"AI will 3x your development speed" — everyone believed it and adopted it.
So why are project deadlines still slipping?
Everyone believes this
The belief isn't wrong, exactly. In GitHub's Copilot research, developers completed isolated coding tasks — like writing an HTTP server — 55% faster on average. Tasks that took 2 hours 41 minutes without Copilot dropped to 1 hour 11 minutes.
The natural conclusion: "Let's use AI to compress our 70-day project down to 35 days." AI compresses the longest phase in the timeline. Makes sense, right?
But the data points the other way
Google's DORA team published something interesting in 2024. While AI adoption raised individual developer productivity and satisfaction, software delivery stability and throughput actually declined in some cases. Individuals got faster, yet team-level reliability dropped. That's a paradox.
The Theory of Constraints explains it. The core principle: the overall performance of any system is determined by its single slowest step. Speeding up anything that isn't the bottleneck doesn't change total throughput — it just creates a pile of unfinished work before the actual bottleneck.
So here's the real question: "Is 'coding' actually the longest step in your project?"
Misidentifying the bottleneck
Even if AI cuts coding to near-zero, if requirements clarification takes 6 weeks, your project won't finish in under 6 weeks. That's Theory of Constraints in practice.
What Toyota knew all along
There's a principle that cuts right to this problem: "Bottlenecks must receive predictable, high-quality inputs." It runs through Eliyahu Goldratt's "The Goal" and Toyota's production system alike.
Here's a real-world example. A ticket arrives: "Send an email to the user once the sale is completed". Before a single line of code is written, you need answers to:
- What's in the email? Order number? Expected delivery date?
- What counts as "completed"? Payment authorized? Inventory deducted? Shipping started?
- What happens on failure? Retry? Alert someone?
- Is the order still valid if the email doesn't send?
Without these answers, developers — and AI — must guess or stop to ask. That 55% coding speed boost means nothing if you're stuck here.
Martin Fowler's "Specification by Example" makes the same case: requirements aren't complete until expressed as concrete examples (input → expected output). Same rule applies to AI — vague inputs make AI quickly build the wrong thing.
| AI alone | Clear requirements + AI | |
|---|---|---|
| Input quality | Still vague | Concrete examples |
| Coding speed | Faster | Faster + less rework |
| Delivery stability | At risk | Maintained or better |
| Actual deadlines | Unchanged or delayed | Can actually shrink |
So where do you start?
- Find the real bottleneck
Honestly track time spent on each phase of a recent project — and why it took that long. Requirements changes? Rework? Waiting for approvals? That's your bottleneck. - Define "done" before writing a line of code
For each feature, write concrete examples: "When user A completes payment → an order confirmation email in format B is sent." This is the quality input that works for both developers and AI. - Move uncertainty earlier
If "how do we handle this?" comes up during coding, you're already late. One hour with your PM/PO prevents two days of rework. - Then add AI
Once requirements reach the concrete-example level, hand the implementation to AI. Speed gains only matter when the input is clear. - Measure what actually matters
Instead of "did we go faster?", ask "did rework go down?" and "did requirements changes drop?" DORA's deployment frequency, lead time, and change failure rate are solid benchmarks.
Want to go deeper?
I don't think AI will make your processes go faster The original article by Frederick Vanbrabant — the seed for this piece. Concrete Gantt chart examples and practical constraint theory application. frederickvanbrabant.com
2024 DORA Accelerate State of DevOps Report Google's annual DevOps research report analyzing AI's impact on development teams with team performance data. dora.dev
Theory of Constraints — Wikipedia Overview of Eliyahu Goldratt's constraint theory — bottleneck management and the Five Focusing Steps. wikipedia.org
SpecificationByExample — Martin Fowler The case for concrete example-based requirements. Why abstract specs fall short and what to do about it. martinfowler.com
Research: Quantifying GitHub Copilot's Impact on Developer Productivity GitHub's official research — the exact experimental conditions behind that 55% figure and its limitations. github.blog
