There is a popular belief that building for safety is at odds with growth. After starting and leading the safety and risk tech team at Lyft, I can confidently say that this is not true.
Growth is about furthering product and business value. Failing to plan for safety measures will erode that product value and certainly cause significant business costs. And in some cases, it’s simply unethical.
Safety issues can impact your business in multiple ways:
Churn: Customers who experience a safety issue typically have such a bad experience that they never come back. Many people who contract food poisoning from dining out will avoid going to the same restaurant again.
Brand Damage: The safety issue doesn’t just affect that customer, it also affects how their social network perceives your product. In the event of negative news coverage or a viral story, hundreds of thousands of people or even millions can have a negative perception of your brand, resulting in less advocacy and a lower willingness to adopt the product. This is part of what Peloton experienced with the fatal child injuries the Tread+ caused.
New Regulation: Regulators generally seek to protect consumers. If there’s a safety issue, manifested or perceived, regulators often step in to set rules. Sometimes these rules are difficult and can affect product market fit. The rules can require new checks, change cost structures with insurance requirements, or even just forbid certain approaches. For a long time, autonomous vehicles were required to have a steering wheel in the vehicle as part of the design — even though no human was supposed to steer.
Cost: Cost can manifest in many ways, including legal settlements, insurance claims, fines, or even new capabilities. As clarified in a BBC report, Roblox has every single piece of content checked by a machine or human. That’s an expensive process to stand up.
Beyond the negative effects on growth, there are benefits to making someone feel safe and comfortable using your product — especially if it’s something as atypical as staying at an unknown person’s house like on Airbnb or entering a rideshare vehicle.
Planning for safety is planning to support growth. Seasoned product managers need to know about this, especially as technology permeates into more and more sensitive use cases.
While this post is written to benefit any business, product managers should be proactively thinking about building for safety if any of the following apply to your product or service:
You are in a regulated industry (example: airlines)
You have a growing community that's interacting in nascent ways that can be harmful to one another (example: social media, games, or cities)
Product failure causes significant harm (example: housing)
Ready to dig in? In the rest of this post, we’ll review…
Preet is obsessed with enhancing safety and societal infrastructure. He started and scaled Lyft’s safety tech area, started a company focused on improving 911 calls (acquired by RapidSOS), and is a co-owner of the daily check-in service for seniors living alone, Snug Safety. He also still wears Ninja Turtle T-Shirts. You can find more of his writing at High Absolute Value.
Perception Matters, and Product Safety is Interdisciplinary
Success in product safety involves both perception and reality.
After all, safety is a feeling.
Behavioral economists have coined a concept called loss aversion, which explains that the psychological pain of losing something is much more powerful than that of gaining something.
Loss aversion leads people to make decisions based on risk/reward ratio calculations. If the ratio isn’t good or overwhelmingly on the side of reward, they won’t convert or adopt a new product.
This ratio is based on perceived risk, not just experienced risk.
One way to improve the risk/reward ratio is to improve reward (the traditional product focus). Another way is to reduce risk — we’re going to focus here!
Note that safety is different from the traditional growth way of sizing the cost/benefit of a user action.
In product growth, converting a user is usually associated with a very clear value. But in safety, the cost of a user experiencing a safety issue isn’t equal to only this user's experience being affected.
There are two hidden factors:
Safety issues are frequently underreported. If one user reported an issue, there are likely more users that have had the same experience and didn’t inform the company.
The initial user will share their safety experience with others, who may adjust their own behavior to avoid that risk altogether
This means you need to try and estimate the impact of the user who experienced the negative event, and also need a modifier for what is underreported and others whose perception will be affected.
Chipotle is an example of this. In 2015, Chipotle had a multi-state outbreak of E.Coli where 400 people got sick, spurring lots of press coverage. Millions of people stopped eating at Chipotle because of the fear of getting sick. Even two years later in 2017, Chipotle’s revenue was still down 22.4%. Ultimately the founder stepped down as CEO to usher in a new chapter. And more recently, the company agreed to pay a $25 million federal fine for the outbreaks.
There’s an important lesson here: 400 people got actually sick, but ~5000X that amount stopped eating at Chipotle just to avoid the perceived chance of getting sick.
