When to use this

After shipping Amazon’s Employee Document Management platform, I led a second phase focused on AI integration across four product surfaces: an M&A dashboard, bulk document generation, the HR admin portal, and the employee portal. The platform worked, but it was reactive. It showed people what had happened without telling them what to do about it.

Use this method when the stakes of a wrong answer are high: HR, legal, finance, healthcare, or anywhere an output affects a real person’s employment, money, or rights. The signals that AI can help are familiar: people interpreting raw data themselves, repeating round trips to fix errors, or reading an entire document to find the one thing that’s wrong.

It’s not the right method when a mistake costs little and speed matters more than accountability. There, a lighter-touch approach is fine.

The steps

  1. Start from a validated finding, not a model capability

    I ran cross-functional workshops to identify where intelligence could close a gap for each user type. Every concept had to trace back to a validated research finding or an existing backlog item. The framing question: what does this person need right now that the system isn’t surfacing for them? Done looks like a short list of concepts, each with the finding it answers.

  2. Turn status into a decision

    Managers tracking acquisition paperwork had one signal: a completion percentage. “57% complete” leaves all the interpretation to them. The redesign added risk scoring with a plain-language explanation, close-date forecasting based on velocity, and anomaly detection that named the actual blocker (three unreleased Critical employees) instead of burying it in a table. Done looks like every metric answering “so what do I do?”

  3. Show the reasoning, the confidence, and the choice

    Bulk document generation used to be a full round trip: upload a spreadsheet, get errors, download, fix offline, re-upload. The self-fix feature detects each error, proposes the most likely correction with its reasoning and a confidence level, and lets HR accept or reject inline. About 75% of common errors became fixable without leaving the page. Done looks like no AI output without an explanation next to it.

  4. Make the “I can’t” state honest

    Some rows can’t be auto-resolved. Those are flagged explicitly, with a clear reason, and they don’t block the rest. In a batch of 72 employees with 8 errors, the other 64 still generate. In compliance, an AI that guesses and gets it wrong is a liability, so the fallback has to be designed as carefully as the happy path.

  5. Keep a person as the decision-maker

    Approvers were opening 14-page agreements cold and reading every line to find problems. The pre-review panel cross-references Workday before they open the document and flags pay band violations, legal entity mismatches, and expired IDs. The approver still makes the call. What AI removes is the archaeology. Done looks like a clear answer to “who approved this, and when?”

  6. Bring in Legal and measurement from day one

    Many of the constraints I designed around were things I already knew from the platform work. Getting Legal and Compliance into the AI workshops from the start would have surfaced edge cases sooner. The same goes for measurement: server-side task timing and a continuous feedback loop give you signal during design, not just after launch.

Common pitfalls

  • Leading with the model: “What can the AI do?” produces demos. “What does this person need?” produces products.
  • Hiding uncertainty: a confident-looking wrong answer does more damage than an honest “I’m not sure.”
  • Letting exceptions block the batch: isolate what can’t be resolved so everything else moves forward.
  • Adding Legal at review time: by then, their constraints mean rework instead of design input.
  • Forgetting the most vulnerable users: new hires on their first day, warehouse workers on phones, and non-native speakers need comprehension, not just completion.

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 defaulted
  • A named person approves every consequential action
  • Legal and Compliance have reviewed the language
  • Measurement is in place before launch

Where this came from

Phase 2 of Amazon’s Employee Document Management platform: 4.8 billion documents a year across 67 countries. On the employee side, the AI layer built on earlier accessibility work, adding plain-language summaries of legal agreements and a section tracker that flags when someone moves quickly past a high-stakes clause, before they sign rather than after.