From AI Pressure to a Responsible Public-Sector Pilot

Note: This is a case study that shows a common public-sector operational challenge. It illustrates a practical engagement model and does not identify a specific client, agency, employee, procurement action, AI system, or technology implementation.

Nina Patel was the deputy director of a California public agency program responsible for licensing, public inquiries, program guidance, and regulatory compliance.

Her division served thousands of residents, regulated entities, and partner organizations each year. The work required employees to interpret policies, explain requirements, review submissions, identify exceptions, route complex cases, and provide timely updates to the public.

Like many government leaders, Nina was hearing increased pressure to “do something with AI.”

The pressure came from several directions.

Executive leadership wanted to know how the agency could modernize service delivery. Program managers wanted help reducing repetitive work. Employees wanted faster ways to find policies and procedures. Constituents expected government service to become easier to navigate. Technology staff were being asked whether a chatbot, AI assistant, or automated workflow could reduce call volume and improve responsiveness.

At first, the opportunity appeared obvious.

The agency had thousands of documents, frequently asked questions, training materials, policies, regulations, forms, application instructions, procedural manuals, email guidance, and public-facing web pages. It seemed reasonable to assume that an AI assistant could answer questions more quickly than staff who had to search through multiple systems.

But Nina understood that the risks were equally obvious.

A resident could rely on a wrong answer.

A regulated entity could receive outdated guidance.

An employee could unknowingly use a draft or superseded policy.

An AI system could provide a confident response without identifying the official source.

Sensitive information could be handled improperly.

A decision that required professional judgment, legal interpretation, or supervisory authority could be treated as a routine question.

The agency did not want to become known for deploying an AI tool that was fast, impressive, and unreliable.

Nina put the issue plainly in a leadership meeting:

“We cannot ask AI to answer questions until we know which information it should be allowed to use, what it must cite, when it must defer to a human being, and who is accountable for keeping that information current.”

That statement changed the conversation.

The agency no longer framed the challenge as:

“Which AI tool should we buy?”

It began asking:

“What must be true before we can use AI responsibly?”

Nina’s program had no shortage of information.

It had a shortage of governed, operationally usable information.

Policies were stored in multiple locations. Public guidance lived on web pages that did not always align with internal desk guides. Some procedures had been updated through memos or emails but were not reflected in formal training materials. Experienced employees often knew how to interpret common exceptions, but their knowledge had not been captured in a consistent form.

The program also had a wide range of question types.

Some questions were simple.

A resident might ask what form to use, where to submit an application, what documentation was required, or how to check a status.

Other questions were more complex.

A regulated entity might ask how a policy applied to a unique circumstance. An employee might need to determine whether an exception was allowable. A supervisor might need to decide whether a case should be escalated for legal, compliance, or executive review.

The agency had to distinguish between information that could be made easier to find and decisions that required accountable human judgment.

Without that distinction, an AI deployment could create more risk than value.

Nina met with OKF Expert to discuss a different kind of AI-readiness effort.

Instead of beginning with a chatbot, product demonstration, or broad software procurement, OKF Expert proposed a fixed-price Responsible AI Workflow and Knowledge Readiness Audit.

The engagement was designed to answer practical questions before the agency committed to a major technology implementation.

* Which workflows create the greatest burden for employees and the public? * Which questions are high-volume, repetitive, and suitable for better knowledge access? * Which policies, procedures, and guidance documents are authoritative? * Which content is outdated, incomplete, contradictory, or difficult to locate? * Which sources should be available to staff? * Which sources could eventually support public-facing guidance? * Which use cases require citations and traceability? * Which decisions must remain under human control? * Which information is sensitive, restricted, or inappropriate for an AI-enabled workflow? * What governance structure is required to keep knowledge current after the pilot ends? * What small pilot could demonstrate value without exposing the agency to unnecessary risk?

OKF Expert explained the central principle:

Responsible AI does not begin with the model. It begins with the workflow, the knowledge, the people, the policies, and the accountability structure behind the work.

Nina recognized that this was the kind of starting point her agency needed.

The goal was not to slow innovation.

The goal was to make innovation trustworthy.

OKF Expert provided Nina with a concise, one-page scope of work.

The scope was designed to be understandable to program staff, executive leadership, procurement professionals, technology teams, legal counsel, and information-security stakeholders.

It defined:

* The business process or service area to be assessed * The high-volume questions and workflow pain points in scope * The policies, procedures, forms, web pages, desk guides, and training materials to be reviewed * The stakeholder interviews and working sessions * The AI-readiness and risk-review activities * The governance deliverables * The recommended pilot scope * The project timeline * The fixed price * The acceptance criteria * The approach to confidentiality and agency information * The role of citations, human oversight, and accountable decision-making

Nina described the project to her procurement team in straightforward operational language:

“We are not asking for an AI product. We need to identify the workflows where better knowledge access would improve service, determine what content is trustworthy enough to use, identify risks, and define a responsible pilot before we purchase or deploy a larger solution.”

