounts of leaked documents using machine learning. And in Kenya 10,000 health related questions by pregnant women are answered daily by SMS. Most of this would have been science fiction in 2022.

However, it is also clear that AI can be stupid enough to delete entire production databases, make people publish summer book reading lists with made up titles, create fake citations in papers or think that it’s ok to make up court cases in your legal arguments.

The “AI Bowtie” Framework

To explain to my 30-person team what I believe good AI usage actually looks like and when to use it, I regularly draw this picture for them on the whiteboard:

Reading it from left to right: The vertical thickness of the bowtie at any point is the amount of AI you should be using at that phase. Wide means lots of AI. Narrow means none. Five phases:

1. Research · Brainstorm · Discover

Lean on AI as hard as you can. Have it summarize twenty papers, generate dozens of competitor pitches, draft customer objections you haven’t thought of, role-play five different personas reacting to your idea. The bigger and weirder the funnel of inputs, the better the raw material later. What AI cannot do here: care which of the twenty options actually matters to you.

1. Synthesize · Prioritize · Distill

Dial AI down. You still use it to cluster and summarize. Then you take over. The model can stress-test your phrasing, flag contradictions, list counter-arguments. It cannot decide which idea is the one you are willing to bet six months on. The closer you get to a single concrete answer, the more it has to be yours.

1. THINK! No AI.

This is the brain-only zone, the soul of the work. The one paragraph you would write on a napkin if you had to explain what you are actually doing. The software architecture you can draw on a whiteboard from memory. The market positioning. The company strategy. The feature scope. The continue-or-pivot call.

If you outsource this to an AI model, you do not actually have a strategy. You have a plausible average of one, generated by something that has read everything and committed to nothing. You neither have a mental model in our head nor will you be able to defend it when challenged, extend it when the situation changes, or argue for it in a room of skeptics.

The soul of your work is what you discuss with co-workers, hone and commit to.

1. Plan · Prepare · Architect

AI comes back, but on a leash. Write the implementation plan in your own words first. Hand it to the model for critique: where are the gaps, what does it not cover, what would a critical reviewer say. Tighten it. Break it into parallel workstreams. The model is now a fast colleague who reads carefully, not a vending machine that gives you an answer.

1. Execute · Implement · Disseminate

Open the throttle all the way back up. Generate the boilerplate. Draft the docs. Spin up agents in parallel. Here is the part most people miss: spend the saved hours on the things that would normally get cut for time. Make the landing page accessible to screen readers. Write the documentation properly instead of leaving it as a todo. Add the integration tests you would normally skip. Localize. Polish the empty states. The whole point of holding the soul yourself is that you earn the right to be lavish here, and your output ends up better than the pre-AI version you could have shipped.

tl;dr: Explore with AI. Define the Soul of your work with your Head. Ship with AI.

Personal plug: Currently my team at Altium is working on making Agentic Requirements Engineering a reality (news coming soon). To make sure that we keep ourselves sharp in our thinking while accelerating our work, I came up with the above concept.