Putting AI to Work: What Texas Public Pension Professionals Should Know About Agentic AI
The biggest change in artificial intelligence may not be its ability to provide a better answer. It may be what happens after you ask the question.
For several years, most people have used tools such as ChatGPT as a more conversational search engine—asking questions, summarizing documents or drafting emails. A newer generation of tools, often described as agentic AI, can carry out a series of steps and produce something ready for review.
That distinction matters in a public pension office.
Trustees and administrators spend considerable time gathering information, comparing documents, preparing board materials, researching service providers and communicating with members. Agentic AI may help with parts of that work, but it should support—not replace—professional expertise, fiduciary judgment or human accountability.
A recent ZDNET article by David Gewirtz offers a practical introduction to ChatGPT Work and the broader shift toward agentic AI. The following adapts several of those lessons for members of the Texas Association of Public Employee Retirement Systems.
From Answering Questions to Completing Work
OpenAI describes ChatGPT Work as a way to delegate a substantial task and receive a result that a person can review. Depending on the plan, platform and permissions available, it can research information, analyze files, use approved tools and create documents, presentations, spreadsheets and other finished materials.
Here is a simple way to understand the differences:
ToolBest suited for:
- Chat: Answering questions, brainstorming, summarizing information and preparing short drafts.
- Deep Research: Conducting detailed, source-based research and producing a documented report.
- ChatGPT Work: Completing a larger, multistep assignment and producing a file or other reviewable result.
For example, Chat might help draft an outline for a board presentation. Deep research might examine a policy issue and assemble a report with supporting sources. Work could review approved source materials, build the presentation, prepare a companion summary, and flag information that still needs human review.
According to OpenAI's guidance on ChatGPT Work, the most effective assignments have a clear outcome, relevant source materials, and an output that can be reviewed.
Where It Could Help a Pension Office
Public pension work rewards accuracy, consistency and documentation, not speed for its own sake. The most useful applications may be assignments that take considerable staff time but still allow a person to verify the result.
Reviewing Board Materials
Pension fund administrators could ask Work to look for inconsistencies in a board agenda, staff reports, presentations, and supporting documents. Conflicting dates, missing attachments, unexplained calculations or references that do not match could be flagged, for example.
The AI should not determine whether the board packet is legally sufficient or whether a recommendation is appropriate. It can provide another layer of quality control before staff, counsel and consultants complete their reviews.
Organizing Policies and Governance Documents
A pension office could use the tool to inventory policies, identify inconsistent file names, locate duplicate documents and prepare an index showing adoption and revision dates.
If the assignment involves renaming, moving or deleting files, work from copies and keep an untouched backup. Do not allow an AI system to make records-retention decisions.
Supporting Vendor and Procurement Research
When permitted by the pension system's procurement procedures, AI may help compare publicly available information or approved proposal materials against the system's established criteria.
It could create a comparison table, identify unanswered questions or point out differences in services and fees. It should not select a provider, make assumptions about missing information, or replace the due diligence work of trustees, staff, consultants and counsel.
Do not upload confidential proposals, member information or other nonpublic records unless the system has specifically approved the platform and its use for that information.
Preparing Member and Stakeholder Communications
AI can help turn technical reports into draft newsletters, frequently asked questions (FAQs), presentations, or plain-language explanations. This may be especially helpful when communicating complex actuarial, investment or legislative topics.
A knowledgeable person should still verify every fact, number and characterization before publication.
Monitoring Public Information
Work may also help monitor approved public sources and compile links to legislative, regulatory or industry developments. However, any important finding should be confirmed through the original source, not simply repeated from an AI-generated summary.
What to Know About Access and Usage
ChatGPT Work is available through ChatGPT, although specific features depend on the user's plan, platform, region, account settings and workplace permissions.
On the web, Work generally operates in a managed cloud environment. When available in the desktop app, local Work can use files, applications, and browser tools on the user's computer, subject to the permissions granted. OpenAI provides a more detailed explanation of the differences between local and cloud work.
Local access should not automatically be considered safer. Security depends on the device, account, permissions, connected applications, data policies and administrative controls surrounding the task.
For official pension-system work, employees should use an organization-approved account or workspace and confirm how information is stored, retained and used. A personal AI account should not be assumed to provide the same protections as an organization-managed environment.
Usage is another area that can be confusing. Tokens are units of information that the system reads and produces, not a fixed number of working hours. The amount of capacity a task uses depends on its complexity, the selected model, the files and context involved and the required tools. OpenAI recommends checking the current usage and pricing information rather than relying on a general monthly estimate.
Ten Guidelines for Responsible Use
Before turning an AI agent loose on official work, pension offices should establish clear boundaries.
- Begin with a low-risk assignment. Start with public information or nonconfidential sample files before expanding to official work.
- Define the result. Tell the AI what it should produce, what sources it may use and what a satisfactory result should include.
- Ask for a plan. For a large assignment, have the AI explain its proposed steps before allowing it to proceed.
- Work from copies. Keep original documents and backups outside the AI's working folder, especially when files may be renamed, moved or reformatted.
- Use the least access necessary. Begin with read-only access whenever possible. Do not grant broad permissions simply because they are available.
- Protect confidential information. Do not provide passwords, authentication codes, member data, personnel records, nonpublic proposals or other sensitive information unless the system has approved that specific use.
- Require human approval for outside actions. A person should review emails, calendar invitations, file-sharing changes, purchases and other consequential actions before they occur.
- Verify facts and calculations. Check citations, formulas, dates, names and conclusions against the original sources.
- Document what the AI did. Ask it to identify the sources used, the assumptions made, the files changed, and the questions that remain unresolved.
- Address records requirements. Consult the pension system's records officer or legal counsel about how prompts, conversations, drafts and final outputs should be handled under applicable retention and public-information requirements.
Keep the Human in Charge
Agentic AI may become a useful tool for public pension professionals, particularly when the assignment involves research, document preparation, organization or repetitive administrative work. Its value, however, depends on how it is governed.
AI can prepare a first draft, assemble a comparison, identify inconsistencies or organize a large collection of information. It cannot accept fiduciary responsibility, understand every local circumstance, or replace the judgment of trustees, administrators, investment professionals, actuaries and legal counsel.
The best place to begin is with a low-risk, time-consuming assignment that has a clear outcome and can be reviewed independently. Learn what the tool does well, document where it falls short and expand its role only when appropriate safeguards are in place.
TEXPERS members who are experimenting with agentic AI are encouraged to share what they are learning—including both the time-saving successes and the cautionary lessons.
Editor's note: AI products, features, plan limits and security controls change frequently. Confirm current capabilities and your organization's requirements before using an AI tool for official pension-system work.
AI Transparency Note: The banner image accompanying this article was created using generative artificial intelligence, with editorial direction from TEXPERS. It is illustrative and does not depict actual people, meetings or events.
About the Author: Allen Jones is director of communications and event marketing for the Texas Association of Public Employee Retirement Systems (TEXPERS), where he leads editorial strategy, member communications, conference marketing and digital engagement initiatives serving Texas public employee retirement systems. He began his journalism career in 1998 and has worked in journalism and communications for more than 25 years. He has completed Vanderbilt University's Prompt Engineering for ChatGPT course and Writing.io's AI Course & Certification for Non-Profits and AI Course & Certification for Government programs.