Back office ops · Production

How AI changes the role of applied scientists at Instacart's Economics Team

The problem

The Economics Team's portfolio was concentrated in standardized Core ML tasks that occupied 73% of PRs, while entire task categories — especially frontend development — were inaccessible due to skill gaps and the coordination costs of delegating to specialist engineering teams.

First attempt

The team's Causal Inference Platform (CIP), a UI-based tool giving company-wide access to standard causal inference methods, was undermined by AI coding assistants that made bespoke causal inference faster and more flexible than any constrained UI could offer.

Workflow diagram · grounded in source
1
AI tool adoption waves
ai_action
“We see a first small increase in the second half of 2023, following the release of new AI-powered productivity tools like Ava and their growing popularity within the team. A much more pronounced jump shows up in the first half of 2025, c…”
2
LLM PR classification
ai_action
“we used an LLM to classify every pull request made by the team between 2023 and 2025. The classifier assigned each PR to one of eight categories (ML/Model Dev, Data Pipelines, Platform/Tooling, Analysis/Research, Infra/DevOps, Frontend/U…”
3
Task redistribution
routing
“The share of Platform/Tooling more than doubled from ~5.2% to ~12.1% of PRs per person and Frontend/UI emerged as an entirely new task”
4
Apex full-stack delivery
output
“Apex, an experimentation dashboarding and NPV-grounded decisioning tool built end-to-end by the Economics team”
5
Agentic causal inference
ai_action
“users can simply type "Use a synthetic control to estimate the impact of our investments in region A on GTV." The agentic interface can leverage our internal MCPs to explore annotated data tables, write Python estimation and visualizatio…”
Reported outcome

AI tools drove a near-doubling of PRs and lines of code relative to the 2023H1 baseline, a 33% increase in task category diversity per team member, and enabled economists to build full-stack web applications like Apex end-to-end — work that previously required coordination with separate engineering teams.

Reported metrics
PRs and lines of codeessentially doubling
Unique task categories per team member33%
New task categories added per memberapproximately 1.3
Work outside primary task category~37%
Show all 9 reported metrics
PRs and lines of codeessentially doubling
unique task categories per team member33%
new task categories added per memberapproximately 1.3
work outside primary task category~37%
Core ML share of PRsfalling from 73% to 64%
Infra/DevOps share of PRsdropping from 7.8% to 4.1%
Platform/Tooling PR share increase132%
Experimentation PR share increase+45.2%
monthly Claude usage daysmore than 400%
Reported stack
AvaCursorClaudeLLMGitHubMCPsPython
Source
https://tech.instacart.com/how-ai-changes-the-role-of-applied-scientists-895192d5e114
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

AI tools drove a near-doubling of PRs and lines of code relative to the 2023H1 baseline, a 33% increase in task category diversity per team member, and enabled economists to build full-stack web applications like Apex…

What tools did this team use?

Ava, Cursor, Claude, LLM, GitHub, MCPs, Python.

What results were reported?

PRs and lines of code: essentially doubling; Unique task categories per team member: 33%; New task categories added per member: approximately 1.3; Work outside primary task category: ~37% (source-reported, not independently verified).

What failed first in this deployment?

The team's Causal Inference Platform (CIP), a UI-based tool giving company-wide access to standard causal inference methods, was undermined by AI coding assistants that made bespoke causal inference faster and more fl…

How is this back office ops AI workflow structured?

AI tool adoption waves → LLM PR classification → Task redistribution → Apex full-stack delivery → Agentic causal inference.