[Part 1, Part 2, Part 3, Part 4]

Harvest’s situation was different from Roomba’s.

We on the Roomba team thought the fact that we all had floors and understood how to clean them gave us a huge advantage.   It meant that without conducting customer studies or surveys or interviews, we intuitively knew the robot’s requirements.   Thus, we figured, if we built a robot that satisfied us it would necessarily satisfy our intended customers.  (We were wrong about that as you’ll see in Dancing with Roomba.  But, with guidance, we overcame our hubris.)

At Harvest Automation we had no such edge.  None of us were veterans of the nursery and greenhouse industry.  None of us were farmers.  And none had managed farm workers.  We well appreciated our shortcomings and set about to temper them.  To accomplish this we talked to industry representatives, we sought advice from potential customers, and we strove to understand customer practices and pain points.  By the time our first HV-100 went into service we thought we’d made a pretty good job of it.  Unfortunately, asking probing questions proved a poorer guide than possessing relevant experience.  The right questions to ask weren’t always obvious, and customers couldn’t tell us what they didn’t know.

Requirements

Our very earliest estimate of the bill of materials (BoM) for our robot was about $750.  Freshly graduated from the frugality of Roomba we felt as though we should be able to build our new robot spending only that much for components.  Then we hired engineers and began developing the robot.  

We never wrote a product requirements document (PRD) for Roomba (in my then-state of innocence I hadn’t learned of that management contrivance).  But the Harvest team would be much larger than the Roomba team and folks with organizational expertise insisted that a PRD was essential for the Harvest robot.  Extensive talks with customers and thoughtful discussions among the founders and others ultimately produced a PRD.  Our newly hired team of engineers—professionals all—took that charter very seriously.  

A couple of stipulations in the PRD were that the robot should have a lifetime of at least 10,000 hours and that customers shouldn’t be asked to make any of their own repairs to our high-tech device.  In pursuit of those and other requirements costs increased.  For example, the lifetime requirement meant that we couldn’t use the relatively inexpensive electric motors designed for the automotive industry as we’d planned but instead needed to develop much more costly custom components.  

But at every stage we vetted with customers the increasing retail price those costs implied.  They didn’t blink.  At each upward step customers told us that our price was reasonable given what the robot would do.  In the end our BoM grew to more than 20 times our initial (naïve) estimate and the retail price climbed to over $30,000.  

The production robot worked great.  After several cycles of early reliability problems followed by redesigns with beefier components, the HV-100 performed just as we had envisioned.  In the benign environment of greenhouses robots could move plants all day without a hitch.  At the opposite end of the spectrum, in outdoor facilities where irregular maintenance of the growing beds left potholes, bumps, and mud, the HV-100s still worked—although they required recurring resets from the designated robot wrangler.

A total of about 20,000 horticultural operations were scattered across the US.  Many were too small to benefit from our robots; but those growers handled relatively few plants.  Most plants were sold by larger growers and with one to two billion plants produced each year in the US alone, there was enough work for a great many robots.  We ultimately found over 30 customers across the country and from them we collected as much data as we could on robot usage.

The hockey stick moment predicted in our investment pitch had not arrived.  Attracting more customers proved to be tough, and surprisingly, none of our existing customers used HV-100s to move more than a small portion of their plants.  Given the robots’ great performance, that was puzzling.

Alignment

Despite our sincere and tenacious efforts to understand our customers’ needs, it turned out that we missed several crucial points.  The uniqueness of the robot made one disconnect inevitable.  Since the HV-100 was the first of its kind, nursery and greenhouse growers had  zero experience with similar products.  So when we asked them what features the robot should have and how they would employ the robot, they followed Barney the purple dinosaur’s advice to his young TV viewers—they used their imagination.  Sadly, imagination didn’t always align with actual practice.

A few of our customers planned meticulously.  On any given day they knew which tasks were going to be done, what resources were required, who would be assigned to the direct task and the associated logistics, and how long each task would take.  But most customers were more growers than planners.  Typically, these customers looked around each morning and assigned workers to tasks on the fly, depending on what was most urgent.  Our robots fit awkwardly into the latter scenario.  The HV-100s needed to have been charged, they needed to be moved to the right growing bed, and the plants they were to space usually needed to be transported and then unloaded from wagons that carried them and placed on the ground.  Also, someone needed to layout the boundary marker, adjust the task configuration parameters, and keep an eye on the robots as they worked.

All that prevented the robots from being a simple drop-in replacement for human spacers as we had passionately intended them to be.  To employ the HV-100s effectively involved more planning than growers were used to and they often skipped that part.  That led to reduced usage.  From the data we collected we discovered an unsettling fact.  Where we had expected growers to use each robot about 2000 hours per year, they were in fact using them an average of only about 500 hours per year. 

We had arrived at the 2000-hour figure through discussions with growers before and during robot development.  But because the robots ended up doing about one-fourth as much work as anticipated, their price would ideally have been about one-fourth what it actually was.  We’d missed that boat.  Our BoM enforced the price; we’d lose money if we charged less.

We discovered another subtle point.  Customers didn’t mind at all doing their own maintenance and repair.  In fact they preferred it.  Most had in-house repair departments that kept their other equipment in working order and were happy to add our robots to that list.  That led us, after the fact, to create repair manuals to allow our customers do things like change out motors and fix broken grippers.  

Had we understood this fact of life we would have designed the robot in a simpler way.  We’d have chosen shorter lived, less expensive components.  Growers would have easily replaced worn out parts and the robot’s price could have been closer to the ideal.

That’s on me.  The reliability number we insisted on in our product requirements document was a guess representing a tradeoff between robot performance and robot price.  We could (and should) have made the equally plausible guess that lower-cost components, replaced during the lifetime of the robot, would be acceptable.  That would have led to a lower priced robot thus lowering the barrier to adoption.  Robot longevity is an abstract concept to early adopters who are uncertain that the robot will work at all.  Price is front and center.

Venture Scale

But so what if our price was higher than optimal and our utilization rate less than desired?   We had a superbly working product.  We had recuring income from robots we rented out and from service contracts with customers who’d purchased HV-100s.  Collectively our robots moved millions of plants per year.  Everything was great, right?

Not the way our investors saw it.  They gave us the cash with the expectation that we would return many times their investment to them.  That didn’t make them greedy, it made them prudent.  Because startups are so risky, venture capital firms go out of business unless the few portfolio companies that do succeed perform spectacularly well.

Spectacular performance would occur only if Harvest Automation grew rapidly.   We’d need more money to make that happen—either through increased sales or through some new approach.  To get there our investors insisted that we hire a new CEO with more fund-raising experience than our founding CEO, Charlie Grinnell.  In late 2012, John Kawola, a veteran of the 3D printing industry, took over as Harvest CEO and Charlie became COO.

Pivot

We learned a great deal about growing plants in pots, about our customers practices’, and about the nursery and greenhouse industry in general.  In hindsight it’s much easier to see what we should have done!  The eternal optimist, I’m convinced that I now know how to redesign the system in a way that would lead to the nursery and greenhouse industry revolution we intended. 

But one rarely gets a do over.  When things don’t go as planned in a startup, standard practice is to pivot.  And pivoting to a different industry was what came next for Harvest.