Data entry ops · Production

Plaid deploys AI Annotator and Fix My Connection agents to accelerate data labeling and repair bank integrations

The problem

Plaid's demand for labeled transaction data across Personal Finance, Credit, and Payments could not be met by manual labeling at scale, limiting model improvement. Separately, maintaining thousands of bank integrations manually was costly, and login-experience updates at financial institutions caused user disruptions that hurt conversions and satisfaction.

Workflow diagram · grounded in source
1
Labeling demand triggers annotation
trigger
“the demand for labeled data spans across Personal Finance, Credit, Payments and other use cases, but has struggled to keep pace”
2
LLM generates transaction labels
ai_action
“Using LLMs to create likely transaction labels with high precision and consistency”
3
Human review for golden dataset
human_review
“Initially involving human reviewers primarily to generate "golden data-set" for benchmarking and then selectively engaging for edge cases or quality spot-checks, rather than bulk labeling”
4
Labeled datasets published
output
“A single environment where internal teams can annotate, review, and scale labeled datasets to support multiple initiatives”
5
Connection quality issue detected
trigger
“Automations spot connection quality issues before user disruption spreads”
6
Agents analyze and generate repair scripts
ai_action
“Intelligent agents analyze bank integrations for potential breaking changes, automatically generating scripts to repair them”
7
Automated repair deployed
output
“automated repairs have already enabled over 2 million successful user-permissioned logins, and reduced the average time to fix a degradation by 90%”
Reported outcome

The AI Annotator produces high-quality labels with greater than 95% human alignment at a fraction of cost and time.
Fix My Connection has enabled over 2 million successful user-permissioned logins and reduced the average time to fix a degradation by 90%.

Reported metrics
Label human alignmentgreater than 95%
Annotation cost and timefraction of cost and time
Successful user-permissioned logins enabledover 2 million
Average time to fix a degradation90%
Show all 5 reported metrics
label human alignmentgreater than 95%
annotation cost and timefraction of cost and time
successful user-permissioned logins enabledover 2 million
average time to fix a degradation90%
manual repairs requiredfewer manual repairs required
Reported stack
AI AnnotatorLLMsFix My ConnectionMCP server
Source
https://plaid.com/blog/ai-agents-june-2025/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The AI Annotator produces high-quality labels with greater than 95% human alignment at a fraction of cost and time.

What tools did this team use?

AI Annotator, LLMs, Fix My Connection, MCP server.

What results were reported?

Label human alignment: greater than 95%; Annotation cost and time: fraction of cost and time; Successful user-permissioned logins enabled: over 2 million; Average time to fix a degradation: 90% (source-reported, not independently verified).

How is this data entry ops AI workflow structured?

Labeling demand triggers annotation → LLM generates transaction labels → Human review for golden dataset → Labeled datasets published → Connection quality issue detected → Agents analyze and generate repair scripts → Automated repair deployed.

Related data entry ops cases