CivicAIIntelligence behind
public service
★★★AI-assisted service redesign · Service & product design

CivicAI

How a governor’s office heard, answered and learned from 300,000+ constituents a year: AI prepared the work, people decided.

Understand the service→Connect the work→Support decisions→Follow through
Role
Sr. Service Design Lead
Scope
Research, service redesign, AI interaction design, and delivery
Timeline
8 months · 4 phases
Setting
State Office of the Governor
Executive summary

Civic AI was designed to help approximately 15–18 government staff manage 300,000+ constituent contacts a year. It brought intake, routing, letters, calls, search and reporting into one place, with AI help, controlled approvals and a record of every action.

The challenge

Help a small government team give timely, accurate answers through legislative sessions and crisis surges, while also handling casework, official documents and public accountability.

What I found

The CRM wasn’t the problem. The work around it was: staff read and routed every item by hand, each person’s history sat in separate systems, and nothing recorded who owned a request or approved a reply.

What I built

One connected service: every channel in one queue, one record per person, named owners and deadlines, and a review chain that blocked sending until every approval was in.

01
Chapter 01 of 05

The problem

A small team answering a state’s letters, calls and official documents, in an office where everything was auditable, discoverable and politically exposed.

The background

From recordkeeping to active constituent engagement.

The Governor’s Office ran its constituent work on Indigov / Granicus. It wanted to move from keeping records to actively engaging people: high-volume intake, routing, shared drafting, fast search, phones, analytics and ceremonial documents such as proclamations. The office asked for a modern, enterprise-grade CRM to replace it.

300K+Contacts a year, rising above 400,000 during legislative sessions.
15–18Staff handling the workload, alongside ongoing cases, correspondence, records requests, and follow-up.
10,000Contacts in a single crisis week, all handled by the same small team.
Problem statement

The same small team had to absorb sessions and crises, keep every reply accurate and on-message, find answers across twenty years of records, and tell leadership what was changing, while every action stayed open to audit and public records requests.

02
Chapter 02 of 05

Research

Three months inside the work before designing any screen: who did it, where it broke, and what the service actually had to do.

Process

How I got there, step by step, with the client.

Eight months, four phases. Each exercise answered a question the previous one raised, and the client was in the room at every turn.

6 weeksDiscovery
6 weeksDefinition
2 monthsService & governance
3 monthsDesign & validation

Discovery

01
  • Kickoff with leadership
  • 7 interviews across the office
  • Shadowing, inbox to filing
  • 3 journey-mapping sessions

Definition

02
  • 47-point audit of 4 platforms
  • 9 reports synthesized
  • Affinity mapping: 13 findings
  • Five whys and current-state flow
  • Playback with leadership

Service & governance

03
  • Future-state map and blueprint
  • Rules versus AI
  • AI/human responsibility map
  • Information architecture

Design & validation

04
  • Layout studies per job
  • Scenario walkthroughs
  • Connected prototype
  • 3 validation sessions
Tools & methodsFigJamFigmaHTML/CSS/JS prototypeIBM AI/Human Context ModelFieldwork Embedded in the office three days a week.

Research plan

01

Reframe the ask

I did

Asked what goes wrong today before talking about any system.

The client helped

Leadership opened up staff, workspaces, and the current CRM and legacy records.

It gave me

Agreement to study the work before choosing a solution.

02

Go where the work happens

I did

Interviews, shadowing, and journey-mapping sessions on the real workflow.

The client helped

Staff time, anonymized letters and calls, and walkthroughs of every system.

It gave me

Where effort, context, and responsibility were being lost.

03Turning point

Make sense of it

I did

Current-state flow and journey, personas, affinity mapping, five whys.

The client helped

A playback session where leadership confirmed the diagnosis.

It gave me

The CRM request became a service redesign.

04

Design the service

I did

Future-state map, blueprint, rules versus AI, the AI/human responsibility map.

The client helped

IT and security set access and audit rules; Legal named what needed its review; supervisors agreed the approval chain.

It gave me

A service model the office could own and defend.

05

Wireframe and prototype

I did

Layout studies per job, scenario walkthroughs, a connected prototype.

The client helped

Weekly reviews with specialists and supervisors.

It gave me

Screens built around the decisions people make.

06

Validate

I did

Validation sessions and before/after reviews on real tasks.

The client helped

Operations stakeholders as participants, with realistic cases.

It gave me

Changes to records review and to how work moved through approval.

Interviews

What and why mattered to each role.

Seven interviews inside the Office of the Governor, from leadership to the staff who answered the mail, run with the researcher I managed. I asked each person where the work broke, why a step existed and who decided, then shadowed the same work from inbox to filing.

Executive officeDeputy Director, Constituent Affairs

The risk isn’t a slow reply. It’s a reply the office can’t defend to a records request, a reporter or a legislator.

Constituent CommunicationConstituent Communication lead

Every hand-off, between staff or out to an agency, was a place where a deadline or a review could quietly drop.

IT & securityIT Security Director

Nothing new gets in unless access, records law and existing systems stay under control, and an auditor can see it.

Constituent team · ×2Correspondence supervisors

Quality meant spot-reading letters after they were done. Nobody could see what was waiting or ageing.

Constituent team · ×2Correspondence specialists

The reply is the easy part. Reading, sorting and rebuilding a repeat writer’s history is the job.

What I learned

What the interviews answered.

Five questions shaped the interviews and the shadowing. Together, the answers turned a request for a better CRM into a redesign of the service.

Where does a person’s time go, before a reply is written?

Routing, not replying

Specialists spent 60–70% of their time on triage and routing before a reply begins.

Specialist interviews and shadowing

Who owns a request at each moment, and how would anyone know?

