Article reader Listen + reading controls

LISTEN + READ YOUR WAY

Article reader

Preparing the reader…

0:00 0:00

Speed

Reading settings

Typeface

Text size

100%

Reading surface Text spacing

The Best Government AI Service May Prevent the Contact¶

Artificial intelligence can translate documents, summarize case histories, route requests, help employees search policy, and provide conversational assistance outside business hours. Those are useful capabilities. They can also make a broken service easier to enter without making it easier to complete.

For public services, the objective should not be a more sophisticated chatbot or a lower call-handle time. It should be a reduction in the effort people must spend obtaining an accurate, lawful, and timely outcome.

The highest-value AI intervention may be the one that prevents the person from needing to contact the agency at all.

At the Senate Homeland Security and Governmental Affairs Committee’s January 2024 hearing on AI and government services, witnesses and members discussed talent, data, procurement, privacy, and leadership as conditions for effective adoption. Chairman Gary Peters noted opportunities including translation, round-the-clock assistance, increased employee capacity, and easier navigation of complex processes.¹

Those opportunities should be evaluated across the full service journey.

Begin With Failure Demand¶

Many contacts occur because an earlier part of the service failed:

a notice was unclear;
an application requested information the government already holds;
a status was unavailable;
two agencies used inconsistent records;
a deadline or next step was ambiguous;
an automated decision was wrong;
or the person could not reach someone authorized to resolve an exception.

This is failure demand: work created by the service’s inability to complete the user’s need correctly the first time.

An AI agent can absorb some of that demand. If the underlying cause remains, the agency has automated the symptom and may conceal the scale of the failure.

Service analytics should classify contacts by root cause and ask which can be prevented through clearer policy, prefilled data, proactive notification, process redesign, or correction of a system defect.

Model the Service as a Journey, Not a Channel¶

People experience a service across websites, paper forms, phone calls, field offices, email, caseworkers, contractors, and other agencies. A conversational interface is one channel within the journey.

A journey map should identify:

the person’s goal and circumstances;
key decisions and evidence;
handoffs among organizations;
wait states and uncertainty;
accessibility and language needs;
common exceptions;
emotional and financial burden;
and points where an error becomes costly or irreversible.

AI should be placed where it reduces total effort and improves outcome quality. That may mean assisting an employee behind the scenes, detecting an inconsistent record, generating a plain-language notice, or predicting that a case is likely to stall.

The visible front end is not always the leverage point.

Use AI to Increase Staff Capacity, Not Just Deflect People¶

Public-sector chatbot business cases often count calls avoided. Deflection can be beneficial when the person receives a complete answer. It can be harmful when the system makes human help difficult to reach.

AI can instead augment staff:

summarize a long case history with links to source evidence;
retrieve policy and identify conflicting guidance;
draft correspondence for review;
translate while preserving human escalation;
identify missing information before a case reaches adjudication;
prioritize cases based on urgency and harm rather than ease;
and expose recurring failure patterns to service owners.

The employee remains accountable for consequential decisions, but the system reduces search and administrative burden.

Evaluation should measure whether staff resolve more needs correctly, not simply whether each interaction is shorter. Some complex cases deserve more time.

Knowledge Infrastructure Is the Core Dependency¶

A service agent needs authoritative, current, and context-specific knowledge. Federal policy often exists across statutes, regulations, manuals, directives, notices, local procedures, and case-specific records.

Retrieval-augmented generation can help only when the knowledge layer provides:

authoritative source identification;
version and effective dates;
jurisdiction and applicability;
relationships among rules and exceptions;
provenance from answer to source;
access control and privacy;
and a process for resolving conflicting material.

A vector database of documents is not a knowledge strategy. The agency needs owners responsible for meaning, currency, and change.

Answers should link to the supporting source and communicate uncertainty. If the system cannot determine which policy applies, it should route the case with the relevant evidence rather than produce a plausible synthesis.

Intergovernmental data sharing can prevent repeated requests and improve coordination. It also creates privacy, consent, quality, and purpose risks.

Agencies should define:

which specific decision requires the data;
the legal authority and permitted purpose;
the source and quality expectations;
how conflicting records are resolved;
what the person can see and correct;
retention and secondary-use limits;
and accountability when shared data causes an error.

Sharing fewer, better-governed attributes may produce more value than copying complete records across agencies.

Data provenance should remain visible to caseworkers and affected people where appropriate. An incorrect fact should not become more authoritative merely because several systems copied it.

Accessibility and Human Escalation Are Product Requirements¶

Government services must work for people with disabilities, limited English proficiency, low digital literacy, unstable connectivity, and complex circumstances.

An AI interface should be tested with representative users and assistive technology. It should not require people to phrase a need in the model’s preferred vocabulary. Alternative channels should remain substantively equivalent, not slower punitive paths.

Human escalation must be designed:

clear triggers based on user request, uncertainty, distress, legal consequence, or repeated failure;
transfer of conversation and evidence so the person does not start over;
routing to someone with authority, not merely another general queue;
and service-level expectations for response.

“Talk to a human” should not be treated as model failure. It is a normal capability of a responsible service.

Measure Outcomes and Burden¶

Useful service measures include:

successful completion of the user’s goal;
first-contact resolution;
total elapsed time and number of handoffs;
repeat contact and failure-demand rate;
accuracy and consistency of outcomes;
accessibility and language performance;
human escalation rate and quality;
appeal, correction, and remedy;
staff workload and judgment;
privacy and security incidents;
and differences among affected populations.

User satisfaction matters but is insufficient. A friendly agent can provide an incorrect answer; a lawful adverse decision can be communicated respectfully and still produce low satisfaction.

Measures should connect experience, legal correctness, operational efficiency, and public outcome.

Procurement Should Buy a Learning Service¶

Agencies should resist procuring a fixed “AI customer service solution” based primarily on demonstration quality.

Contracts should require:

integration with authoritative knowledge and case systems;
evaluation against representative service journeys;
citations and provenance;
accessibility evidence;
privacy and security controls;
version and change management;
telemetry for outcome and failure analysis;
government control of service data and configurations;
and the ability to change model providers without rebuilding the journey.

The vendor can supply technology. The agency must own the public service, its policy, and its evidence.

The Strategic Inference¶

AI can make government services more responsive. It can also create a polished layer over fragmented systems and transfer the burden of navigating that fragmentation to a machine that sometimes guesses.

The better strategy is to redesign the service journey. Prevent avoidable contact, give staff decision-ready knowledge, make status and evidence visible, preserve human escalation, and learn from every failed interaction.

The goal is not to keep people talking to an AI for longer. It is to help them reach a correct outcome with less effort and greater dignity.

This essay was substantially revised in July 2026. The original page contained an accidentally overlong headline and a short hearing summary; this version preserves the URL while replacing the body with an evidence-based service-design analysis.

For related work on human-centered AI, knowledge systems, and public-sector product strategy, see my portfolio or connect on LinkedIn.

References¶

Senate Committee on Homeland Security and Governmental Affairs, “Harnessing AI to Improve Government Services and Customer Experience,” January 10, 2024.
Office of Management and Budget, M-24-10: Advancing Governance, Innovation, and Risk Management for Agency Use of Artificial Intelligence, March 28, 2024.

Senator Gary Peters, Opening Statement: Full Committee Hearing—AI and Service Delivery, January 10, 2024. ↩