Perception in Action: Health Safety at Lyft
Let’s look at an example from my time leading product for Lyft’s COVID-19 response, and use it throughout the rest of this post. When my team started designing the Health Safety Program at Lyft, COVID had just arrived in the U.S., but we still knew very little about the virus that caused it. People were sent home out of an abundance of caution, and mask mandates had yet to be ordered by the CDC.
Still, Lyft’s revenue and ridership had fallen over 70% in three weeks. Both our riders and the Lyft team were still learning about data on the efficacy of masks and how COVID is transmitted, but my team made a decision to reinforce our commitment to safety in the product.
Our customers were worried. Was it possible for them to catch COVID in the car? Would other people wear masks? If they caught COVID, would they infect someone else in their family? Could they die? Many had stopped using our product to avoid the potential risk of COVID.
While many customers had the privilege of remote work, Lyft also served many essential workers who needed to get to work to help others. For the drivers who wanted to keep working, there were worries about how to best protect themselves. How do they guide riders about what to do? Can they get PPE reliably? Were there any changes they could make to their car to enhance their safety?
And it wasn’t just customers. Lyft and Uber stopped shared rides across the country. There was a fear that rideshare altogether might be suspended. Government officials were worried about how to approach it since rideshare served an essential transportation purpose.
There was a lot of anxiety, to say the least, all driven by the perceived danger of catching COVID at a time when little was known. We knew our plans would have to address that.
How to Identify and Categorize Product Safety Risks
There are different types of risks — possible, perceived, and experienced — and you need to catalog the ones that can affect your customers and business. What may affect one type of product might not affect another. For example, misinformation is a concern for social media products but not really for a new type of housing.
Step 1: Make a list of possible risks with an interdisciplinary group.
If you already have data and known areas of improvement, start there. If not, use your interdisciplinary group to generate a set of hypotheses. This group should represent your customer’s experience but also the perspectives of your business’s various stakeholders. This may be uncomfortable, but get as specific as possible. Place the risks and concerns along the user’s journey.
For example at Lyft, we brought together leads from policy, design, legal, marketing, customer support, analytics, and engineering.
Step 2: Start to categorize these risks.
You should bring multiple lenses to this step of the process — at least as many lenses as there are different types of risk. Think about your product and community in particular.
Here are some lenses that can offer a unique point of view on product safety:
Product Risk vs Behavioral Risk
The first lens is whether a risk is a product risk or a behavioral risk.
Behavioral risk comes from users of a product using the product incorrectly or maliciously. The risk is created by people's behavior. You mitigate this by clearly and visibly messaging proper behavior and having mechanisms to detect incorrect behavior. An example of a behavioral risk is Twitter Doxing.
Product risk comes from products being incorrectly designed and failing during reasonable use. You mitigate this with a safety-informed product design process and very rigorous testing processes. An example of product risk is the Chevy Bolt’s batteries catching on fire. There was no user error there.
Product risk and behavioral risk can also intersect. For example, Tesla’s Autopilot is a product that helps lower the cognitive load of driving. It safely takes over many, but not all, driving tasks.
However, there were fatal collisions that occurred seemingly due to Autopilot failure. In some of those cases, it became clear that the driver was no longer engaged with the driving experience at all, including being asleep. So when autopilot disengaged, per design, the driver wasn’t using the product as it was supposed to be and wasn’t able to take over driving.
Abuse and Bad Actors
Abuse and bad actors are a more extreme form of behavioral risk. Whereas some users can inadvertently cause behavioral risk from a lack of knowledge, abusive users are ones who are deliberately subverting the product’s design. Examples can be spam, fraud, or harassment.
Identity and Data Privacy
This might not be what people normally think about with safety, but it’s so pervasive and fundamental it needs a specific callout. Customers using your product often are providing very sensitive data.
For example, in Airbnb’s case, it is structured when someone is away from their home and in a different place. That information can be problematic in the wrong hands.
Further, if the person is not who they claim they to be, because of an account takeover, that’s risky to other parties and a painful experience for the user. Identity verification and protection is something to be planned for with any product of sufficient scale and intimacy.
Support and Claims
It’s bad enough that a safety incident has happened. It’s worse when the follow-up support is slow, misinformed, incorrect, or unavailable. The risk here is making a bad situation worse, losing even more user trust, and re-traumatizing someone.