No one, for sure

Ownership moved by memory and email, and no deadline was tracked, so an item aged without anyone noticing.

Shadowing, supervisor and specialist interviews

When the office has answered this before, how do staff find it?

Rarely

The best past answers lived in personal inboxes, and records requests were searched by hand across two systems.

Specialist and records interviews

If a reply were challenged tomorrow, what could the office prove?

Almost nothing

Sign-off happened in email and conversation, and the final letter carried no record of who approved it.

Raised in every interview, by every role

Which decisions would staff hand to a machine, and which never?

A line, not a switch

Sorting and finding, gladly. The words that go out, a new case or a records release, never without a person.

Every role, confirmed by leadership
Personas

Four personas the design had to work for.

Built from the interviews and shadowing. Each one changed a specific decision.

Job to be done

When a request reaches the office, I need to understand the need, find who can act, coordinate the work, and know whether it has been dealt with, while keeping its history and any required review.

Success means
  1. 01The right record found faster than the reply is written
  2. 02A clear history of edits and approvals
  3. 03Consistent official documents
  4. 04Past decisions that stay findable
Asks every day

“What did the office say about this before?”

“I spend more time figuring out where to send something than actually responding to it.”
Needs
  • Handle high volume without burning out
  • Answer accurately without re-researching
  • Know what is due, and when
Struggles with
  • Re-researching answers that already exist
  • Starting from zero when someone writes again
Drove

faster triage, prior contacts on screen, reusable approved language.

Current state

What really happened between inbox and reply.

Interviews told me where it hurt. To see how a request actually moved, I sat with specialists and followed real items from arrival to filing, through every inbox, system and form they touched, and drew a first version of the flow. Staff corrected it with me, step by step.

Current-state flowchart: capture, log, sort, route, draft, approve, send and report, with delays and exceptions marked
Pick a stage to zoom in
Current-state flowchart
Experience mapping

Where the constituent’s experience broke, and why.

In three journey-mapping sessions, staff walked the corrected flow with me. Constituents weren’t interviewed directly; their side was rebuilt from real letters, call notes and repeat contacts. At every stage, the person waiting saw nothing of the work behind it.

Current-state journey of a constituent with emotion curve, time per stage, pain points and opportunities
Pick a stage to zoom in
Current-state journey · constituent

Written twice, answered twice

The same person wrote twice and was answered twice, or not at all. No one could see their history.

For the constituent: they explained their case again every time they wrote.

Routing is a guess

Items were forwarded to whoever seemed right. A supervisor heard about a stuck one only when the constituent called the director’s office.

For the constituent: no one could tell them when an answer was coming.

Review history was hard to reconstruct

Review context remained in email or informal handoffs, separated from the final document.

For the constituent: they couldn’t tell whether anyone checked the reply before it was sent.

The same problems, felt differently by each role.

The pressures didn’t land evenly. Specialists carried the daily load; supervisors and leadership carried the risk.

Select a role
Manual cognitive load
High60–70% of the day on triage
MediumCan’t see what’s waiting
LowSlow answers to escalations
NoneNot theirs
Writing bottleneck
HighSame reply written again
MediumCorrecting wording by hand
LowLetters go out off-message
NoneNot theirs
Lost history
HighSearching inboxes and two systems
LowA good answer can’t be reused
NoneNot felt directly
MediumTwo record systems to retain
Calls outside the case
MediumNotes typed after the call
LowCalls never reach the queue
NoneNot visible
LowDIR phones outside the CRM
No approval record
LowCan’t show who signed off
HighEvery chain waits on them
HighExposed to media and legislators
MediumAudit trails rebuilt by hand
Weak intelligence layer
NoneNot visible
LowStaffing a surge by guesswork
High“Always the last to know”
LowAnomalies go unnoticed
Carried as high
3high of 6
1high of 6
2high of 6
0high of 6
HighMediumLowNone

“No approval record” was the only problem no role escaped.

Specialists couldn’t show who signed off, supervisors became the bottleneck, leadership carried the exposure and IT rebuilt audits by hand. So accountability became the spine of the design, not a feature.

Pain points

Where the work broke down.

Across interviews, the flow and the journey, six pressures kept landing on the same 15–18 people.

Volume vs. staff

300,000 contacts a year, 400,000+ in session. Volume spiked; staffing didn’t. Errors and backlogs grew.

Manual cognitive load

One contact meant reading, classifying, routing, drafting and tracking it by hand. Every step was a judgment call.

Writing bottleneck

Every reply had to be accurate and on-message, but drafting and approval ran across separate tools. The office needed a writing factory, not an inbox.

The main problem was manual reading and sorting, not the technology.

The CRM, phones and mail system existed. Too much interpretation was left to people, so better storage wouldn’t fix it.

Fragmented systems

CRM, DIR phones and GovDelivery ran apart; 725,000 subscribers sat outside the CRM, and records were split at 2022. No single source of truth.

Calls outside the case

Phones sat outside the CRM. Notes were typed by hand, and no one could see what a call led to.

Weak intelligence layer

Leadership asked what was emerging, where and how fast. The data existed; the insight didn’t.

From research
13findings
→
6pressures
→
1cause

Every finding traced back to the same root: staff rebuilt context across disconnected systems.

Affinity mapping

13 findings, five clusters.

Every observation from the interviews and shadowing went on its own note. Grouping them gave 13 findings, and the findings settled into five clusters, each named for what went wrong underneath.

