Twelve navigators and a spreadsheet
Maya had a problem worth solving. She did not yet have a product, and she knew better than to let the team start building one.
Instead, she and Lena Park, Brindlepath’s head of design, recruited twelve outpatient navigators from three health systems: two existing customers and the hospital that had said no. They did not show them a prototype. They asked each navigator to walk them through the last three patients who had “fallen off the radar,” what had happened, when they had found out, and what they had done.
The stories were remarkably consistent. The navigator usually found out late, from the patient, a family member or an emergency department visit. The information she needed existed somewhere: in the discharge summary, the inpatient care manager’s notes, the specialist’s referral. It just never arrived in one place, at the right time, with a name attached.
Then Dev Raman, the engineering lead, did something that made the engineers on his team uncomfortable. Rather than build anything, he set up a shared spreadsheet for one navigator team at a customer site, and for two weeks a Brindlepath analyst filled it in by hand every morning: newly discharged patients, who the inpatient contact was, what the discharge summary said, and a suggested owner. The navigators used it every day. By the second week, they were asking for it before 8 a.m. One navigator caught two patients who would otherwise have been missed.
Nobody at the customer knew there was no software behind it. Two weeks and one analyst’s time had answered the most expensive question Brindlepath faced: if we deliver this, will they use it? Now the team could build with confidence.
Discovery is continuous, not a phase
In older ways of working, discovery happened once, at the start of a project, and produced a requirements document. Then the team built for months and found out whether they had been right.
Modern product teams treat discovery as a continuous practice that runs alongside delivery. Every week, the team learns something about its customers and tests at least one assumption. Teresa Torres calls this continuous discovery and offers a simple benchmark: a product team should be talking with customers at least weekly.1
Continuous discovery matters for three reasons:
- Markets move. What was true at the start of a project may not be true at the end.
- Small bets beat big ones. Testing an assumption in a week is cheaper than building on it for a quarter.
- Teams that hear from customers make better small decisions. Hundreds of tiny choices about wording, defaults and edge cases go better when the people making them have heard real users recently.
The rhythm does not need to be elaborate. A product trio of a product manager, a designer and an engineering lead, holding one or two customer conversations a week and reviewing what they learned, will outperform most formal research programs.
Interviewing well
Customer interviews are the most important discovery tool and the most commonly done badly. The most common mistake is asking people to predict their own future behavior or to evaluate your idea. People are polite and optimistic. They will tell you your idea is great and that they would definitely use it. Then they won’t.
Rob Fitzpatrick’s rule of thumb is to talk about their life, not your idea; ask about specifics in the past, not generics or opinions about the future; and talk less, listen more.2 In practice:
Ask about the last time, not “usually.” “Tell me about the last patient you lost track of after discharge” produces a real story with real details. “How do you usually handle discharges?” produces an idealized process that matches the policy manual.
Follow the workaround. When someone mentions a spreadsheet, a sticky note or a daily phone call, ask to see it. Workarounds are evidence of a problem worth solving and a blueprint for the solution.
Ask what it cost. Time, money, stress, consequences. If they cannot name a cost, the problem may not be urgent.
Ask what they have already tried. Someone who has never looked for a solution probably does not feel the pain strongly enough to buy one.
Avoid pitching. The moment you describe your product, the interview becomes a sales call and the data becomes polite.
Interview the people who said no. Lost deals and churned customers teach more than happy ones. Maya’s most useful interviews were at the hospital that had rejected Brindlepath.
After each interview, capture the evidence while it is fresh: quotes, stories, workarounds and surprises, separately from your interpretation. Your interpretation will change; the evidence should not.
Map your assumptions
Every product idea rests on a stack of assumptions. Most teams never write them down, which means they never test them. A simple exercise changes that.
List every assumption that must be true for the idea to succeed, grouped by type:
- Desirability: Navigators will check a daily list of newly assigned patients.
- Usability: They can confirm or reassign ownership in under a minute per patient.
- Feasibility: We can get discharge data from the customer’s health record within a few hours of discharge.
- Viability: Health systems will pay per navigator seat, at a price that supports our margins.
- Trust and compliance: Customers’ security teams will approve the data flow without a six-month review.
Then rate each assumption on two scales: how important it is (if it is wrong, does the idea fail?) and how much evidence you already have. The assumptions that are important and unproven are the ones to test first. This sounds obvious, yet many teams spend their first months polishing the user interface, the part they are most confident about, while the riskiest assumption, often a feasibility or viability question, waits untested until late.
