When to use this
After we shipped Amazon’s Employee Document Management platform, I led a second phase focused on adding AI to four parts of the product: an M&A dashboard, bulk document generation, the HR admin portal, and the employee portal. The platform worked well, but it mostly reported what had already happened. People still had to figure out what to do about it on their own.
This approach makes sense when a wrong answer is expensive: HR, legal, finance, healthcare, or anything that affects someone’s job, pay, or rights. You’ll usually spot the need in a few familiar places. People are interpreting raw data by hand, sending the same work back and forth to fix errors, or reading a whole document to find the one thing that’s wrong.
If mistakes are cheap and speed matters more than accountability, you probably don’t need all of this.
The steps
-
Start from a research finding, not a model capability
I ran workshops with cross-functional partners to figure out where AI could actually help each type of user. The rule was that every idea had to connect to a research finding we’d validated or something already on the backlog. The question we kept coming back to was: what does this person need right now that the system isn’t showing them? By the end, you want a short list of concepts, each tied to the finding it addresses.
-
Tell people what to do, not just where things stand
Managers tracking paperwork for acquisitions had one number to go on: a completion percentage. “57% complete” doesn’t tell you much. You still have to work out whether that’s a problem. We added a risk score with a plain-language explanation, a close-date forecast based on how fast things were moving, and anomaly detection that called out the actual blocker (for example, three Critical employees who hadn’t been released yet) instead of leaving it buried in a table. A good test is whether every number on the screen answers “so what do I do?”
-
Show the reasoning and the confidence, and let people choose
Bulk document generation used to mean a lot of back and forth. HR would upload a spreadsheet, get a list of errors, download it, fix things offline, and upload it again. The self-fix feature finds each error, suggests the most likely fix, explains why, and shows how confident it is. HR can accept or reject each one inline. About 75% of common errors can now be fixed without leaving the page. My rule here: no AI output shows up without an explanation next to it.
-
Be upfront when the AI can’t help
Some rows can’t be fixed automatically. We flag those clearly, say why, and let everything else go through. So if a batch has 72 employees and 8 errors, the other 64 documents still get generated. In compliance work, an AI that guesses wrong creates legal risk, so the fallback needs as much design attention as the main flow.
-
Keep a person in charge of the decision
Approvers used to open 14-page agreements with no context and read every line looking for problems. The pre-review panel checks the document against Workday before they open it, and flags things like pay band violations, legal entity mismatches, and expired IDs. The approver still makes the call. The AI just saves them from digging through the whole thing to find what’s wrong. You should always be able to answer “who approved this, and when?”
-
Bring in Legal and measurement from day one
A lot of the constraints I designed around were things I already knew from the platform work. Still, if Legal and Compliance had been in the AI workshops from the beginning, we would have caught some edge cases sooner. Same with measurement: if you set up server-side task timing and a feedback loop early, you get useful data while you’re still designing, not only after launch.
Common pitfalls
- Starting with the model: if the first question is “what can the AI do?”, you’ll end up with demos. Start with what the person needs.
- Hiding uncertainty: a wrong answer that looks confident does more damage than one that admits it isn’t sure.
- Letting exceptions hold up the batch: set aside what can’t be resolved so the rest can keep moving.
- Bringing Legal in at review time: by then, their feedback means rework instead of shaping the design.
- Forgetting the people with the least context: new hires on their first day, warehouse workers on their phones, and non-native English speakers need to understand what they’re signing, not just get through it.
Checklist
Before an AI feature in a regulated workflow goes to review:
- Every AI output traces back to a research finding
- Every suggestion shows its reasoning and confidence
- Users can accept, reject, or edit inline
- The low-confidence and “can’t resolve” states are designed, not left to defaults
- A named person approves every consequential action
- Legal and Compliance have reviewed the language
- Measurement is in place before launch
Where this came from
This came out of phase 2 of Amazon’s Employee Document Management platform, which handles 4.8 billion documents a year in 67 countries. On the employee side, the AI work built on the accessibility work we’d already done. It added plain-language summaries of legal agreements and a section tracker that flags when someone moves quickly past an important clause, so they catch it before they sign instead of after.