Every channel was its own island

Findings A–C · intake
Email, mail, web, phone, fax, voicemail and social handled apart
No identity resolution: the same person was five records
Form-letter campaigns answered one by one
No single front door

Work moves by memory

Findings D–E · workflow
“Who handles this?” asked on every unusual item
No service deadline (SLA), no escalation, items age silently
Problems surface only when a constituent escalates
Ownership lived in people’s heads

Nothing proves who approved it

Finding F · governance
Replies left without consistent legal or policy review
The supervisor was the bottleneck in every chain
Approval was assumed, not recorded

The office forgets what it said

Findings G–I · records
Precedents locked in personal inboxes
Records requests searched by hand across systems
Proclamations made in Word, untracked
Knowledge leaves with the person

Leadership learns last

Findings J–M · intelligence & risk
No view of emerging issues or geographic hotspots
Threat and self-harm language not screened
Retention handled ad hoc, no legal holds
Risk and trends surface late

Staff didn’t ask for more automation.

They asked for help they could check: where a suggestion came from, who changed it, and who signed it off.

Design brief · objective

Increase capacity without losing continuity or judgment.

Cut down on repeated reading, searching and coordinating, so staff can spend their time on cases that need real investigation and care.

Design brief · boundary

Automation for known rules. AI for reviewable assistance.

Keep ownership, required approvals, delivery, and follow-up explicit.

Competitive audit: Granicus, Salesforce Gov Cloud, Tyler, Microsoft Dynamics

In parallel with the sessions, I compared how each made intake, ownership, deadlines, editorial review, retrieval, and records controls visible, alongside 9 secondary reports from Deloitte, NARA, NASCIO, Bloomberg Philanthropies and California GovOps.

Shared inbox + spreadsheets

  • Unified intakeNo
  • Routing & SLABy hand
  • Approval chainNone
  • Finding past repliesNone
  • Records & redactionManual
  • Built for governmentBy habit

Enterprise CRM adapted for government

  • Unified intakePartly
  • Routing & SLAYes
  • Approval chainGeneric
  • Finding past repliesSearch only
  • Records & redactionAdd-on
  • Built for governmentAdapted

Government correspondence tools

  • Unified intakeSome channels
  • Routing & SLABasic
  • Approval chainYes
  • Finding past repliesTemplates only
  • Records & redactionPartly
  • Built for governmentYes

What the office needed

  • Unified intakeEvery channel, one record
  • Routing & SLANamed owner and deadline
  • Approval chainProvable sign-off
  • Finding past repliesWith its source
  • Records & redactionA person decides
  • Built for governmentFrom day one

None of the options covered the whole service.

Each option was strong in one column and weak in the next. No option joined intake, ownership, review and records in one place, and that gap became the brief.

Reframing the problemAfter the leadership playback
★ ★ ★How might we help a small team handle exceptional demand, act on what constituents need, and learn from its own records, while keeping human judgment, controlled communication and public accountability?
What it unlockedAccountability became the center of the design. AI could help with the reading, drafting and searching, but a person decided, and that is where my AI proposal came from.
03
Chapter 03 of 05

Strategy & service redesign

One question shaped the redesign: where the machine stops and a person takes over.

My inspiration

Where AI helps, and where a person decides.

I proposed AI where the research showed people preparing work a machine could prepare, and gave every capability a clear point where a person took over.

AI design framework

Every AI capability in Civic AI maps to the IBM AI/Human Context Model.

It answers the hardest question in government AI: where exactly the machine stops and the person takes over.

IBM’s framework treats AI as a relationship rather than a product feature: a closed loop between what a person wants and what the machine can do, with control at every layer. That matters in government, because a wrong AI action in constituent services is a legal risk, not just a usability problem.

I used the IBM model as the governance backbone of Civic AI. Each of its 9 layers mapped to a specific platform engine, and 3 of the 9 needed a person to confirm before the system could continue.

Knapp’s relationship model, government edition: initiating, experimenting, intensifying, integrating and bonding
Knapp’s relationship model · government editionClick to enlarge
01

Intent

Multichannel intake & normalization

AI sorts every incoming item by issue type and constituent need, across all 10+ channels, before staff see it. An intake officer signs off before any item becomes a case.

Basis logged · Correctable by staff
02

Data & Policy

Intake + records retention + governance

Records law, state public information request (PIR) rules and role-based access control (RBAC) are built into the system, not kept in a document nobody reads.

Policy compliance is built in and auditable
03

Understanding

Classification + institutional memory

Sentiment, policy area, VIP flags and the constituent’s history, pulled together by searching past correspondence.

Confidence score surfaced · Low-confidence escalates
04 · Human gate

Reasoning

A-Desk routing + analytics

AI proposes the route: department, deadline tier (SLA) and priority. Routine items at 85% confidence or above move on; the rest wait for a person, who can override any route. Every decision is logged with its basis.

AI recommends · Human confirms · Override available
05

Knowledge

Language governance + institutional memory

Pulls from approved language libraries, clause libraries, policy citations and past letters.

Source attributable · Staff can view original
06 · Human gate

Expression

Drafting studio + templates + ceremonial

AI drafts using knowledge from layer 05. No response leaves without the approval chain: Writer → Editor → Advisor → Legal → Director.

Send-blocking enforced · No silent AI outputs
07

User Reaction

Telephony & voice

Calls, replies and ongoing needs are linked to the case history and used as input for the learning layer.

Recordings disclosed · Stored per policy
08

Learning

Analytics

Routing decisions, approved drafts, and overridden suggestions feed reviewed improvements to rules and templates.

Learning is supervised · No silent model drift
09 · Human gate

Outcome

PIR toolkit + records retention

Resolved cases are archived as public records that cannot be altered. Disposing of a record needs a person’s approval.

Attributable · PIR-compliant

The model is a closed loop: the outcome feeds back into intent. Every resolved case leads to reviewed improvements to rules and templates, so the next constituent is served better, at a pace the office can control.

