AI First
AI first doesn't mean AI before people. It means you redesign the work as if the capability had always been there.
AI first is a design rule, not a ranking. A new capability has arrived, so you go back to first principles and ask what the work would look like if you had it from the start. Not “where can we add AI to the process?” but “what process would we build today?” That question reaches the work you never did because it cost too much: the prototype you skipped on the way to a ticket, the test suite nobody had time for, the customer conversation that waited a quarter. The cheap version of AI first deletes people. The useful version deletes steps.
The API-first parallel
API first never meant “only build APIs.” Around 2002, Jeff Bezos sent Amazon a mandate that Steve Yegge later retold in his accidentally public “Platforms Rant.” Every team exposes its data and functionality through service interfaces. Teams talk to each other through those interfaces and nothing else: no direct reads of another team’s database, no back doors. Every interface is designed so it could be opened to outside developers. Anyone who doesn’t comply is fired.
The mandate said nothing about what to build. It changed how teams were allowed to talk, and that one constraint forced every team at Amazon to redesign how it worked. It was an Inverse Conway Maneuver at company scale: Conway’s Law says the architecture follows the communication, so Bezos changed the communication. Each team ended up with a hard version of what Team Topologies later called a Team API, and the architecture that came out the other side became AWS.
AI first works the same way. The capability is the constraint you design around. You don’t keep the old process and hand one of its steps to a model.
Bolted on versus rethought
Bolt AI onto today’s steps and you get a familiar result. A coding agent makes development faster. The pull requests pile up in front of reviewers who read at last year’s speed. The story still waits for refinement and still waits for a slot. Lead time barely moves, because development was never where the time went.
A value stream shows you this before you buy anything. The collection’s Software Feature Delivery map spends 5.5 of its 45 hours on work and the rest in queues. Make development five times faster and lead time drops from 45 hours to 42.6. Flow improves when you remove waits, and the way to remove a wait is to remove the step that creates it.
Idea to production, twice
Here is one order mapped twice: add CSV export to a reporting dashboard. Times are working hours. The first map is a good lean team without AI. They work test-first in small batches, ship through a pipeline, and hide unfinished work behind feature flags. Nobody would call this team slow.
The team does almost everything right, and the order still spends 100 of its 132.5 hours waiting. Two of the waits sit in front of the build: three days for refinement, four for a WIP slot. Both exist for good reasons. Building is expensive, so you refine before you commit a pair of developers, and you limit work in progress so the pair finishes before it starts something new. The last wait is a week for usage data, because the first customer to see the feature sees it in production.
Now redesign it with a coding agent in the room. This is not the same map with faster steps.
The product manager and an engineer sit with a customer and build a working prototype with a coding agent in two hours. The session becomes the spec: the engineer turns what the customer accepted into evals and acceptance tests, and the agent builds against them while the engineer watches. A senior developer reviews the intent along with the diff. The feature ships behind a flag to the customers who were in the room, and they already know what to look for.
| Streamlined | AI first | |
|---|---|---|
| Steps | 11 | 8 |
| Waits | 4 | 2 |
| Process time | 32.5 h | 8.5 h |
| Lead time | 132.5 h | 26.5 h |
| PCE | ~25% | ~32% |
Lead time drops by 80%, and process cycle efficiency barely moves. If the agent had only sped up the build, the waits would have stayed put and PCE would have fallen. Instead the waits shrank as fast as the work. The refinement queue and the WIP queue are gone, 56 hours between them. The agent’s typing speed didn’t remove them. The prototype did, because it replaced the ticket. You don’t refine a story the customer has already used, and a three-hour build doesn’t sit four days waiting for a slot. The usage wait shrinks from 40 hours to 16 because the first users are the people who asked for the feature.
Look at the work that appeared, too. Prototyping with a customer before every story used to cost too much, so the lean team never did it. Now it costs two hours, and it is the step that deletes the others.
The biggest wait left is still learning from customers. AI first doesn’t remove that wait. It points everything else at it.
Why we use it
We bring up AI first when a leadership team asks where to put AI, because that question already assumes the current process. In Effective AI, when handoffs are the bottleneck, the cohort maps the value stream before it decides where AI goes, and we ask what the map would look like if they drew it today. The interesting difference between today’s map and that one is never a faster step. It is the steps that are gone, and the work that shows up for the first time because it finally got cheap: the prototype, the eval suite, the usage summary nobody had time to write.
If your AI plan leaves every step in your value stream where it was, you bought a faster typist and kept the queue.
Resources
- Steve Yegge, “Stevey’s Google Platforms Rant” (Google+, 2011) — the Bezos mandate as Yegge told it, and what it did to Amazon
- Value Stream Mapping — map the stream before you decide where AI goes
- Value Stream — why the queues, not the steps, set your lead time
- Team API — the modern name for what the mandate forced on every team
- Compounding Engineering — the AI setups worth building once the steps change
- Loop Engineering — designing the agent’s cycle inside the new steps
- AI Without ROI — what bolted-on AI looks like from the budget side