Intuit · Sr. Content Designer · 2023
An AI Bill of Rights
We were asking people to let AI act on their money, so before we wrote a single prompt, we defined what it would never be allowed to do.
Fintech · AI integration- Problem
- In 2023, customers were being asked to let AI act inside their own books, and the biggest question was what it should be allowed to do without asking first.
- Decision
- Write an AI Bill of Rights before shipping anything, setting limits on what AI would do with financial data. Then score generated drafts on a Likert scale so feedback was based on data, not opinions.
- Complications
- The models at the time were only good enough to sort output into bad and not bad. After a layoff, I also lost access to the document and the reasoning behind it.
- Outcome
- Product teams adopted the principles for governance on Intuit Enterprise Suite, a product line I never worked on.
Setting limits for AI before anything shipped
Intuit’s public guide for writing AI content. The internal principles came first.
In 2023, the models were new, there was a lot of hype, and customers were being asked to let software act on their behalf inside their own books. People weren’t worried about tone. They wanted to know whether a company with access to their bank feed would make moves without asking them.
So our content design team wrote an AI Bill of Rights, a statement to leadership about what AI in our products would and wouldn’t do with someone’s financial data. It was a set of limits, written before the features they applied to had shipped.
Measuring quality with a Likert scale
- Your books are looking great! Let's keep the momentum going.2bad
- We sorted 42 transactions. Check the 6 we weren't sure about.4not bad
- Effortlessly manage your finances with smart automation.1bad
- This expense looks personal. Move it out of business?4not bad
These are representative examples I rebuilt, since I no longer have access to the original spreadsheet.
Getting there took a lot of testing, collaboration, and spirited debate. What helped most was giving everyone a shared way to evaluate the output.
We put every generated draft on a Likert scale in a very large spreadsheet, scoring quality, voice, and tone one row at a time. Once stakeholders were looking at numbers, feedback shifted from “I don’t like it” to “I scored it a two, and here’s why,” which made for much more productive conversations.
In 2023, the results were pretty binary: output was either bad or not bad. The models weren’t good enough yet for finer distinctions, and being upfront about that helped us move forward and ship.
The principles went on to guide a product I never worked on
Product teams adopted them when they built governance for Intuit Enterprise Suite, the company’s move upmarket from small businesses toward larger, more complex ones. I wasn’t part of that work, but our principles were.
For me, that’s the real outcome: our work traveled to a part of the business none of us had worked on.
What I learned from losing access to the principles
The principles are behind a login I lost in a layoff. I helped write them, but I can’t fully reconstruct them. I could approximate the rules, but the reasoning is gone: why we drew each line where we did, which debate each one settled, and what we were worried about at the time.
This kind of loss is common in tech, and very few teams plan for it. Institutional memory usually lives with a handful of people who still work there, and when they leave, the reasons behind decisions go with them.
Since then, everything I build keeps its decisions and the reasoning behind them in plain files I own. Every rule carries its why, because I know what it’s like to lose it.
Project details
- Company
- Intuit, QuickBooks
- Role
- Sr. Content Designer
- Timeline
- 2023
- Tools and methods
- Content principlesPrompt evaluationLikert scoringDesign system contributionCross-functional governance