AI‑SDLC · Two Lifecycles
Side by side/Two lifecycles, one product

The AI writes the code either way. The lifecycle decides whether it ever ships.

Lane A An AI coding tool, driven by you. Stalls at 77%.
Lane B The same AI, run as a pipeline. Ships on week 10.

Same idea, same models, same week one. Watch where they split.

AI‑SDLC vs an AI coding tool alone
same scope, same product
~45%
less money
~72%
less engineering time
1.8×
product per dollar
See the model →

// our cost analysis of a real ~1,800-story-point build · illustrative model, inputs tunable — not a measured guarantee

01/The race

14 weeks, one shared clock, two lifecycles.

Both lines plot how much shippable product exists — not how much code got written.

Lane A · the sawtooth
Every drop is finished work coming undone.
why it flattens at 77%
A regression nobody caught, an architecture the model quietly rewrote, an edge case found by a user. The climbs are real; so are the losses. Net progress stops moving.
Lane B · the staircase
Slower for two weeks. Never gives ground back.
what the first two weeks buy
A frozen spec and a design — no code yet, and it looks slower. That's the trade. After that nothing regresses, because nothing passes a gate until it's done. Bugs still happen; the pipeline catches them, not you.
The point
It's not better AI. It's the same AI.
then what wins
A lifecycle around the model — analysis, design, independent review, automated tests, gates. The jobs a chat window was never doing.

// Illustrative timeline, drawn from our cost-and-effort model and observed failure patterns — not a measured guarantee. Your mileage varies with scope and domain.

02/Where it stalls

Six places the solo AI build gets stuck.

Not one of them is the model being bad at code. Each is a job a lifecycle does and a chat window doesn't.

Entrepreneur + AI, alone
AI‑SDLC with agents
STUCK 01 · no spec
The requirements live in your head
detail
You prompt what you remember to prompt; the model fills the gaps with plausible guesses. Nobody can say what "done" means, so nothing gets finished — only abandoned.
PHASE 0a · feature analysis
"Done" is written down before "start"
detail
Product, business analyst, domain expert, market, legal, architect and founder agents each pass over the idea and produce one frozen analysis with an explicit recommendation — before a line is written.
STUCK 02 · regression
Feature 3 quietly breaks feature 1
detail
Nothing tells you. You find out weeks later, by accident, in front of someone who matters — then fix it and break something else. That whack‑a‑mole eats the back half of the calendar.
PHASE 6 · feature testing
A break fails a test the same day
detail
The QA agent writes the tests, then runs them forever. Coverage grows with the code instead of lagging a year behind it — so the context is still fresh and the fix is cheap.
STUCK 03 · amnesia
The model forgets the architecture
detail
The context window fills, the session resets, and the same AI invents a second way to do what it already does. Three months in: four patterns for one problem, and no idea which is load‑bearing.
PHASE 2 + 5 · design & doc-sync
Memory is a file, not a context window
detail
The implementation design is agreed and written down before build, and docs re‑sync to the code every cycle. The architecture lives in the repo — the next agent reads it instead of guessing.
STUCK 04 · no reviewer
The model marks its own homework
detail
You're not equipped to disagree. It looks right, it runs, you ship it. The security hole and the money‑path edge case get found later, by someone else, for free.
PHASE 4 · dev review + gates
A different agent reviews — a human signs
detail
Code reviewer, architect, security reviewer and specialists check work they didn't write. Nothing advances until a senior engineer clears the gate. Judgment stays human; the volume doesn't.
STUCK 05 · unhandoffable
Nobody can pick it up — including the AI
detail
No docs, no tests, no decision trail. Hiring a real engineer now means paying them to reverse‑engineer your own product first. The "free" build bills you at the worst possible moment.
PHASE 5 + 8 · doc-sync, release check
You get the paper trail, not just the code
detail
Source, tests, docs and the reasoning behind every decision — yours to own. A new engineer (or a new agent) reads the repo and starts on day one instead of month two.
STUCK 06 · the wall
A demo that will not become a product
detail
All five compound. The build is perpetually "almost there" and has been for two months. This is where most solo AI products quietly stop — not at the idea, not at the demo, but at the last mile nobody scoped.
PHASES 6 ⇄ 7 · the contained loop
The last mile is a phase, not a surprise
detail
Test → triage → fix → re‑test loops until zero actionable defects remain, then a release check clears it. The messy part of software is planned for and bounded — which is exactly why the line gets crossed.
03/The two loops

Four steps. Or ten phases.