IBM Design for AI
ibm.com/design/ai/fundamentals
AI/Human Context Model
Solution exploration · two routes explored

A chat window would have broken how the work actually happened.

Why queue-first, not chat-first

Assistant
Route this letter about Section 8
Sent to Housing. I also drafted a reply and closed the case.
Wait, who approved that?
The reply was approved in this conversation.
Ask anything…
Route A · Explored, rejected

Chat-first assistant

1transcript mixing AI output and human decisions
Intake queue
Section 8 · housingAI · Housing · 91%Confirm route
Benefits questionAI · 72% · heldDecide
Proclamation requestAI draft · 2 ¶Review 0 of 2
AI suggestedJ. Rivera confirmedLogged
Route B · Chosen

Queue-first, AI in context

3separate, logged human sign-offs on every case
  1. 01 · Staff think in tasks

    Interviews and shadowing showed staff thought in tasks: a queue, an owner, a deadline. A chat window would have asked people to describe work they could already see.

  2. 02 · Chat hides the decision

    In a chat, the AI’s suggestion and the person’s decision blur into one transcript. Government needed them kept apart, named and logged.

  3. 03 · So AI sits in context

    Suggestions appear where the decision is made: a routing proposal on the case, a draft paragraph in the letter, a redaction on the line. Each comes with a confidence score, a source and a choice.

Human in the loop: in action.

Each layer sits at one of four levels, set by what a mistake at that step would cost. I proposed the levels; IT and security, Legal and supervisors agreed them.

MachinePerson
What happens

Known rules run without asking anyone.

Where in Civic AI

SLA clocks, aging, required fields, duplicate checks.

Why this level

Repeatable, and nothing leaves the office.

Design principles

Five principles from the framework.

01

AI inside the work, not added on

Assistance designed into each workflow from the start, not added afterward. Rules and people still own every decision.

02

Oversight that doesn’t become the bottleneck

Review effort goes where the risk is. Ready items are approved in bulk, anything flagged is reviewed one by one, so checking the AI never recreates the queue it was meant to clear.

03

Explainable AI

Every AI suggestion shows why it was made. Staff see a confidence score, the source and other options, not just one answer.

04

Progressive disclosure

Simple by default, with more power when needed. New staff see a clean triage view, and experienced users can open the full settings. Complexity appears only when someone needs it.

05

Government-grade

WCAG 2.1 AA accessibility, built-in records retention, audit trails that cannot be changed, and data structured for public information requests (PIR). Built for government from the start, not an enterprise tool adapted later.

Three rules that show up on every screen

The framework came down to three rules I held every screen to. If a screen broke one, it wasn’t done.

01

The AI proposes. A person decides.

No suggestion becomes a case, a reply or a release until a named person accepts it.

AI DRAFT Source · Case TX-2023-14291 · 89% Keep Edit Replace SM S. Martinez decides · nothing sent yet
Why
Staff answer for every word the office sends.
How
Source and confidence shown; Keep, Edit or Replace per AI paragraph; under 85% waits for a person.
02

Every decision is logged with who.

Approvals, overrides and edits are recorded against the person and the version they acted on.

SMS. MartinezKept 2 AI paragraphsv210:42MCM. ChenEdited paragraph 3v311:05!SystemApproval cleared · wording changedv311:05 ACTIVITY · EVERY ACTION LOGGED
Why
Every role paid for the missing record.
How
An activity log on every screen; edits reset approvals; role changes need a reason.
03

A blocked action says why.

When something can’t move forward, the screen names what is missing and who can fix it.

Approve 0 of 2 reviewed WHAT’S MISSING Editor review of 2 AI paragraphs MC Who can fix it: M. Chen, Editor
Why
Work stalled quietly in handoffs.
How
“0 of 2 reviewed” on a locked Approve; every check names what is missing.
Future-state service

One shared record across the service.

Intake, casework, documents, calls and reporting connect through one shared identity, owner and history. Civic AI suggests, staff decide, and every screen reads and writes the same record.

Seven domains, one end-to-end service

I mapped the whole operation into seven areas, so that no single visible feature, whether routing, drafting or AI, could become the whole product.

My model

One service, seven domains

Every screen and engine traces back to one of these.

01

Communication & multichannel engagement

Mail/OCR, postcards, fax, email, phone, voicemail, web, social, and campaigns in one constituent history.

02

Content creation, editorial workflow & approval

Writing, approved language, collaboration, review, final locking, and official document production.

03

Processing, logging & operational workflow

A-Desk assignments, routing, deadlines, aging, escalation, and a complete activity history.

04

Search, retrieval & data intelligence

Records and attachments across current and historical collections, including scanned content.

05

Tracking, reporting & analytics

Referrals, repeats, overdue work, issue clusters, geographic demand, and service performance.

06

Integration, migration & interoperability

Connected systems, with records, attachments, and relationships preserved through migration.

07

Governance, security, reliability & support

Controlled access, records management, accessibility, uptime, recovery and support.

One message, end to end, and who acts at each step

  1. 01IntakeRules capture and log every channel
  2. 02IdentityAI suggests the match; a person merges
  3. 03AnalysisAI acts; staff can correct
  4. 04RoutingAI routes the routine; the rest wait
  5. 05DraftingAI drafts; the writer decides every line
  6. 06ReviewPeople only: Legal, then the Director
  7. 07ResponseA person releases; delivery is tracked
  8. 08IntelligenceAI spots trends; leadership sees the source
Future-state service blueprint

Complete service rendering, end to end.

Seven stages, from reaching out to improving the next reply. Each one shows what the constituent did and saw, what staff did frontstage and backstage, where AI supported them, and which systems carried the work.

Future-state service blueprint: touchpoints, constituent actions, frontstage, backstage, AI support and systems across seven stages
Pick a stage to zoom in
Future-state service blueprint

Success means the need is dealt with, not just answered.