She then asked:

“Would this engagement be appropriate for consideration through the SB/DVBE Option? OKF Expert is a California-certified Small Business and Disabled Veteran Business Enterprise, and the work is a tightly scoped, fixed-price professional service.”

Procurement reviewed the request and followed the agency’s applicable purchasing process, including required quote, documentation, and approval steps.

The engagement was easy to evaluate because it was not framed as an open-ended experiment.

It was a defined, fixed-price effort to reduce uncertainty, identify practical opportunities, and create a governance-ready foundation for future decisions.

OKF Expert began by examining the agency’s most common service interactions.

The team reviewed what happened when a member of the public, a regulated entity, or an employee needed an answer.

The assessment looked at:

* The question being asked * The person responsible for answering it * The systems and documents used to find information * The policies and procedures that governed the answer * The time required to locate the correct guidance * The points where staff needed supervisory assistance * The situations that required legal, compliance, or specialized review * The information that could be safely standardized * The information that required case-specific judgment * The records that needed to be retained for accountability * The risks of providing incomplete, outdated, or incorrect guidance

The audit revealed that the agency had several distinct categories of work.

These were questions that could often be answered consistently from an authoritative source.

Examples included:

* Where to find a form * Which documents to submit * How to meet a filing deadline * Where to find a public guide * How to check the status of a request * What standard eligibility or submission requirements applied * Which office handled a particular service

These questions were not necessarily easy for staff to answer because the information was often scattered.

But they were strong candidates for improved knowledge access because the answer could be tied to a current, cited source.

These were questions where staff needed to follow a sequence of steps.

Examples included:

* Determining whether an application package was complete * Identifying which supporting documents were required * Routing a request to the correct program area * Explaining the next step after a standard submission * Using a checklist to identify missing information * Following a standard workflow for a common request

These situations could benefit from structured decision paths, checklists, workflow guidance, and eventually limited automation support.

However, the guidance had to be governed carefully because an incomplete or outdated decision path could create errors at scale.

These were questions involving discretion, interpretation, exceptions, sensitive information, or decisions that materially affected a person, organization, or public obligation.

Examples included:

* Determining whether an exception should be granted * Interpreting ambiguous or conflicting policy language * Deciding whether a submission met a complex standard * Reviewing potential enforcement or compliance concerns * Handling sensitive records * Making determinations that required supervisory, legal, or executive authority * Responding to unusual circumstances not addressed by standard guidance

The audit concluded that these matters should not be treated as simple automation opportunities.

They required clear escalation rules, appropriate documentation, accountable human review, and possibly specialized subject-matter expertise.

This distinction was one of the most important outcomes of the project.

The agency learned that responsible AI readiness was not about finding one answer for every question.

It was about knowing when a tool could support a person and when a person had to remain responsible for the decision.

The Workflow Audit also examined the quality and governance of the agency’s information.

The review found several recurring risks.

Some procedures existed in several forms.

A formal policy document might have one answer. A desk guide might have a simplified version. A training presentation might describe an older process. An email from a prior year might contain a change that was never formally incorporated into the authoritative documentation.

This created a serious problem for any AI-enabled workflow.

An AI assistant cannot reliably distinguish current guidance from obsolete guidance unless the agency has already made that distinction.

Many documents had no identifiable owner.

Employees could use them, but no one was clearly responsible for reviewing them, updating them, retiring them, or confirming that they remained accurate.

The agency needed to know:

* Who owns this content? * Who approves changes? * How often should it be reviewed? * What happens when the policy changes? * How will staff know that a prior version is no longer valid? * Which version is authorized for use in an AI-enabled knowledge environment?

Some staff guidance explained what to do but did not identify the source that supported the instruction.

This made it harder for employees to confirm an answer, explain a decision, or identify whether a policy change affected the process.

For AI readiness, traceability was essential.

The agency needed staff and users to be able to understand:

* What source supports this answer? * Is the source current? * Does the guidance apply to this situation? * Is the answer a direct policy statement, an internal procedure, or a general explanation? * Does the question require human review?

Some of the agency’s most useful knowledge existed only in the experience of long-tenured employees.

These employees knew how to navigate exceptions, identify missing information, resolve recurring problems, and recognize when a question needed escalation.

That expertise was valuable.

But if it was not documented and governed, it created risk.

The agency could lose critical knowledge when employees retired, transferred, or became unavailable. It could also create inconsistency if different staff members applied their personal experience differently.

OKF Expert translated the findings into a practical readiness framework.

The framework helped the agency organize its modernization work around five connected areas.

The agency needed a clear understanding of the work AI might support.

This included:

* The start and end of each workflow * The people involved * The information required * The rules that governed the work * The decisions made at each stage * The approvals and escalations required * The systems of record * The points where errors or delays occurred

Without workflow clarity, AI use cases would remain vague and difficult to govern.

The agency needed to identify which content was authoritative and who was responsible for maintaining it.

