Mike Litman
They Are Cooking
A TALK BY MIKE LITMAN

They are cooking.

Why some companies are shipping at 10x velocity – and what the rest are doing about it.

The question

"How is AI impacting your velocity?"

Your board has been asking this for over a year. Most exec teams – regardless of industry, size, or AI budget – are giving some version of the same answer. And in the time they've been crafting that answer, the companies moving fastest have stopped caring about the question entirely. They're just shipping.

The honest answer
What most execs say
"We're piloting Copilot."
"PRs are up 20–30%."
"We have an AI task force."
"We're evaluating options."
What's actually happening
Agents shipping hundreds of PRs per week
Cycle time down from weeks to hours
Internal tools built by EAs and HR teams
CFOs backing token spend like infrastructure
The gap
10x

The velocity gap between the companies that are cooking and the companies that are piloting. Not 20%. Not 2x. A full order of magnitude. And it's widening every month.

This is about you

If your honest answer to the board is "we're piloting this" – you're in the bell curve.

Not failing. Not slow. But not cooking either. And the gap between you and the companies that are is not a technology problem. It's a decision problem. The tools exist. The models exist. The question is whether you've actually decided to change.

05
Why the bell curve is so crowded

Smart people have been in pilot mode for 18 months. Here's why.

Middle management protecting headcount. Every AI efficiency gain is a threat to someone's team size and status. The people who need to implement change are often the ones with the most to lose from it.

Procurement slower than model releases. By the time a security review completes on a tool evaluated six months ago, three better models have shipped. The enterprise buying motion is structurally incompatible with the pace of AI development.

Fear of being the cautionary tale. The incentive structure punishes visible failure harder than invisible mediocrity. CEOs who move fast risk becoming the next Amazon-lost-6M-orders story. CEOs who move slowly just stay in the bell curve – quietly, anonymously.

Measuring the wrong things. When you only track inputs – seats activated, training hours, adoption rates – you can never prove it's working. Without evidence, you can't escalate. Without escalation, nothing changes.

05
Not the companies you expect

It's not just the AI-native darlings.

Shopify: 14,000+ employees. In April 2025, CEO Tobi Lütke posted his internal memo publicly: "Reflexive AI usage is now a baseline expectation." Before requesting new headcount, every team must prove the work cannot be done with AI. That's not a pilot. That's a restructure.

Block: 10,000 employees down to 6,000 in February 2026 – 40% cut, explicitly citing AI. They built an internal coding agent called Goose on Anthropic's MCP. One engineer reported 90% of his code output now comes from it. Production code shipped per engineer up 40%+ since September 2025. Jack Dorsey: this is what "AI-first" actually looks like.

Revolut: grew its customer base 34% in 2024 while limiting customer support headcount growth to just 5%. The gap between those two numbers is AI. That's not an efficiency saving. That's a business model restructured.

The Pattern: a daily culture intelligence brief scanning 150+ feeds. What used to require a team of researchers now runs on a cron job. No editorial team. No headcount. That's a department that no longer needs to exist.

The visibility problem

You can't see the gap if you're measuring the wrong things.

Vanity metrics (what laggards track)
Copilot seats activated
PRs up 20–30%
Hours of AI training completed
Number of AI experiments running
Signal metrics (what winners track)
PR per R&D head, per week
Cycle time: idea to shipped feature
Feature velocity per engineer
Token spend as % of build cost

DORA benchmark: elite performers ship in under 1 hour. Industry average: 1–4 weeks. Your cycle time tells you which camp you’re in.

What separates them

Nine things the fast companies have in common.

01
Top-down edicts, not bottom-up experiments
Not "feel free to try AI." Mandatory. Shopify-style: prove why a human is needed before you get one.
SHOPIFY
02
Token budgets that would scare your CFO
Not a £10K pilot. Six-figure annual spend per team, treated as infrastructure, not experiment.
BLOCK
03
Public leaderboards. Unapologetically.
Who's shipping the most? Whose cycle time is shortest? Posted internally. Made competitive.
STRIPE
04
At least a dozen deep AI obsessives
Not an "AI team." Twelve people embedded across every function who live and breathe the frontier.
BLOCK
05
Non-technical teams building real tools
EAs building internal platforms. HR running custom ATS. Finance building their own dashboards.
SHOPIFY
What separates them — continued

The harder half of the list.

