IN BRIEF
All 645 households received a site with personalized information, filing instructions and evidence for an appeal; half also received an AI chatbot.
Seventy-eight percent of households offered the chatbot started a conversation.
The probability of filing without a human agent rose by 9.1 percentage points, from 41.4% to 50.5%, but the effect was smaller among less-advantaged households.
WHY IT MATTERS
Administrative and tax procedures often require interpreting information, deadlines and options in ways that favor people who can afford professional help.
This experiment measured real behavior rather than stated willingness to use a service: more households actually filed on their own.
WHAT MADE IT POSSIBLE
The chatbot answered questions and provided guidance tailored to a household's circumstances in the appeal process.
AI did not determine the tax outcome or replace the authority; it added conversational information and decision support on top of a website that already supplied personalized evidence and instructions.
THE STORY
The researchers created a useful comparison for isolating the AI contribution. Every household already had personalized information and evidence supporting a possible appeal; only half could also converse with the chatbot.
With AI access, the share filing without an intermediary rose from 41.4% to 50.5%. Clickstream and conversation evidence suggests the chatbot helped people evaluate information and decide how to act.
The result also contains an important distributional warning: the increase was smaller among less-advantaged households. And the paper measures filing an appeal, not whether the appeal was successful or generated savings.
WHAT WE KNOW
The experiment was randomized and preregistered. [1]
The main outcome is observed administrative behavior: filing an appeal without an agent.
The chatbot's incremental effect was measured against a control group that already had personalized information.
WHAT WE DON'T KNOW
The reported result does not tell us how many appeals were successful or how much money households saved.
Whether the effect replicates in other counties or administrative processes is unknown.
The smaller benefit among less-advantaged households needs further analysis and replication before drawing broader distributional conclusions.
WHAT HAPPENS NEXT
Measure appeal outcomes, not only filing.
Replicate the intervention across different jurisdictions and administrative procedures.
Test designs intended to reduce socioeconomic gaps in use and benefit.
SOURCE TRACE
The field experiment recruited 645 households in Dallas County; all received personalized appeal information and half were randomly assigned access to an AI chatbot that could answer questions and tailor guidance.
Seventy-eight percent of households given chatbot access initiated a conversation.
AI chatbot access increased filing a property-tax appeal without a human agent by 9.1 percentage points, from 41.4% to 50.5%.
The filing increase was smaller among less-advantaged households, which the authors describe as suggestive evidence that the intervention widened existing disparities.
SOURCES
- Taxpayer Behavior in the Age of AI: A Field Experiment on Property Tax AppealsOpen source
CURRENT EVIDENCE
EARLY EVIDENCE
preregistered randomized field experiment with observed appeal-filing behavior