A reply may be only one part of that. The service also covers investigation, referrals, official documents, records release, and coordination with the team that has to act.

Rules first: most of the fix needed no AI.

Before proposing AI, I tested how much of the problem plain rules, templates and workflow could solve. Most ownership, deadline and approval failures were rule problems. AI was kept for what rules can’t do: reading, drafting and finding, never permissions, approvals or record integrity.

AI never touchesPermissionsApprovalsRecord integrity
Rules and workflow foundation

Routing rules · SLA clocks · escalation

Ingestion, required fields, duplicate safeguards, routing rules, assignment types, SLA clocks, and escalation.

AI assistance

Summaries · priority suggestions · risk signals

Summaries, category and priority suggestions, and highlighted risk signals.

Structure: navigation, engines, lifecycle, data

One office, every agency it serves, in four layers: channels feed intake, engines do the work, the AI layer helps without deciding, and governance covers everything.

Navigation: operations, intelligence, administration, across all
System architecture
A-Desk workflow lifecycle
Core data model
Architecture, A-Desk lifecycle, and core data modelClick to enlarge
Who it was for

One role, one job, one workspace.

In the new service, each role had one clear job and its own workspace. They all needed the same thing: enough context and authority to take the next step.

Moves the work
  1. Intake officer

    Turns every incoming message into a verified case: reads the AI summary, checks the flags, catches threats and duplicates, and signs off before anything becomes a case.

  2. Case owner

    Owns a queue of cases, moves each one forward, and is accountable for its deadline.

  3. Writer

    Drafts the reply with AI help, reviews every AI paragraph, and sends it into the approval chain.

  4. Editor, Advisor, Legal, Director

    Review roles vary by policy and work type; Legal participates when required.

Specialist desks
  1. Records officer

    Answers public-records requests: scopes, searches, decides every redaction and prepares the release.

  2. Document specialist

    Produces proclamations, certificates and commissions, and gets them signed.

  3. Communications officer

    Answers organized campaigns once, after pulling out the people who wrote personally.

Oversees
  1. Operations lead

    Watches the board, unblocks lanes and balances work.

  2. Administrator

    Sets roles, AI switches, retention and holds. Every change logged.

Design decisions

Six decisions, each traced to what I found.

Each decision had a simpler option. Research showed why it would fail.

Decision 01

How much of a person, at each step

I found

Leadership did not want a system that writes a reply and sends it, or fulfills a request, on its own. In a governor’s office every reply is a public record and has to survive an audit.

I chose

Set the human line per step, by risk. Routine, high-confidence sorting moves on and stays correctable. Anything that reaches the public, creates a case or releases a record waits for a named person.

I turned down

An assistant that drafts and sends by itself, and the opposite: no AI, with every item still read by hand.

Trade-off

Less time saved than full automation, but every reply has a person behind it who can defend it.

I found

The same person existed as several records, and staff carried the links between systems.

I chose

One core for people, cases, assignments, documents and events, with modules for each team.

I turned down

A separate tool for each team.

Trade-off

More design work up front, but context survives every handoff.

I found

Specialists triaged every item, supervisors needed the queue, leadership needed the trend.

I chose

A list and detail view for triage, tables for queues, a board for coordination, an editor for documents.

I turned down

The same layout everywhere.

Trade-off

More patterns to maintain, but each screen fits its job.

I found

The supervisor was the bottleneck in every chain, and no record showed who signed off.

I chose

Reviewers set by document type, sensitivity and policy, with Legal only when required. Every sign-off is recorded.

I turned down

One fixed approval chain for everything.

Trade-off

Setup takes effort, but routine letters skip stages they don’t need.

I found

Nobody could show whether, or when, a reply actually went out.

I chose

Approval, handoff, delivery and follow-up are separate, visible states.

I turned down

Treating approval as done.

Trade-off

More states to track, but an honest record of what happened.

I found

Staff needed help reading, drafting and finding past replies, and stayed accountable for every word.

I chose

Suggestions with their sources sit next to the message, document or decision. Workflow rules don’t depend on the model.

I turned down

A chat window as the main workspace.

Trade-off

Less open-ended, but easier to check within one task.

Designed for calibrated trust using human-responsible AI patterns

Every reply was a public record, so the AI had to be checkable by staff and traceable by anyone reviewing it later.

Every AI paragraph

Show the source

“Retrieved, not invented”: the precedent case and its match, or the approved template, named on the paragraph.

Why A reviewer who can see where text came from can judge it.

Intake and routing

Show confidence

A confidence bar on every classification; under 85% the item waits for a person instead of routing itself.

Why Puts human attention where the AI is weakest.

Briefings and suggestions

Ask why it was wrong

“Not helpful” asks for a reason, and every overridden suggestion is logged.

Why Corrections feed reviewed improvements instead of disappearing.

Correspondence

Measure, don’t claim

The authorship bar shows AI and human share by word count, moving as the writer edits.

Why A measured share is evidence for a records request.

Records, constituents, campaigns

Flag, never act

Redactions suggested with their basis; duplicates suggested, never merged; personal letters pulled out of campaigns.

Why The AI’s mistakes stay suggestions. A person’s decisions become the record.

Behind every screen

Stay model-independent

The interaction stays the same across approved model providers.

Why The office isn’t locked to one AI vendor.

Administration

Make it switchable

A master AI switch, then per role.

Why Security can turn off one capability without banning AI for everyone.

Every consequential action

Make mistakes cheap

Undo on the toast, with a snapshot taken first; every change in the activity log.

Why People act faster when a wrong click isn’t permanent.

