Unshipped concept

Dara and Connie: controls for trustworthy AI.

After the redesign, I explored how two AI assistants, Dara and Connie, could help with payments across New York and New Jersey without adding financial risk. These assistants were internal concepts we used for testing; they didn’t ship, so none of the redesign’s results apply here. What shipped was guided AI billing, which did better than chat in those prototype tests. The screens below show Dara.

Open chat with Dara

Billing help screen with the Dara chat button in the bottom right corner
Dara closed: a chat button in the corner.
Dara chat open on a phone, answering whether the $161.50 bank payment has posted
Dara open: the same question asked in chat.
Dara chat open on the E-ZPass website beside the bill and payment details
Dara on the website.
How I thought about itWhat the assistants could and couldn’t do, and whyOpen overview

An assistant is only as reliable as what surrounds it. What it knows about the account, what it’s allowed to do, and what happens when it’s wrong decide whether people can trust it.

Why this matters to the business

A wrong answer here isn’t just bad copy. It can mean someone pays twice, misses a deadline, or has their account changed without asking. So the goal was never “add a chatbot.” It was to handle more routine questions without putting anyone’s money at more risk.

My principle

I let AI give me more options than I’d come up with alone. Then the payment rules, accessibility and each state’s policies decide which one ships.

Explore the AI systemSeven pieces: what it knows, what it’s allowed to do, what it remembers, its screens, my review, its tests, and how it all fits Open collection
1
Context packWhat the assistant gets before it answers anything.
View artifact
Artifact 1 / Context pack

Give the assistant the right context before asking it to help.

A good model with the wrong context still gives bad answers. So I put what we already knew in one place the assistant reads before every answer: the E-ZPass research, the payment rules, where New York and New Jersey differ, how we sound, and the accessibility rules. The design system stops being something only the team can see. The assistant follows it too.

Why it matters Without it, the assistant can sound sure of itself while ignoring the account, the state, what the driver agreed to, or what a wrong answer costs them.

Open artifact image
E-ZPass Assistant/ AI Context Pack CP-01 · v1.0 Draft for review
Assistant for violation questions · New York & New Jersey

E‑ZPass AI Context Pack

Pending legal review
Owner Product designReview CX · Legal · AccessibilityScope NY + NJ account supportUpdated May 2026
Main question
“I paid this toll yesterday. Why do I still have a violation?”
Who
A driver short on time or money
Goal
Understand the charge and take the safest next step
Biggest risk
Paying twice, or missing a deadline
What it gets with every question
  • StateWho issued the notice: NY or NJ
  • AccountIs this the account holder, and is the account in good standing
  • PaymentSubmitted, processing, posted, rejected, reversed
  • ViolationNotice number, status, amount, due date
  • RulesThe current support rules for that state
01 / Check first
  • Confirm whether NY or NJ issued the notice before applying any rule.
  • Confirm it’s the account holder before talking about their account.
  • If either is unclear, ask only for safe details and pass it to a person.
02 / Every answer
  • Says what’s known and where it came from.
  • Keeps payment status and violation status apart.
  • Says what isn’t confirmed yet.
  • Offers one safe next step, and never says it’s resolved when it isn’t.
03 / Accessibility
  • Plain language with short, clear headings.
  • Status is never shown by color alone.
  • A summary a screen reader reads in order.
  • Confirm and hand-off buttons work from a keyboard.
04 / Never
  • Make up a balance, date, rule or outcome.
  • Promise a violation goes away, or give legal advice.
  • Change an account, dispute or payment without a clear yes.
  • Suggest paying again while the first payment is still open.
Every answer, in this order
  1. StatusWhat the record confirms
  2. MeaningWhat that means, and what it doesn’t
  3. Not sure yetWhat can’t be checked yet
  4. Next stepThe safest thing to do now
Release criteria

The driver can say what happened, tell what’s known from what isn’t, avoid paying twice, and pick a next step without guessing.

4 of 4 criteria Pending sign-off
2
Action boundariesWhen it can answer, when it has to ask, and when it has to stop.
View artifact
Artifact 2 / Action boundaries

Telling it not to isn’t enough. It has to be unable to.

