WorkSkillsRecognitionAboutContactRésumé ↓
Case Study 05 · Research Synthesis

Where the bottleneck moved

Four senior engineers on where the slow step went once AI took over the coding, and why they disagreed.

Status
Live
Type
Research synthesis
Method
4 interviews, ~45 min each

What everyone agreed on

Everyone is asking whether AI makes engineers faster. It is the wrong question, and it tells you nothing useful. I asked four senior engineers a narrower one: if writing code is no longer the slow part, what is? They agreed completely on one thing, coding stopped being the bottleneck, and then pointed in four different directions about what replaced it. That disagreement turned out to be the most interesting thing they said.

The numbers they reached for, unprompted, landed close: roughly 90% AI and 10% manual, almost zero manual for another, 70% coding down to 40%, and work that used to take a week now done in two days. The thing that used to eat the day, writing code by hand, mostly does not anymore.

Four people, four answers

I asked each the same follow-up: if coding is fast now, what is the slow step? Here is where they walked in different directions.

Abhiraj
Lead Backend Developer
WWEXprev · Razorpay
The delay is your customer interaction, and PM to engineer interaction.
Pointed upstream to alignment on what to build, and downstream to infrastructure deployment SOPs that are not automated.
Manish
Full-stack Engineer
Uber
The testing part is the bottleneck right now. I go testing manually, to be sure before production.
The AI writes it in minutes; making sure it is actually right is what takes the time now.
Harish
Data Scientist
FFreshworks
An AI-written prompt lands at not more than 40% accuracy; refining it to 95 to 99% is the human part.
Pointed at judgment, the last stretch from passable to correct that does not automate away.
Pabitra
Frontend Engineer
Target
What is slowest is creating the task. It always comes from the product managers.
Pointed furthest upstream, at the product function itself: when the roadmap thins out, the team runs out of defined work.
The disagreement is not noise. It is a map of where each person sits in the system. Each identified the bottleneck nearest to them.

What this means for product

Two of the four bottlenecks sit inside the product function: Pabitra’s creating the task and Abhiraj’s alignment on what to build. That is not a mandate, two of four is not everything. But the direction is consistent with what the whole conversation kept gesturing at: as AI compresses the act of building, the expensive part drifts upstream, toward deciding what is worth building and exercising the judgment to know when the output is actually right.

A note on method

Four conversations are not a study. This is a set of careful anecdotes, recorded with permission, each engineer asked the same core questions in the same order. And since the irony is the point: I did not synthesize these transcripts by hand. I built a tool to do it, one that keeps every claim tied to the exact quote that supports it. That tool is Receipts, and it is its own case study.

Next case study
GroundTruth