The whole difference, flat: not a better model — more stations, and a checkpoint between every one.

Lane A · the chat loop
Prompt → paste → click → hope
1
Describe the feature from memory
No written spec. No second opinion.
2
The AI writes the code
Fast, confident, and unreviewed.
3
You click around to see if it works
This is your entire QA department.
4
Something else broke → back to step 1
The loop that never converges.
Missing entirely
Analysis · design · independent review · automated tests · docs · release check · anyone to sign off
Lane B · the pipeline
Ten phases. A gate between every one.
0a · analysis
founder · product · BA · domain · market · legal · architect
0 · task creation
architect · teamlead · backend · frontend · BA
1 · UX review
ux · frontend · BA
2 · implementation design
teamlead · architect · backend · frontend
3 · implementation
build agents, in parallel, around the clock
4 · dev review
code reviewer · architect · security · tech-debt sweep
5 · doc sync
docs re-aligned to the code that shipped
6 ⇄ 7 · test / triage / fix
loops until zero actionable defects remain
8 · release check
PM · devops · BA · docs — ship or don't
9 · harvest
what the build learned goes back into the backlog
Between every phase
A gate a senior engineer has to clear.
Machines do the volume. A human owns the judgment.
04/What the difference costs

The sawtooth has a price tag.

One real build — a regulated trading platform, ~1,800 story points, about a year of conventional engineering — costed both ways.

1.8×
product delivered per dollar, vs AI‑assisted coding alone.
~45%
less money spent
~72%
less engineering time paid for
The gap isn't the coding.
It's the months of manual testing and review the pipeline automates away. Lane A pays for those in people.
Modeled total cost
same scope
$141k
$78k ↓ ~45%
AI coding only · Lane A
AI‑SDLC · Lane B
// illustrative model, inputs tunable — not a measured guarantee
Modeled engineering time
same scope
277d
≈ 13 months of work
76d ↓ ~72%
≈ 3.5 months of work
AI coding only · Lane A
AI‑SDLC · Lane B
// person-days of paid engineering & QA time, from the same model

And that's the generous reading of Lane A — it assumes the solo build eventually gets there. See the full case study →

Not just a model — already running
A second build went through the same pipeline, in regulated healthcare.
what came out of it
A live client project in a regulated medical domain, run end to end through the process by a delivery team — not a to‑do‑app demo. Real code, real test suites, and documentation the team reads. Client and product withheld.
~1,800
code files shipped
177
automated test suites
7+
impact‑analysis & audit cycles
live
API docs, glossary, QA runbook
05/The honest bit

Lane A isn't wrong. It's just a different job.

Need something clickable in front of five people this week? Open the chat window and go — you don't need us for that.

Where the lanes actually diverge

A prototype has to work once, for you. A product has to keep working, for strangers, while you change it.

Only one of those is solved by writing code faster. The wall arrives the moment it has to be real — real users, real money, someone else maintaining it.

Week eight of Lane A isn't a personal failure. It's a missing lifecycle — and it's fixable.

06/Switch lanes

Bring your build. We'll tell you which lane you're in.

Fifteen minutes, no slides. From a blank idea or from a stalled 80% build — same process either way.

  • ✓ A concrete plan for your build.
  • ✓ A timeline, a budget range, and an honest read on feasibility.
  • ✓ A fixed‑scope first sprint — real output before any big commitment.
DS
Dmitry Shalin
Founder & engineer — you'll be talking to me on the call.
LinkedIn ↗
Book a 15‑minute call →
15‑minute intro call
UTC
// connect your Google booking page — see src/README.md
MON
TUE
WED
THU
10:00
14:30
09:15
11:45
Confirm a time →