Another risk are fraudulent claims. This can result in big, incorrect payouts for the company and also incorrect corrective action with any other party. When a 911 call is received about a reported fire, dispatchers are trained to always send firefighters to respond but to make sure those firefighters do their own independent assessment
Frequency vs. severity
Some concerns manifest in daily anxiety, which is very frequent. Some very severe things, such as car batteries spontaneously catching on fire or sexual assault, are extremely infrequent and may only happen once in the history of your product. Of your safety concerns, it’s most likely that the worst of the worst aren’t happening the most often — but you still need to plan for them.
Controls and Preferences
The same way that customers may have different needs from a product, they may have different levels of anxiety and worry. Accordingly, they may need additional controls and measures to help them feel safe while using the product. An example of this is parental controls and what type of content is discoverable. Another is supplementary insurance.
Personas
Different communities experience different risks at different frequencies and worry about them differently. For example, unfortunately, women are generally victims of harassment much more frequently and as such is a much more top-of-mind concern. Similarly, seniors are more likely to get phished while certain ethnic groups face more verbal abuse.
Additionally, there are some personas for whom you need to change how your product works in order to be safe. For example, there is some content inappropriate for younger users to see. That wouldn’t be considered a safety concern for adults, but it is for children.
If you’re unsure about what risks are top of mind for which community, dig in. Do user interviews, take stock of other societal trends that may be affecting a specific group, and categorize your data with that persona in mind.
Impact of Risk
Each of these risks can have a tangible business impact as we have listed above (churn, cost, brand, etc.). As you build out your list of risks, make sure the impact is noted as well.
This process of reviewing risks and concerns should be recurring the same way that you would refresh your growth roadmap. New safety risks arise that you need to categorize and make a plan to mitigate.
Categorization in Action: Health Safety at Lyft
Rideshare is about spontaneously matching a rider with a driver — who rarely know each other beforehand — to get the rider to their destination as quickly as possible.
Using the lenses above, the risks the product needed to address were:
Behavioral: How do we make sure people keep wearing masks and know what’s expected as part of the new policies to keep everyone safe? How do we maximize airflow and limit how close people are to each other?
Abuse: How do we address users that repeatedly don’t wear masks and create additional risk for others in the community?
Support: What is the response when a driver experiences a rider (or vice versa) not wearing a mask? How should they report this? How do you help a driver that catches COVID?
Frequency: Since people are riding with someone unknown nearly every ride, this was an incredibly frequent risk we needed to address.
Why It’s Important to Gather Data to Help Prioritize Product Safety Measures
Now that you’re aware of the different types of risk that can affect your customers and the importance of perception, you’re probably getting ready to start roadmapping and getting specific. The good news is that while safety has some differences, a lot of the fundamental concepts from growth work map over nicely — especially the need for data.
You need good data to prioritize.
Data is critical in both growth and in safety. You can’t fix what you don’t know. Build out analytics for both leading and lagging indicators of your set of risks. This will start to give you data to inform your funnel and what you build first in your roadmap.
If you’re unsure where to start, start with how users report a concern. The product’s report function should have specific, structured concerns in addition to free-form text for any additional context. This will give you specific issues to size and understand what happened end-to-end. Look at the reporting experiences on products like Lyft and Airbnb for inspiration of what to emulate.
Another area of great, clear data is quality assurance (QA) testing. Especially for product risks, QA testing is a critical way to find potential safety issues before they reach your customers, especially as it relates to product risks. You can help your testers here by doing a safety focused pre-mortem before an upcoming product is launched. An example of safety and security focused QA is vulnerability testing where your own teams try to find the conditions in which your product can fail.
The funnel still applies, but the shape should be different
A funnel, and its extension the growth loop, are tried and true frameworks of growth. They map over to safety, but in the inverse. For growth, you are trying to maximize the percentage of people who get to the end of the funnel.
In safety, it’s the opposite. You want 100% of your audience to be aware of what you’re doing to enhance safety and the rules of the community, whether they are customers or just brand observers. However you want as few users as possible to have something bad happen to them, and even fewer to experience or cause repeated abuse.
