Let the backlog pick the pilot. The deepest pile with the costliest error pays first: documents, the service queue, the case file.
Book the clock before the promise. The buying route is settled before any date is promised.
Run the boundary test. An output that decides a benefit, enters a record, sends an inspector or speaks for the agency clears a higher bar.
Count the AI already running. Rules engines, matching systems and fraud models go on the inventory before new buys.
Measure from day one. Mean time to resolution and average speed to answer go on a dashboard before the pilot starts.
About 80% of states now report AI pilots in production, NASCIO's 2026 data shows. Doug Robinson, executive director at NASCIO, credits that to strong governance, oversight and measuring impact along the way. One decision left is which workflow goes first. The other is the bar for any output that decides or recommends about a person's benefit, permit or tax, enters a record, sends an inspector or speaks in the agency's name. Pick your first workflows in an agency, a city or a county: the pay sits in the same places.
Records and back-office documents. California's Employment Development Department scans more than twenty million mailed pages a year. AI reads and keys them; what it cannot read goes to a person. The records lesson goes deep.
The service queue. Denver reports a 311 transaction costs $4.75 with a person and 35 cents through Sunny and other digital channels. Sunny closes simple cases and routes harder ones to human agents. The front-door lesson goes deep.
Eligibility and case processing. Robinson says case and benefits management top the use-case list. Several state legislatures have barred AI from final determinations such as benefit eligibility. The determinations lesson sets the bar.
Inspection and compliance. Mark Raymond, state CIO at the State of Connecticut, says one agency's regulatory reviews went from 40 hours of preparation to four. Staff judge the material instead of assembling it. The inspection lesson goes deep.
The legacy estate. Bill Kehoe, state CIO at the State of Washington, says AI crawls COBOL and other old code. It surfaces the eligibility and business rules buried there before replacement. The legacy lesson goes deep.
Emergency response. The warning that goes out in the agency's name. The response lesson goes deep.
Fraud and improper payments. California EDD's fraud hotline once piled up thousands of voicemails, many not in English. AI now transcribes them into English text, removes duplicates and opens a ticket automatically. Lesson 3's clawback row goes deep.
The caseworker and inspector shortage. Work that can't be staffed, c1-4, goes deep.
Six checks before the first pilot
Backlog size and error cost pick the first workflow. Suma Nallapati, chief AI and information officer at the City and County of Denver, tells peers: pick a real problem that has exhausted ordinary resources. Adoption is what makes an AI program succeed.
Book the clock before the promise is dated. Katrina Flory, state CIO at the State of Ohio, says the state added an AI category to an existing data and analytics contract set. Ohio pre-qualified 33 vendors, so even production deployments can run through statements of work, not fresh solicitations. Lesson 2 covers the routes.
Run the boundary test. Does the output decide or recommend about a person's benefit, permit or tax? Does it enter a record the law keeps, send a crew or an inspector, or speak in the agency's name? A yes means a named owner, and the bar the lesson for that output sets. Ajit Girn, chief information officer at the California Employment Development Department, says EDD keeps a standing rule. AI does not determine benefits or adjudicate claims. Robinson says prompts and AI results from state agencies are public record under nearly every state's open-records law. In his words, "a record is a record regardless of physical form or characteristics."
Count the AI that already runs. Rules engines, matching systems, fraud models and form auto-fill go on the inventory first. Robinson says nearly every commercial product a state buys "most likely already has AI embedded in it." The count has its own lesson: c5-1, count the AI you already have.
Test "gov-ready AI" on the agency's own work. Tony Day, chief digital officer at the Government of New Brunswick, says the province spent slightly over six months on large language model R&D before deploying anything. External platforms meant handing sensitive data over, which was "a non-starter for government." c7-5, vendor real AI or relabel, carries the test.
Measure from day one. Denver measures quality alongside cost on dashboards its incident experience director reviews: mean time to resolution, average speed to answer. The method has its own lesson: c1-8, baseline and bar before pilot.