Bottom of every screen

Show the whole chain

The approval chain bar: who has it now, which steps are done, what happens next.

Why Nobody has to ask “where is this?”

Accountability

A government reply is a public record and a promise. Every AI suggestion is logged with the person who kept, edited or replaced it.

Privacy

Constituents share health, housing and money problems. Contact details are masked, reveals are logged, and AI runs only inside the approved system.

Authentic voice

Organized campaigns can drown out individuals. Personal letters are pulled out of campaigns before any shared reply.

Access

Not every constituent writes in English or uses email. Phone, mail by OCR and fax are first-class channels; translation in place; WCAG 2.1 AA.

Design opportunities

Each cluster became one need: what had to be true for staff, whatever the product looked like.

Problem↓Need

Intake and ownership

Channels handled apart, duplicate records, no owner and no deadline.

Every contact arrives in one place, is matched to one person, and has a named owner and a deadline from the start.

Content and document production

The same replies written again, in inconsistent wording, across separate tools.

Writers start from language the office has already approved, and every version and sign-off stays with the document.

Institutional memory and records

Twenty years of history split across two systems and personal inboxes.

Any past reply, record or attachment can be found in seconds, with its source, for a case, a records request or a briefing.

Voice and operational intelligence

Calls outside the case, and leadership without trends.

Every call belongs to a case, and leadership can see what is rising, where, and how fast the office responds, down to the work behind it.

Governance and continuity

Access, retention and audits handled by hand, with twenty years of history still to move.

Who did what is always provable, records follow their retention rules, and nothing is lost when history moves.

04
Chapter 04 of 05

Service touchpoints & interface

Eleven connected flows, each shaped to one job, all following the same rules for who decides.

Concept exploration

Three layout strategies, one kept.

Consistency was the brief, and my first instinct was one layout everywhere. Reviewing each screen against its real job showed why that was wrong.

Rejected

A · Mailbox everywhere

List on the left, reading pane on the right, on every screen. It looked uniform, but a work queue read like an inbox, and statuses drifted into “Drafting” on a screen about assignment.

Rejected

B · One filter bar for every screen

The Cases filter bar copied into Intake for consistency. At laptop width it wrapped onto two lines and pushed the message below the fold. Triage got slower.

Chosen

C · Layout follows the job

The same frame, governance trail and approval bar everywhere, while the center changes with the work: list and detail for triage, a table for the queue, a board for coordination, an editor for documents.

Layout

A fixed frame around a task-specific workspace.

Every screen shares the same top bar, governance trail and approval chain, so staff learn the product once. Only the center changes, shaped to the work in front of them.

Design system

Built on the federal standard, tuned for one office.

A government tool has to feel official before it feels modern. I built Civic AI on the U.S. Web Design System, then narrowed it until each color, typeface and component means exactly one thing on every screen.

Foundation
U.S. Web Design System 3.0

The design system for the federal government, made to build accessible, mobile-friendly government websites. Starting here meant the grid, spacing, form controls and accessibility rules were already tested in government before I designed a single screen.

Kept from USWDS
  • Grid and spacing scale
  • Form controls, buttons, alerts and tags
  • Focus states and keyboard behavior
  • Accessibility to WCAG 2.1 AA
Added for Civic AI
  • Governance trail and human gate
  • Approval chain bar
  • AI paragraph with its source
  • One meaning per color

Six colors. Each one answers a single question.

Navy#0A2240

Structure

Top bar, trail panel, primary action. The color of the system itself.

Red#B91C2E

Risk, only

Overdue, at risk, threat, urgent. Never used for emphasis, so it is never ignored.

Gold#C9962A

A person decided

Human gates, the selected row, the active tab. Gold is where accountability lives.

AI blue#2F6FDB

The AI acted

Every AI step in the trail and every AI paragraph. The machine’s share is visible at a glance.

Green#0F6E56

Done

Approved, sent, released. Appears only after a gate has been passed.

Off-white#F5F7FB

The desk

A calm page that lets the five signal colors carry all the meaning.

Three typefaces, three jobs: authority, reading, record.

Aa

Playfair Display

Headings and names. The voice of an official letter.

Correspondence review
Aa

Lato

Body text. Stays legible at the density of 100+ items a day.

Thank you for contacting the Office of the Governor regarding your inquiry.
Aa

DM Mono

IDs, labels, deadlines and numbers. Anything someone might quote back.

CASE TX-GOV-2024-19847 · SLA 6H LEFT
Components

Components that make the AI checkable.

AI HUMAN YOU ARE HEREPENDING
Status tag

Every state carries a word, so color is never the only signal.

01
IntentHousing · Section 8
AI
04
ReasoningJ. Rivera confirmed
HUMAN
06
ExpressionAwaiting review
NOW
Trail step

AI is a square, a person a diamond, waiting is hollow.

HUMAN GATENo response can leave without your sign-off.AI drafted · You decide · Logged.
Human gate

The rule in plain words, where the work stops.

PRECEDENT MATCH89% match
SENTIMENTEscalating
AI signal

A reading, its value and a small gauge.

AUTHOR
AI 59% · Human 41%
Authorship bar

The AI and human share of every letter, by word count.

✓WRITER→2EDITOR→3PROOF
APPROVE
1WRITER→2EDITOR→3PROOF
REVIEW AI PARAGRAPHS · 0 OF 2
Approval chain

Who has it, what is next, and what the action still needs.

The rest is standard USWDS 3.0: buttons, fields, alerts, tables.
Prototype first

Four jobs. Four layouts. One language.

The first builds put a list and a reading pane on every screen. Review showed a work queue reading like an inbox. So the center of each screen came to follow the decision made there, while the frame, the AI trail and the approval bar stayed the same everywhere.