For Brindlepath, the riskiest assumption was not whether navigators wanted a list. The field work had largely answered that. It was whether discharge data could flow from five different health record configurations quickly and reliably. So the team’s first technical spike was an integration test with one customer’s health record, not a user interface.
The evidence ladder
Not all evidence is equal. An enthusiastic comment in an interview is evidence; a customer changing their behavior is stronger evidence; a customer paying money is stronger still. Figure 3.1 arranges common validation methods by the strength of evidence they produce and the cost of running them.
The principle: match the strength of evidence to the size of the decision. A small feature change may need nothing more than a few interviews and a prototype test. A new product line with a dedicated team deserves evidence from the top of the ladder: commitments, pilots with success criteria, or paid pre-orders.
The experiments in detail
Interviews and observation. The base of the ladder. Cheap, fast and essential for understanding the problem; weak as evidence that people will adopt a specific solution.
Surveys. Useful for sizing a problem you already understand, measuring how widespread a behavior is or prioritizing among known options. Poor for discovering problems, and poor at predicting behavior.
Prototype tests. Put a clickable prototype in front of users and give them a task. You learn whether they understand the design and can use it. You do not learn whether they will use it every day.
Smoke tests. Describe a product or feature as if it exists, on a landing page or in an in-app prompt, and measure how many people try to use it or sign up. Common in consumer products; in business software, a “request early access” button for existing customers works well.
Concierge tests. Deliver the outcome by hand, openly, to a small number of customers. You learn exactly what the service needs to do, and whether customers value it enough to keep using it.
Wizard of Oz tests. Deliver the outcome by hand behind something that looks like a product, as Brindlepath did with its spreadsheet. You learn whether customers will change their behavior when the experience feels real.
Letters of intent and pilots with success criteria. A customer formally commits time, data or money to a trial with agreed measures of success and a decision date.
Paid pilots and pre-orders. The strongest evidence short of a full sale: someone gives you money for something that is not finished.
Minimum viable product, done right
The term minimum viable product is one of the most misused in the industry. It does not mean “the first release, with fewer features.” It means the smallest thing you can build that delivers real value to real users and lets you learn whether your core assumption is true.
Two words carry the weight. Minimum means you cut everything that is not needed to test the core value. Viable means what remains must genuinely work for the user; a broken or embarrassing product teaches you nothing, because users reject it for the wrong reasons.
There is good evidence that most teams build far more than they need. In a 2019 analysis of usage across 615 of its software customers, the product-analytics company Pendo found that 80 percent of features in the average product were rarely or never used, and that about 12 percent of features generated 80 percent of daily usage.3 The data comes from one vendor’s customers and is several years old, but it matches what most product leaders see in their own analytics.
Every feature that does not earn its place costs money to build, test, document, support, secure and maintain, for as long as the product exists. The cheapest feature is the one you were disciplined enough not to build.
Brindlepath’s MVP
After the spreadsheet test, Maya’s team defined the minimum viable product in one sentence: Every morning, each outpatient navigator sees the patients discharged in the past 24 hours who are now theirs, with the discharge summary and the inpatient contact attached, and can accept or reassign each one.
What they deliberately left out of the first release:
- Scheduling, the original product’s core feature. Navigators could keep using the existing tool.
- Analytics dashboards for managers.
- Patient-facing messaging.
- Support for every health record vendor. They started with the two configurations used by their pilot customers.
What they insisted on including, even in the MVP:
- Role-based access controls and full audit logging, because patient data was involved and the customer’s security team would not approve anything less.
- A way to reassign a patient in one step, because the field work showed ownership disputes were common.
- A measure of the outcome: the share of high-risk discharged patients with a confirmed owner within 24 hours.
The list of exclusions is as much a part of the MVP as the list of features. Write it down, share it, and defend it.
Pilots that end in a decision
In business-to-business software, and especially in healthcare, validation often takes the form of a pilot. Pilots are valuable. They are also where many promising products go to die, in what people in the industry sometimes call “pilot purgatory”: a trial that is never declared a success or a failure and simply fades.
A pilot that ends in a decision has five things agreed in writing before it starts:
- The problem and the measure. What outcome are we trying to change, and how will we measure it? For Brindlepath: the share of high-risk discharges with a confirmed owner within 24 hours, and missed first follow-ups.
- The baseline. What is the measure today? If nobody knows, measure it before the pilot begins.
- The success threshold. What result would justify a purchase? Agree on a number, not on “a good result.”
- The scope and duration. Which teams, how many users, how long. Long enough to see the outcome; short enough to keep momentum. For most workflow products, 60 to 90 days is enough.