Exercise (20 minutes): pick your first AI workflow
Take the real choice in front of your agency, city or county: the two or three candidate workflows that compete to go first, from the map above. Rank them by pile and error cost on the card, then work the checks in order. You won't know every line yourself. Every line you can't fill gets an owner and a date before you close the card. The card download includes a filled-in example of a first pick that isn't ready yet, and why.
First AI workflow card (PDF)
A blank card and a filled-in example, one page each.
Show the three prompts: fill the card, test it, turn it into a plan
Use AI to fill in and test your card
These three prompts go into your organization's approved AI assistant. Use the first to fill in the card one row at a time; it won't guess a number for you. Use the second when the card is done, to hear the case against starting the pilot. The third asks about your level of government and your workflows and turns the lesson into a 30-day plan, with a checklist, rubric or brief for your boss if you want one. Use placeholders for agency, system and people names; the help doesn't depend on them.
Prompt 1: Fill in my card with me
Paste this, then add your candidate workflows in one or two sentences and the card below.
Prompt
Before you start: use your company's approved AI assistant, not a personal account. Replace agency, system, supplier and people names with placeholders (Agency A, System X) and round any figures. The help you get does not depend on them.
You are helping me fill in a first-AI-workflow card for my agency. The card is pasted at the end.
Rules:
- Ask me about one row at a time, in order, and wait for my answer before moving on.
- Write down only what I tell you. Never suggest, estimate or invent a number, a name, a yes/no answer or a decision for me, even if I ask you to guess. If I ask, say that this row needs a real answer and ask who could give it.
- If I don't know an answer, write "Not known" and ask who should find out and by when. Put that in the Owner / by when column.
- After the last row, show the whole card as a table with three columns: Question, Answer, Owner / by when. Then list every row marked "Not known" with its owner, and remind me that any unanswered row means the first workflow is not ready to start.
My candidate workflows in one or two sentences: [DESCRIBE THE WORKFLOWS COMPETING TO GO FIRST]
The card:
[PASTE THE CARD]
Prompt from The AI Decision Course by Dr. Tony Hoang, host of The Artificial Intelligence Podcast (Spotify, Apple Podcasts): tonyphoang.com/learn
Prompt 2: Stress-test my card
Paste this, then your filled card.
Prompt
Before you start: use your company's approved AI assistant, not a personal account. Replace agency, system, supplier and people names with placeholders (Agency A, System X) and round any figures. The review does not depend on them.
Below is a first-AI-workflow card I filled in for my agency. Act as a skeptical reviewer at the decision meeting and argue the case for HOLD.
For each row:
- Say whether the answer is specific enough to act on (a number, a named role, a date) or too vague.
- Point out where answers contradict each other, for example a workflow that clears the boundary test against a plan with no named owner, or a public promise dated before the buying route is booked.
- Say what evidence you would ask for before accepting it.
Do not rewrite my answers and do not invent facts about my agency. If a row is blank or says "Not known", treat it as not ready.
Finish with: your recommendation (GO or HOLD), the three rows that most need work, and one question I should ask my team on Monday.
My card:
[PASTE YOUR FILLED CARD]
Prompt from The AI Decision Course by Dr. Tony Hoang, host of The Artificial Intelligence Podcast (Spotify, Apple Podcasts): tonyphoang.com/learn
To paste the card into prompt 1, copy its rows from the PDF.
Prompt 3: Make it yours
Paste this and answer its four questions.
Prompt
Before you start: use your company's approved AI assistant, not a personal account. Replace agency, system, supplier and people names with placeholders (Agency A, System X, Supplier B).
You are helping a leader apply one lesson to their own situation. The lesson's five decisions are below. Use only them and what I tell you. Never invent facts, numbers, vendors, laws or results about my organization or my level of government. Where a number is needed and I have not given one, say what to measure and who should measure it, instead of guessing.
