High-Duty-of-Care Product Design
Design work-critical products around repeated exposure and real-world consequences.
- Difficulty
- Advanced
- Time to result
- ~ongoing to results
- Steps
- 7
- Confidence
- 96%
High-duty-of-care product design starts by distinguishing casual consumers from people whose livelihood, safety, or essential services depend on the product. A worker may use an application for many hours each day, so a statistically rare failure can become a routine lived experience. Teams therefore examine tail outcomes such as the 95th-percentile experience rather than relying on averages, which can conceal harm. They also classify decisions by consequence. Reversible interface changes may suit ordinary A/B testing, while choices analogous to compensation or benefits require slower judgment, direct listening, and stronger governance. The framework does not reject analytics; it places experimentation inside a broader duty of care. Success means improving aggregate marketplace performance without treating vulnerable or highly exposed participants as interchangeable optimization units.
Origin
Khosrowshahi described Uber's shift from applying standard business-to-consumer optimization methods to recognizing that drivers and couriers spend hours in the app and depend on it for earnings.
Core principles
- 01Products used to earn a living require a higher duty of care.
- 02Frequent users repeatedly encounter rare failure states.
- 03Averages conceal harmful tail experiences.
- 04Not every employment-like decision belongs in an A/B test.
- 05Listening and humility must complement quantitative optimization.
How to run it
- 1
Identify Dependency
Determine which users rely on the product for income, safety, health, mobility, or another essential outcome.
Pro tip Map both direct users and people affected indirectly by product decisions.
Watch out Do not infer low stakes merely because users are legally classified as customers or contractors.
- 2
Calculate Cumulative Exposure
Measure how often and how long each user group encounters the product and its decision systems.
Pro tip Translate event probabilities into expected weekly or monthly exposure per user.
Watch out A rare event at transaction level may be common for a high-frequency user.
- 3
Inspect the Tails
Examine poor outcomes and edge cases at relevant percentiles rather than relying only on averages.
Pro tip Pair percentile metrics with concrete user journeys.
Watch out Averages can remain healthy while a meaningful population experiences repeated harm.
- 4
Classify Experiment Risk
Separate reversible convenience experiments from decisions that materially affect earnings, access, benefits, or safety.
Pro tip Create a review threshold based on consequence and reversibility.
Watch out Statistical significance does not make an ethically inappropriate experiment acceptable.
- 5
Listen Before Optimizing
Collect direct qualitative feedback and incorporate participant context before changing consequential systems.
Pro tip Include users who report poor outcomes, not only highly engaged advocates.
Watch out Feedback from internal employees may not represent people who depend on the product daily.
- 6
Ship With Guardrails
Roll out consequential changes slowly, monitor tail outcomes, and maintain a clear rollback path.
Pro tip Use cohort and geographic limits to contain unforeseen harm.
Watch out Marketplace-wide averages may delay detection of localized failures.
- 7
Audit Distributional Outcomes
Check whether the system works for different user groups and address recurring pockets of dissatisfaction.
Pro tip Track both the percentage and absolute number of affected people.
Watch out A small percentage can still represent hundreds of thousands of users at scale.
In the wild
Uber recognized that optimizing its driver application like an ordinary consumer app was inadequate. Drivers may spend four to six hours in it daily and experience statistically uncommon failures several times each week.
→ The company began slowing consequential changes, listening more closely, and treating earner products with greater care.
Khosrowshahi compared some earner-side experimentation to testing whether employees should receive a three percent or six percent retirement match based on turnover. The analogy exposed a class of decisions that should not be delegated solely to conversion experiments.
→ The team gained a clearer boundary between routine optimization and livelihood-sensitive product governance.
Common mistakes
Trusting the Average
Aggregate results can hide repeated failures experienced by the most active or vulnerable users.
Testing Every Decision
Some choices affect livelihoods deeply enough that an A/B test is not the appropriate decision mechanism.
Designing From Employee Experience
Employees who briefly use the consumer product may underestimate what daily earners experience.
Is it for you?
Best for
It is best for labor platforms, financial tools, healthcare systems, and other products with frequent use and material consequences.
Not ideal for
It is not ideal as a blanket reason to avoid low-risk, reversible experimentation on ordinary convenience features.
From the transcript
“when you're building a product that people are making a living off of or are earning money that they have to earn with there's a…”
“you build you don't look at p50 because averages lie and then you look at P95”
“there's been a pretty important culture change of the company which is like higher duty of care actually slowing down in terms of how we…”
From the episode
Uber CEO Dara Khosrowshahi
Dara Khosrowshahi