Data in Action: Health Safety at Lyft
When we were first launching the Lyft Health Safety Program, we didn’t yet have structured data to inform our approach. Instead, we used surveys and customer interviews to inform needs.
We knew our first requirement was to set the policies. But as we built both the messaging experiences for those policies and ways to report violations of the policies, we ensured that we built out instrumentation to inform dashboards and reporting. This data was critical to inform subsequent improvements.
Here’s what that funnel looked like after we went live:
A specific note with regards to reporting: We built out multiple ways to report violations of the policies. Underreporting is a systemic pattern with regards to safety issues all throughout society. We knew there were always reports we weren't going to get. So, we built multiple ways to report, did surveys, and even developed other checks to get as close to the true violation rate as possible.
How to Start Building Your Safety Roadmap
When it comes to building a safety roadmap, the ultimate goal is to increase comfort and reduce risk. Already, we’ve specified the various risks that can affect your product and hopefully have some data on the relative importance of each risk.
As investor Charlie Munger says, “Invert, always invert.” Look at those specific risks and invert them into potential solutions that could mitigate those risks and bring comfort. It’s best if you can tie these all together into a top-line narrative and experience that you’re aiming toward.
Ultimately, you should prioritize executing on the most impactful ideas toward that narrative, and beware you may need multiple solutions. That’s where the concept of safety in layers comes up. We did this on Lyft’s Health Safety Program to ensure people followed our policies.
Building Lyft’s Health Safety Program
Lyft’s COVID safety response plan for minimizing risk in the product involved three key pieces:
Setting policies for expected behavior, and reinforcing those policies
Providing drivers access to PPE (as there was a big shortage in the early days of the pandemic)
Adapting our support processes to run smoothly and effectively
We wanted our community to feel safe and be equipped to act safely.
Policies
Our policies focused on three behaviors, each informed by the established public health protocols that were applicable to rideshare:
Maximizing the distance between people in the car
Ensuring ventilation in the vehicle
Wearing a mask
We relied on public health protocols for two reasons: First, we were reinforcing common messages users were hearing outside of the product, leading to an easier behavior shift for our community. Second, there simply wasn’t (and still isn’t) a way to detect the real-time transmission of COVID — we needed to focus our policies on the upstream prevention measures that were known to reduce transmission likelihood.
Introducing our policies started with our Health Safety Commitment. It was a community-wide modal that detailed our policies and had the community agree. This was critical because these policies were still pretty new to everyone, so we wanted to ensure broad awareness. Further, since customers had deliberately agreed, we could draw upon this to reinforce the policies.
The need to follow these policies was critical every ride. Accordingly, we reinforced these policies with contextual reminders and even targeted feedback if someone was reported to not be following the policies.
Repetition
One of our goals in Lyft’s Health Safety Program was to maximize the percentage of people wearing masks in our vehicles. We didn’t just have one solution for this. We used multiple, overlapping mechanisms to maximize efficacy here.
Here’s what our safety looked like, in layers:
At the top of the funnel, we informed all of our customers and had them commit one-time to our health safety protocols.
Before requesting a ride, we gave contextual reminders of the requirement to wear a mask. This was initially shown on every ride.
When a vehicle was approaching, we gave riders a reminder to put on their mask before entering.
If someone wasn’t wearing a mask, either a rider or driver could cancel the ride without getting in and specify that reason. It gave them some control over their experience, which reduced anxiety.
The canceled ride, and other report types, would be used to give feedback to the party not wearing a mask (unfortunately this could be abused). Some would even be challenged to provide a selfie wearing a mask to prove the presence of a mask.
If someone still persisted in not adhering to the policies, they were suspended from the platform.
These multiple features all worked together towards the goal of ensuring mask wearing. And the product reinforced Lyft’s safety measures repeatedly to foster a new habit of the desired behavior.
Altogether, our team’s health safety program had an immediate impact when it launched. Lyft had its highest week of rider reactivations with increased trust, which helped to accelerate the business’ pandemic rebound.
“We are encouraged by the recovery trends we are beginning to see, with monthly rideshare rides in July up 78% compared to April,” said Logan Green co-founder and chief executive officer of Lyft, in a press release accompanying the company’s Q2 2020 results.
And, best of all, the team did it by helping to protect the community.