- The decision and the decision-maker. Who decides, on what date, and what happens next if the threshold is met: a signed contract, a wider rollout, a budget line.
Get the economic buyer, not just the enthusiastic user, to agree to these terms. A pilot sponsored only by the people who love the product will produce a happy report and no purchase order. Chapter 13 returns to pilots as part of go-to-market.
Brindlepath in practice: the pilot agreement
The spreadsheet test answered the desirability question. The next step was a pilot that could turn into a contract. Maya, Lena and the sales lead wrote a one-page pilot agreement with the second health system, using the template in the Appendix.
The measure was the share of high-risk discharges with a confirmed owner within 24 hours, along with missed first follow-up visits. The baseline took two weeks to establish, because nobody at the health system had measured it before: about 45 percent of high-risk patients had a clear owner within a day. The success threshold was 75 percent, sustained for the last four weeks of a 90-day pilot. The scope was one hospital and its outpatient navigator team of fourteen people. The health system paid a modest pilot fee and committed a navigator manager’s time each week. The decision-maker was the chief nursing officer, with the CFO’s agreement, and the decision date was written into the agreement, along with the price and timeline for a rollout to the system’s other four hospitals if the threshold was met.
Three things went wrong during the pilot, which is normal. The discharge feed missed patients from one unit for a week. Two navigators were on leave in the second month. And a change to the hospital’s risk-scoring rules briefly flooded the queue with low-risk patients. Because the pilot had a clear measure and a weekly review, each problem was caught and fixed within days.
By day 90, the measure stood at 82 percent. The decision meeting took twenty minutes, because nothing had to be negotiated: the terms had been agreed at the start. The rollout to four more hospitals began the following month.
Common mistakes
Running discovery as a phase. Teams that do a burst of research at the start of a project and then stop are surprised by what they learn at launch. Discovery is a weekly habit.
Pitching instead of listening. The moment you describe your idea in an interview, the data becomes polite. Save demos for later, and keep early conversations about the customer’s world.
Testing the safe assumptions. It is tempting to test what you are already confident about, such as the interface, while the riskiest assumption, often feasibility or willingness to pay, waits. Test the assumption that would kill the idea first.
Building an MVP that is all minimum and no viable. A product so thin that it does not solve the problem teaches you nothing, because users reject it for the wrong reasons.
Treating the MVP as the finished product. The opposite error: launching the first slice and moving on. An MVP is the beginning of learning, not the end of the project.
Running pilots without a decision. A pilot with no agreed measure, threshold or decision-maker drifts into “pilot purgatory,” consuming time and goodwill on both sides.
Try this
List the five assumptions your current idea depends on most, one each for desirability, usability, feasibility, viability and trust. Score each from one to five for how important it is and how much evidence you have. Pick the assumption with the highest importance and the least evidence, and design the cheapest experiment that could disprove it within two weeks. Use the evidence ladder in Figure 3.1 to choose the method. Then schedule the experiment, and decide in advance what result would change your plan.
How AI changes this
Discovery is one of the areas where AI helps most, and also where it can do the most damage if used carelessly.
Synthesis is now fast. Interview transcripts, support tickets and sales-call recordings can be transcribed, tagged and clustered by an AI assistant within hours of collection. A product trio that used to spend a week synthesizing a research round can now spend that week running the next one.
Prototypes are now cheap. AI coding and design tools can turn a sketch into a working prototype in an afternoon. That moves prototype testing down the cost axis of the evidence ladder and makes it reasonable to test several directions instead of one.
Humans talk; agents synthesize. The order matters. Use people for the conversations and the sense-making, and AI for the transcription, tagging, first-pass clustering and drafting. Figure 3.3 shows the division of labor that works in the teams I advise.
Synthetic users are not users. Some tools now offer AI-generated “personas” that answer interview questions. They can help you rehearse an interview guide or anticipate obvious objections. They cannot tell you anything true about your actual customers’ behavior, because they have never had your customers’ problem. Never count a synthetic interview as evidence.
Beware the smooth summary. AI is good at producing coherent themes from messy data, which is exactly the risk. Real customer evidence is contradictory and specific. Keep the raw quotes attached to every theme, and read a sample of the source material yourself. The navigator who caught two patients in the spreadsheet test is worth more than any summary of her.
Keep the patient’s privacy in mind. In healthcare and other regulated settings, interview recordings and support tickets may contain personal or health information. Before you put them into any AI tool, confirm that the tool is approved for that data, and remove identifying details where you can. Chapter 11 covers this in depth.