Tuesday 23 October 2018

Smart marketplaces bridge the implicit and complex

Marketplace businesses are intrinsically linked to the technologies that enable them. There would be no Craigslist without email/SMTP, no eBay without the graphical browser and no Uber without location services and ubiquitous smartphones. As enabling technologies have evolved, marketplaces have grown to facilitate increasingly complex exchanges in new environments and industries. The next level? They will get smart: goodbye marketplace 1.0, hello smarketplace.

The theory is that dense edge computing and machine learning will enable marketplaces to understand more complex demands, and thereby facilitate transactions currently impossible using the present models. Before moving into a discussion on the “next era” of the marketplace, though, let’s review the history, to plot the course of marketplace evolution to date.

In the beginning, there was Craigslist. The progenitor of all digital marketplaces led the way, and its enabling technology was dial-up internet and email! It was the digital reproduction of the printed classified ads that early internet users were used to IRL. Craigslist grew rapidly and organically, but was hard to trust. There was no payments mechanism baked in, meaning users had to set up deals and payments “off marketplace.”

Things evolved. “Web 1.5” brought online payment technologies and better user interaction thanks to glossier programming frameworks such as CSS and HTML+. There were also simply many more users, increasingly ready to pay for goods online. Hence the rise of eBay, which later acquired PayPal (and later still, Amazon).

Alongside eBay rose the generation of the “vertical marketplace.” These had clear limitations: marketplaces were commonly restricted to a niche product, and (outside of eBay motors) users were generally unwilling to pay for high-value items online.

The co-emergence of social media and the camera-phone dramatically altered marketplace models. Social media removed trust barriers between users, and the ease of taking high-quality photographs made selling goods easier, and increased trust on the part of the buyer. This movement saw the emergence of marketplaces such as Etsy,* where users purchased specialist items from a vast range of sellers able to showcase both craft and personal stories.

As transacting became easier, users were further encouraged to sell and buy, and the marketplace platform became the (often vertically focused) clearinghouse that sat between them. Instances include, on the consumer side, Airbnb for travel and Habito for mortgages and Convoy, for trucking, on the b2b side.*

The latest species is the mobile, on-demand marketplace. Uber is the archetype — enabled by location technologies, increased mobile user density, flexible labor and behind-the-scenes supply-facing tools that bridge atoms to bits: scheduling, direct deposits and earnings dashboards.

The next evolution in marketplace businesses will be substantial. Enabled by sensor technologies and AI, marketplaces will be able to understand and satisfy complex multivariate needs.

The next generation is intelligent

Until this point, marketplaces relied on small numbers of variables to inform matches and transactions. For example: I tell Airbnb I would like to rent an apartment in Paris for a week in July and let it know my budget, and it shows me the options. While the pricing in the background may be extremely sophisticated, the customer’s explicit “expression of need” is a fairly straightforward equation.

Now, a new set of technologies will enable a next generation of marketplaces. AI will allow marketplaces to process richer data, and to therefore understand complex multivariate needs. So in the Airbnb example, combining previous trips to infer taste + available flights to balance costs + local events of interest to surface a packaged journey that is more likely to convert.

Smart marketplaces directly link complex demand to supply by understanding multiple needs.

Or take a common manufacturing process such as laser cutting — we have been meeting several online entrants into the space this year. Historically, this has been a non-trivial quoting process which can require expert CAD engineers, and a lot of back and forth between customer and machinist. To determine the right price, these experts consider the (available) laser-cutting machine, the desired material and specifics of the work. This is understandable for an offline over-the-counter transaction.

This is where machine learning could enter the conversation. With a rich history of orders and their resulting price, you could train a neural net that would ingest the 3D file and spit back out an accurate cost, without the customer explicitly defining the parameters. This would leave the factory to set a target margin, and by extension, bring the supply-side online to a manufacturing platform where the Boschs of the world could source parts.

In this way, smart marketplaces directly link complex demand to supply by understanding multiple needs.

The double cold start problem

Building a marketplace is not without challenges, notably achieving liquidity. Starting a (non-smart) marketplace business is difficult if you need to begin the supply and/or demand from a “cold start.” For example, consider the problem of starting a Deliveroo competitor with no customers and no restaurants on the platform.

Smart marketplaces are even more complex to set up. In addition to needing demand and supply-side engagement; the matching algorithm needs to be trained. To use a real example, Uber Pool could not exist without Uber providing the training data: It needed the pickup and routes data from the solo rides to optimize the communal rides.

Therefore, smart marketplaces face a “double cold start.” Not only do they have to populate supply and demand, but they also have to train their matching algos before they can be effective.

Alternatively, as platforms scale, concomitantly increasing the volume of training data, does the platform choose to eat the prediction errors (and consequently their profit margin), the better to train the model? If so, when does “break even” happen? It is worth it?

In the short term, this all prompts the question: Are horizontal smart marketplaces feasible, given the quantity and quality of training data required? If so, which categories could currently support it? The model may be tested in hyper-specific and data-rich markets such as manufacturing before it becomes more widespread.

These are not the only considerations for and limitations to the model. Builders of smart marketplaces need to ask themselves: Is AI a core competency “baked into” the platform, or is AI a “partner” technology that is brought in? If the latter, what does this imply for defensibility?

What next?

Smart marketplaces are coming: They will first emerge in analogue industries producing complex, high-dimensional data that humans are not good at describing: industry (customer part orders), content (stock footage exchanges), healthcare (MRI/CT related) and finance (commodities trading).

The current crop of marketplaces are likely to layer AI on top of their current offerings to reduce friction and increase efficiencies, but the truly smart marketplaces will be those reliant on AI to broker previously impossible transactions.

*Full disclosure: Mosaic Ventures is currently invested in Convoy and Habito. One of my partners, Simon Levene, was previously an early investor in Etsy.



from TechCrunch https://ift.tt/2ysWExC

No comments:

Post a Comment