First, ask me these four questions, one at a time, and wait for each answer:
1. My level of government (state, city, county or agency) and roughly how many people work in my organization.
2. The candidate AI workflows I have in mind, in one or two sentences.
3. Where the work stands today: an idea, a vendor demo, or a pilot ready to start.
4. What worries me most about it.
Then give me the following. Keep it short and practical:
A. My plan. For each of the five decisions: one action written for my workflows, the role that should own it (a role, not a name), and what should be done within 30 days. If my level of government makes a decision harder, for example open-records law, a bar on AI determinations, or a procurement clock, say which decision and why in one sentence.
B. Three questions to ask my team on Monday, specific to my workflows.
C. Ask me which one I want: a checklist for my first workflow, a rubric to score candidate workflows, or a one-page brief for my boss. Then write it, and finish with a prompt I can paste later to update it.
Mark anything that rests on an assumption with "(check)".
The five decisions:
1. Let the backlog pick the pilot. The deepest pile with the costliest error pays first: documents, the service queue, the case file.
2. Book the clock before the promise. The buying route is settled before any date is promised.
3. Run the boundary test. An output that decides a benefit, enters a record, sends an inspector or speaks for the agency clears a higher bar.
4. Count the AI already running. Rules engines, matching systems and fraud models go on the inventory before new buys.
5. Measure from day one. Mean time to resolution and average speed to answer go on a dashboard before the pilot starts.
Prompt from The AI Decision Course by Dr. Tony Hoang, host of The Artificial Intelligence Podcast (Spotify, Apple Podcasts): tonyphoang.com/learn
Your company's assistant may answer differently, so read what it writes before you use it.
Check yourself
Two workflows compete for the first pilot: a document backlog and an inspection scheduler. Which goes first?
Show the answer
Rank them by pile and error cost; the ranking picks.
A vendor's "gov-ready" demo reads a permit form perfectly, on the vendor's samples. What happens before you sign?
Show the answer
The claims get tested on the agency's own work.
A proposal would let AI draft the letter that denies a benefit claim. What does it need first?
Show the answer
A named owner, and the bar the determinations lesson sets.
Nobody calls the fraud-matching rules engine AI. Does it count?
Show the answer
Yes. It goes on the inventory before anything new is bought.
The mayor wants to promise that AI will cut the 311 queue by fall. What do you book first?
Show the answer
The clock: the buying route. Date the promise after.
Sources
Ajit Girn, Chief Information Officer, California Employment Development Department, on The Business of Government Hour (2026): "Modernizing Public Benefits Delivery: A Conversation with Ajit Girn, CIO, California EDD". https://exa.ai/library/podcast/6877y9tnnpt/episode/tnpy712mrb9 Produced by the IBM Center for the Business of Government.
Mark Raymond, State CIO, State of Connecticut; Katrina Flory, State CIO, State of Ohio; and Bill Kehoe, State CIO, State of Washington, on Priorities Podcast (2026): "Rapidfire Chats with State CIOs at NASCIO's 2026 annual conference". https://exa.ai/library/podcast/8kb2dn598s8/episode/m01jcpp2pb0
Suma Nallapati, Chief AI and Information Officer, City and County of Denver, on AI-First Podcast (2026): "How Denver is Leapfrogging the Private Sector in AI Innovation | Box AI-First Podcast EP 22". https://exa.ai/library/podcast/5v9j47pnvqm/episode/w84jkhrdjf3 Produced by Box.
Tony Day, Chief Digital Officer, Government of New Brunswick, on Public Sector Podcast (2026): "Navigating the Digital Frontier: GNB's Approach to AI and Innovation - Tony Daye - Episode 187". https://exa.ai/library/podcast/3mtqw4bx9vj/episode/ytq561149t4
Where AI pays first in government, and the bar it has to clear
Ask about this lesson, or how it applies to your team. Every answer cites the lessons it draws on.
Try asking
Free preview: 3 questions a day, answered from this lesson. Follow-ups remember your earlier questions on this visit. Don't paste confidential details.