06
Investment in developer experience, not just AI tools
They didn't just buy Copilot seats. They rebuilt tooling, CI pipelines, and local dev environments to be AI-legible. The infrastructure has to match the ambition.
BLOCK
07
No sentimentality
The 15-year engineer who reviews every line as if they wrote it. The process from 2018. The team structure built for a different era. The companies winning have no emotional attachment to any of it. This is the hardest one to say out loud. Block eliminated a middle management layer when Goose shipped. They didn't renegotiate it. They just stopped hiring for it.
BLOCK
08
Skill – real AI literacy, not tool familiarity
Not knowing how to prompt. Knowing what models can and cannot do. Building intuition for when to trust outputs. This spreads through a team like any craft – but only if someone skilled enough is teaching it.
THE PATTERN
09
Will – the decision, made and communicated
You can mandate licences. You can hire an AI lead. But will is different. It's the executive decision to actually change – not manage the appearance of change. It comes from the top, stated plainly, and it makes every other item on this list possible.
SHOPIFY
08
The thing nobody says out loud

The bottleneck is not the tools.

Every developer in your organisation already has Copilot. The bottleneck is psychological permission to truly delegate. To accept a PR you did not write line by line. To let an agent make an architecture decision. To trust the output, not just the suggestion.

The companies moving fastest have given that permission explicitly, publicly, and from the top. The middle manager who reviews every AI-generated line as if they wrote it themselves is not a quality gate. They are the bottleneck.

Psychological permission Not a tool problem The real blocker
09
The second thing nobody says out loud

You can't bolt AI onto a two-week sprint.

When cycle time drops from weeks to hours, every downstream process breaks. Code review. Stand-ups. QA. Product prioritisation. Sprint planning.

The companies winning did not add AI to their existing process. They rebuilt the process around what AI can actually do. That's a fundamentally different project than buying a seat licence.

Amazon mandated 80% usage of their internal AI coding tool Kiro with adoption tracked as a corporate OKR. They didn't update the safety and review infrastructure to match. The result: 6.3 million lost orders when AI-generated code caused cascading failures. They now require senior approval for every AI-assisted code change. Speed without process redesign is just a faster way to break things.

The shift happens at the individual level first. Early on, you review every Claude output line by line. Then the volume makes that impossible. At some point you stop – not because the quality dropped, but because checking everything became the bottleneck. You move from line-by-line review to spot-checking outputs, trusting the system, intervening on exceptions. That is the process redesign. It's psychological before it's organisational. And every company that's actually cooking has made it.

How to cross the gap

Five moves. In this order.

01
Find your twelve obsessives now
They're already in your organisation, doing this without being asked. Not an AI team – people embedded across every function who live on the frontier. Find them, name them publicly, give them air cover. Everything else depends on this.
02
Give them a genuinely scary budget
If no one winced when you named the number, it wasn't scary enough. Track ROI on token spend the same way you track any infrastructure cost. Publish it internally. This is how you signal you're serious.
03
Start with internal tools, not customer-facing products
Block's Goose started as an internal coding agent before it was handling 90% of engineering output. Permission to experiment is lowest here. Failure is invisible. Wins compound fast.
04
Measure cycle time, not activity
Track the elapsed time from "idea" to "shipped and live." Not PRs merged. Not story points. If you don't know this number today, finding it is your first task. You can't see the gap without it.
05
Write the memo
Tobi Lütke didn't say "try AI." He said prove why a human is needed before you hire one. Write your version. Say it out loud. Post it. That single act changes more culture than any training programme.

"The window is open. But it closes not when the technology changes – when the talent, the benchmarks, and the client expectations permanently reset."

The companies that started in 2024 have had six quarters of compounding. Their cycle times are shorter. Their engineers are faster. Their codebases are more AI-legible. Their non-technical teams have already built and shipped internal tools. Every quarter you wait, catching up gets structurally harder – not because the technology changes, but because their lead becomes embedded in the org.

The window is open. The question is whether you've decided to use it.

A Talk by Mike Litman

Now go cook.

One question to take back: what did your team ship this week that AI wrote? If you can't answer it, you know where you are.

mikelitman.me · hello@mikelitman.me

Continue from slide ?