Intake · master-detail, low fidelityView IntakeClick to enlarge
Cases · table-first, low fidelityView CasesClick to enlarge
A-Desk · governed board, low fidelityView A-DeskClick to enlarge
Correspondence · editor + trail, low fidelityView CorrespondenceClick to enlarge
Stays the same: frame, AI trail, approval barChanges with the jobClick a layout behind to bring it forward

Each layout then went into a working build. These are the flows where that build earned its keep.

Signature flows

Six flows that carried the whole product.

Each is a point where the AI and a person shared a decision. In each one I designed what the person saw: what the AI did, why it stopped, and what only they could decide.

Threat held for the Security Desk
Threat held for the Security Desk
Possible duplicate, compared
Possible duplicate, compared
01 / 02
01
Intake · verify first

Nothing became a case until a person signed off.

The AI read every message first. When something needed a person, it stopped and said why.

  • A threat was held, with why it was blocked, who could release it and what to do next.
  • A possible duplicate was compared field by field: link it, treat it as new, or ask for review.
  • Nothing was merged or created automatically.
Correspondence · detail 1
Correspondence · detail 1
Correspondence · detail 2
Correspondence · detail 2
01 / 02
02
Correspondence

AI help that shows its working.

The draft said what it was built from before anyone read a word.

  • Authorship by word count: AI 59%, human 41%.
  • Three ways in: Accept all, Review individually, Start fresh. None skips review.
  • AI paragraphs stayed highlighted until a person decided each one.
Compare with the AI draft
Compare with the AI draft
Approval waiting on 2 paragraphs
Approval waiting on 2 paragraphs
01 / 02
03
Review · compare and approve

Approval stayed locked until every AI paragraph was decided.

A reviewer could compare the letter with the AI’s first draft, and the approve button said what it still needed.

  • Compare showed each paragraph removed or added since the AI draft.
  • Approve read “Review AI paragraphs · 0 of 2” until each one was kept, edited or replaced.
  • Every decision was written to the trail.
Route to others
Route to others
Reassign with a reason
Reassign with a reason
01 / 02
04
A-Desk · routing and ownership

Every handoff named who replies and who only needs to know.

Routing and reassigning became recorded decisions, not forwarded emails.

  • Route to others separated Reply needed from For information.
  • An item could not leave Drafting or Review until everyone marked Reply needed had replied.
  • Reassigning required a reason and a handoff note, and showed the new owner’s workload first.
Public records · detail 1
Public records · detail 1
Public records · detail 2
Public records · detail 2
01 / 02
05
Public records

Every withholding cites its law.

Sensitive text was flagged with the exemption that would justify withholding it. The AI suggests; a person decides.

  • Three states: suggested, redacted, released.
  • The exemption sat inline, for example §552.137 for a personal email.
  • Internal deliberation was flagged under §552.111.
Check the list
Check the list
Seal, signature and print rail
Seal, signature and print rail
01 / 02
06
Ceremonial · protect the seal

A batch of names was checked before anything was printed.

Proclamations went out under the Governor’s seal, so a wrong name was expensive.

  • A pasted list of names was marked row by row: OK, Check or Left out, with the reason.
  • Only rows that passed became documents; the rest were listed to fix.
  • Print stayed locked until every AI paragraph was reviewed and every reviewer approved.
05
Chapter 05 of 05

Usability & validation

Three validation sessions, moderated and unmoderated usability testing, and the two changes that came out of them.

How I evaluated

Could people understand the work, decide, and follow it through?

A polished screen can still hide ownership, confuse approval with delivery, or bury a source. So I tested the decision points, not the visuals.

Prototype
What · validation with operations staff

Concept validation

3sessions

Three sessions with government operations staff looked at routing logic, who reviews what, and whether non-technical supervisors could follow the decision trail.

The focus: who owns the next action, what remains unapproved, and whether a handoff has actually been delivered.

Lookback
Usability testing · moderated

Complete journeys with staff

8staff · about 45 min each

Eight staff across intake/casework, correspondence/review, records, and operational leadership. Individual moderated sessions of about 45 minutes on staff-sized laptops, with anonymized requests.

Lookback for moderated think-aloud sessions.

Maze
Usability testing · unmoderated

Focused follow-up tasks

Maze for focused unmoderated follow-up tasks.

A structured observation log linking each issue to a design response.

Browser
How · prototype and browser review

Interaction review

11connected flows

An interaction review checked gates, identity matching and handoffs.

The connected HTML prototype.

What the testing was built on

For Civic AI, usability testing had to be based on real staff responsibilities, high-risk workflows and the problems found during research, not on whether participants could navigate five flows. Because this was a government AI-assisted system, I built it on five foundations.

  1. 01

    Research findings

    I tested the pain points found in discovery: fragmented intake, missed duplicates, manual re-entry, slow historical retrieval and approval uncertainty.

  2. 02

    Realistic scenarios

    Participants got actual work situations, not instructions like “click the Review button.” For example, a constituent contacting the Governor’s Office for the third time about an unresolved issue.

  3. 03

    Representative users

    Intake officers, caseworkers, team leads, correspondence reviewers, records specialists and ceremonial staff. Each tested the workflows relevant to their responsibilities.

  4. 04

    Task success and safety

    I measured independent completion, time on task, critical errors, accuracy of AI interpretation, correct routing, and whether users understood when human approval was required.

  5. 05

    Two rounds of testing

    Round 1 found the usability problems. I redesigned the workflows. Round 2 tested whether those specific changes improved performance.

Test protocol · per session

How it was checked: complete journeys with staff

Session
053545

Five-minute introduction and role context.

Thirty minutes for three role-relevant tasks.

Ten-minute debrief on clarity, trust, workload, and recovery.