Being able to talk about something isn’t the same as being allowed to do it. The assistant can explain and ask questions. Anything that changes an account or moves money needs a confirmed identity, the driver’s clear yes, and sometimes a person.

Why it matters A rule written in the instructions can still be broken. With money involved, the assistant only gets the access it needs, every action is confirmed on screen and logged, mistakes can be undone, some cases always go to a person, and the system itself blocks anything outside those lines.

Open artifact image
E-ZPass AI assistant action boundary map showing automatic answers, information gathering, confirmation gates, human escalation, blocked actions, and audit logging
3
Memory and contextWhat it keeps, what it checks again, and what it forgets.
View artifact
Artifact 3 / Memory model

Remember enough to help, not enough to create a new risk.

Not everything in a conversation should be remembered. Each thing the assistant keeps has a reason and an end date. A long conversation gets cut down to the few confirmed facts it needs to carry on. Anything that might have changed, like a payment status, gets checked again. And the driver decides whether anything is kept after the conversation ends.

Why it matters What it remembers is a design decision, not a backend setting. Remembering saves people from repeating themselves. Remembering too much can expose account details, carry a wrong guess forward, or shape a later answer after things have changed.

Open artifact image
E-ZPass AI assistant memory and context model showing policy information, customer input, verified account records, session-only memory, context compaction, expiration, and approved retained context
4
Chatbot statesFive screens: sure, not sure, needs a yes, something broke, a person takes over.
View states
Artifact 4 / Chatbot states

Show uncertainty before it becomes a financial mistake.

I carried those rules into the five moments that matter most. The assistant explains what’s confirmed, says what isn’t, asks for a clear yes before changing an account, recovers when something breaks, and hands the whole case to a specialist when it should.

Why it matters A confident wrong answer can mean paying twice or missing a deadline. The assistant still has to help when information is missing, a change needs permission, a system fails, or a person needs to step in.

5
Human reviewAn AI draft that failed, what I changed, and the version I approved.
View review
Artifact 5 / Human review

AI drafts. I decide what ships.

I treated every AI draft as a suggestion. Here’s one from start to finish: the draft, my review, the business rules it broke, my fix, and the final accessible version.

Why it matters A fast draft only helps if the review catches what’s wrong with it. This one claimed things the record didn’t support, including that a violation would be removed. Catching that is my job, along with the cases the AI never thought of.

Open AI draft Open designer review
01 / AI draft blocked by business rules
AI-generated E-ZPass response with unsupported payment, violation removal, and fee claims flagged and blocked from customer delivery
02 / Designer review and approved correction
Designer-reviewed E-ZPass response showing corrected claims, uncertainty disclosure, safe next action, accessibility checks, and approval status
6
Evaluation scorecardThe test cases, what passing means, the results, what still blocks release, and who owns it.
View scorecard
Artifact 6 / Evaluation

Define “good” before measuring the assistant.

I turned the goals into pass or fail checks. Every answer goes through the same steps: automated checks, an expert review, a check that it says what it doesn’t know, then final approval. Sounding natural doesn’t pass. An answer passes only when the facts, the permissions, the accessibility, the uncertainty and the handoff are all right.

Why it matters AI makes answers faster to produce. Checking them is what makes them safe to show a driver. Clear pass rules, review in stages and a named owner for every problem catch confident mistakes before a customer sees them.

Open artifact image
E-ZPass AI assistant evaluation scorecard showing test scenarios, acceptance criteria, results, severity, evidence, owners, and unresolved release blockers
7
End-to-end systemOne driver’s question from start to finish, including what happens when things go wrong.
View system
Artifact 7 / End-to-end system

The interface is one part of the product.

This map shows everything behind the chat window: what the assistant knows, what it’s allowed to do, what it remembers, how it’s tested, what happens when it fails, and which person owns each exception. Each part answers a different reason someone might not trust it.

E-ZPass violation-resolution journey map showing authority identification, payment context checks, permission decisions, uncertainty, duplicate-payment risk, human escalation, system failure, and safe resolution

People can trust the assistant when everything around it makes the safe answer the easy one, and a person stays responsible for the exceptions.

Open journey map