TREAS · TREAS-117 · record R2804

Local GenAI Pilot

Department of the Treasury · Internal Revenue Service (IRS)

Overview

Development stage: PilotSource: b) Pilot – The use case has been deployed in a limited test or pilot capacity.
High-impact designation: Not high-impactSource: c) Not high-impact
Impact justification: The output is not presumed to be high-impact and is not used as the principal basis for significant decisions/actions
Start date: 2025-03-25 (day precision)Source: 3/25/2025
Withheld from public reporting: NoSource: a) No

Mission

Topic area: Information Technology
Operational functions: Summarization, Software development assistance, Translation and transcription Derived by keyword rules; see methodology

Problem the AI is intended to solve

Local Large Language Models (LLMs) will help in local no software/hardware (SW/HW) cost. LLMs will position IRS for sustainable AI services and AI tool integrations and mitigate many risks [Several Risks Mitigated Include: vendor lock-in, licenses shortages or over purchasing, SW costs or per Application Programming Interface (API) call costs, contract gaps and access loss with outsourcing, and data privacy policies and interagency agreements restricting data leaving the IRS (Auditing/Compliance Algorithm's Source Code)]

Expected benefits

Reducing human workload by leveraging LLMs increasing development speed, automation, and operational efficiency. Ensuring alignment with internal Information Technology (IT) policies and security protocols by operating entirely within IRS infrastructure. Eliminates licensing costs reducing operational expenses. Maintains full control over sensitive data, enhancing data governance and compliance with federal privacy mandates. Reducing maintenance costs and technical debt due to legacy code and technology. Aging/departing workforce and lack of expertise with legacy systems increases risk of system failures and operational bottlenecks. Resource optimization, efficiency, and sustainable solutions required to modernize critical IRS operations and services

System outputs

Contextually grounded responses derived from the given input prompt. Examples: Text Generation, Summarization, Code Completion, Unit Tests, Text Translation, Text-to-Text Transformation.

Technology

AI classification: Generative AISource: Generative AI: AI that generates new or synthetic content (e.g., images, videos, audio, text, code).
System name(s): Not reported
Custom-developed code: No
Public source code: github.com ↗

Sourcing

How it was built: Developed in-houseSource: b) Developed in-house
Vendor (as reported): Not reported
Vendors (standardized): None on the standard list

A vendor is the supplier named by the agency. It does not identify the underlying model or AI technology.

Data and privacy

Involves PII: No
Privacy Impact Assessment: Not reported
Authorization to Operate: Yes
Demographic features: k) None of the above
Training and evaluation data: We have LLMs that are fully pre-trained. Each model is specifically trained on various use case categories like code translation, text summarization etc. We do not perform additional training for these models at this time.
Federal Data Catalog: Not reported

Governance

0 of 8 minimum-practice questions answered. A blank answer means the agency reported nothing; it does not mean the practice is absent.

Pre-deployment testingNo answer reported

AI impact assessmentNo answer reported

Independent reviewNo answer reported

Ongoing monitoringNo answer reported

Operator trainingNo answer reported

Fail-safeNo answer reported

Appeal processNo answer reported

User and public consultationNo answer reported

Potential impacts and how they were identified

Not reported