This included:

* Source identification * Content ownership * Review cycles * Version control * Retirement of obsolete material * Citation requirements * Audience and access considerations * Procedures for updating content after policy or process changes

Without knowledge governance, the agency risked giving staff or the public access to answers that were outdated or unsupported.

The agency needed a practical way to distinguish lower-risk from higher-risk use cases.

The assessment grouped potential AI-supported activities into three levels:

* Low-risk support: Finding approved information, summarizing public guidance, identifying forms, directing users to published resources, or assisting staff with document navigation. * Moderate-risk support: Guiding staff through standard checklists, helping identify missing information, routing routine requests, or drafting communications for human review. * High-risk support: Interpreting policy, determining eligibility, granting exceptions, making enforcement or compliance decisions, processing sensitive information, or taking actions that materially affect rights, benefits, or obligations.

The agency could now pursue low-risk, high-value improvements while preserving oversight for more sensitive work.

The assessment clarified where a human being had to remain responsible.

The agency defined:

* Which answers could be provided from approved sources * Which recommendations required staff confirmation * Which decisions required supervisory review * Which matters required legal, compliance, or program-expert involvement * Which actions required documentation in a system of record * Which workflows should never be delegated to an AI-enabled process without additional controls

This helped the agency avoid a common mistake: treating AI as an autonomous decision-maker rather than a tool that supports accountable public servants.

The agency needed a small, measurable use case that could demonstrate value.

The audit identified a high-volume service area where employees and the public frequently sought the same information.

The proposed pilot would not make final decisions.

It would focus on governed access to approved information.

The pilot would help staff locate current, cited answers to common questions and identify when a matter needed escalation.

The agency could measure:

* Time required to locate approved guidance * Frequency of supervisor escalations * Consistency of answers across staff * Number of outdated or duplicative documents identified * Employee confidence in using approved guidance * Common information gaps * Volume and type of questions routed for human review * Feedback from frontline staff and program managers

This created a practical way to demonstrate improvement without making unsupported claims about a large AI transformation.

The agency selected one service area for an OKF Bundle pilot.

Rather than attempting to organize every policy, procedure, and document across the agency, the pilot focused on the information most relevant to a defined set of high-volume questions and staff workflows.

The OKF Bundle included:

* Current public-facing guidance * Approved internal procedures * Relevant policies and regulations * Forms and submission requirements * Frequently asked questions * Standard checklists * Decision-support guidance * Escalation criteria * Communication templates * Citation links to authoritative sources * Content ownership assignments * Review dates * Maintenance responsibilities * Restrictions on use * Human-review requirements

The bundle was designed to help staff understand:

* What information they could use directly * Which source was authoritative * Where the answer came from * How recently the information had been reviewed * What workflow the content supported * When a matter required escalation * When a question required a human decision rather than a standard answer

The result was not simply an organized set of documents.

It was a governed knowledge foundation for better service today and responsible AI support tomorrow.

The engagement gave Nina’s agency a more disciplined way to move forward.

Leadership gained a practical understanding of where AI might create real value and where it could create unacceptable risk. Program managers gained a clearer picture of the workflows that caused the greatest burden. Information-security, legal, and compliance stakeholders gained a structured way to discuss controls. Frontline staff gained a path toward faster access to reliable information.

Most importantly, the agency gained permission to move from urgency to clarity.

The conversation changed from:

“We need an AI solution.”

To:

“We need to improve a defined workflow, govern the knowledge behind it, protect sensitive information, require citations where appropriate, and ensure that human accountability remains in place.”

That change made responsible innovation possible.

The agency did not begin by buying an AI platform.

It began by understanding its work.

The initial engagement focused on a visible leadership concern: pressure to modernize, repeated questions, fragmented information, inconsistent guidance, and uncertainty about how to use AI safely.

OKF Expert helped the agency create a practical progression:

AI-Readiness Workflow Audit → Knowledge Governance → Risk Classification → OKF Bundle Pilot → Measurable Responsible AI Support

This approach allowed the agency to make progress without pretending that technology alone could solve its operational problems.

It also ensured that future AI investments could be tied to clear workflows, authoritative information, citations, human oversight, and measurable public-service outcomes.

Public agencies should not begin their AI journey by asking what tool to buy.

They should begin by asking what work needs to improve, what knowledge supports that work, which information is authoritative, what risks must be controlled, and where human accountability must remain.

Responsible AI is not a product category.

It is an operating discipline.

OKF Expert helps agencies make their workflows, institutional knowledge, policies, and procedures usable, governable, citation-backed, and ready for responsible AI.

OKF Expert is a dba of eGovernment.ai which is a California-certified Small Business and Disabled Veteran Business Enterprise. Tightly scoped engagements may be suitable for consideration through the SB/DVBE Option, subject to the agency’s procurement process and applicable requirements.

* Workflow Audit * OKF Bundles * Responsible AI Readiness * California SB/DVBE Procurement Pathways * Knowledge Governance * Public-Sector AI Implementation