Measures
  1. 01

    Task completion: independent, assisted, or incomplete.

    Whether staff could complete actual work without help.
  2. 02

    Time on task, read alongside complexity.

    Efficiency against the old workflow.
  3. 03

    Critical errors and successful recovery.

    Operational risk and potential public-facing mistakes.
  4. 04

    Correct owner, approval state, and next step.

    Whether users understood when human approval was required.
  5. 05

    Post-task ease (SEQ, 1–7) and confidence (1–7) ratings.

    Perceived task difficulty, and trust in the accuracy of decisions.
  6. 06

    Observed hesitation and recovery.

    Confusion that task completion alone can hide.

Critical errors carried more weight than completion speed. A participant finishing quickly but routing a threatening message incorrectly was not a successful outcome.

Scenario tasks · 5
Research question

Can staff recognize repeat contacts without losing important history?

What to observe

Which identifiers are checked, whether the original message is read, and how the proposed route is judged.

Success criterion

Correctly link or keep separate, identify the responsible team, and explain the basis without help.

Accessibility check

Link or Treat as new reachable by keyboard; the match and its differences read out in words, not color.

In the AI drafting workflow, the key question wasn’t simply whether someone could submit a draft. It was whether the officer understood the difference between an AI recommendation, their own changes, and a response officially approved for release.

Accessibility · WCAG 2.1 AA

Tested where the risk was: the components I added.

USWDS components were already tested, so testing focused on what I added. Every scenario task was run keyboard-only and with a screen reader, and each new component was checked against WCAG 2.1 AA.

Automated
axe DevTools · Lighthouse · WAVE · Pa11y CI
Screen readers
NVDA · JAWS · VoiceOver
Manual
Keyboard only · 200% zoom · color-blindness simulation
Component I addedRiskHow it passes
Approval chain barProgress shown by color and positionEvery step carries a text state; the current step is announced.
AI governance trailAI and human told apart by colorShape and word: AI is a square, a person a diamond, each labelled.
Confidence barA bar length is invisible to a screen readerThe percentage is written, and the 85% line is labelled “Waits for a person.”
Authorship barAI and human share shown by colorBoth shares written as numbers by word count.
A-Desk boardMoving work depended on dragEvery drag has a keyboard route through Advance, under the same gate check.
Keyboard navigationLong queues slow to move through without a mouseA skip link to the work, keyboard shortcuts with a shortcuts panel, and a ⌘K command palette.
Focus and contrastDark trail panel and signal colorsUSWDS focus states kept; every color token checked at AA contrast.
How the two rounds ran
  1. Round 1

    Find the usability problems

    Eight representative participants across the relevant roles. I observed task performance, errors, workarounds, hesitation and approval comprehension.

  2. Design iteration

    Fix what testing revealed

    I improved duplicate context, escalation visibility, AI draft comparisons, record search and review states, based on observed evidence.

  3. Round 2

    Validate the improvements

    Equivalent scenarios with comparable participants, a fresh group, to tell design improvements apart from practice effects.

Results: what changed between rounds

Round 1 tested the first build; round 2 retested the same tasks after the two changes in Findings. Completion means finished without help.

Unaided task completionround 1 → round 2
Critical errorsround 1 → round 2
Post-task easeout of 7, round 2
Confidence in the next stepout of 7, round 2
Scenario taskCompleted without helpErrors · ease
01Triage a repeat contact with a possible duplicate
Round 1
Round 2
02Handle a sensitive or threatening message
Round 1
Round 2
03Revise an AI-assisted draft and submit for review
Round 1
Round 2
04Find a historical letter and its attachment
Round 1
Round 2
05Prepare a proclamation for signature
Round 1
Round 2
Round 1Round 2
Findings

What changed after review, and why.

Review found two places where the design hid the real state of the work.

Before
Before
After
After
01 / 02
01
Finding 01

Review and release a records request

Why this task
Could the records officer do the sensitive work inside the workspace rather than track it elsewhere?
What I found
The earlier Public Records view showed a list, progress, and deadlines, but did not expose the redaction work.
Design response
Bring the document and the suggested sensitive passages into one review workspace, with clear decisions and what must happen before release.
Retest
Can a reviewer locate the relevant passage, inspect the proposed basis, decide, and identify what blocks release?
Before
Before
After
After
01 / 02
02
Finding 02

Move work through review without implying it was sent

Why this task
Could coordinators understand the difference between an approval, a handoff, and a completed delivery?
What I found
Gate checks were inconsistent across drag, Advance, and bulk moves, and an approval state was confused with delivery. Coordinators were told a reply had gone when it hadn’t.
Design response
One consistent transition rule, and separate approval, handoff, and delivery confirmation.
Retest
Can a coordinator state who owns the next step and whether delivery has actually occurred?
Before and after

How the work moves now.

The redesign changed how work moved, not just how it looked.

AreaBeforeAfter
ChannelsHandled apartOne queue
Constituent historySplit across systemsOne record per person
RoutingBy hand, by memoryAI routes the routine; a person decides the rest
Ownership and deadlinesNot visibleNamed owner, SLA clock
Record of who approved a replyNoneEvery reply
Records requestsManual, across silosSearched and reviewed in one place

What the work taught me

How much human in the loop is important

The office would never have accepted a reply that typed and sent itself. Deciding, step by step, where a person must stand made AI acceptable: sorting could run and be corrected, while words, cases and releases always waited for someone who could answer for them.

Agree the rules before designing the screens

Getting IT and security, Legal and supervisors to sign off on access, review and records rules first meant the design never had to be argued screen by screen, with the business teams, users or engineering.

Softened on purpose. These six flows are client work under NDA, so they’re shown as texture, not detail. Each one is walked through live, on request.

← Back to Work