Rise of Autonomous Shopping

Previously in the Future of Retail series, we set the stage for retail automation.? In this installment we will look at how ordering is already moving into a strange new automated world.

Retail’s selling but who’s buying?

The days of human shoppers drifting through endless sites and offers to find the perfect selection are numbered with the emergence of AI.? We are on the cusp of autonomous ordering and many of the precursors are already happening.

As discussed in the previous installment, AI will help reinvent the way we shop.? While this won’t happen overnight, automating consumer purchasing will result in:

1. better buying decisions for consumers
2. greater predictability for sellers
3. greater efficiencies for manufacturing, procurement fulfillment and delivery.??

So what is autonomous purchasing and what needs to happen to achieve it?

Goals of Autonomous Purchasing

One way to summarize the goal of anonymous purchasing solutions (APS)? is to apply a variation of the Turing Test.??

The Turing Test is a classic AI test invented by legendary computer scientist, Alan Turing:?

Turing Test: “A test for intelligence in a computer, requiring that a human being should be unable to distinguish the machine from another human being by using the replies to questions put to both.”? —- Oxford Languages

If we modify this for autonomous purchasing:

Autonomous Purchasing Test:? automated purchases should be indistinguishable from the purchases a human would make when presented with the same choices.

Achieving Autonomous Purchasing

Ultimately, for an APS to be as effective, it must? evaluate all the offers a human could and apply the same criteria to make a selection.? Currently humans purchasing decisions are made after a review across many disparate offers (multiple brands, marketplaces, sellers, etc).? Replicating this process for an AI system will not be easy as a number of things need to change before a APS can mimic - let alone exceed - what a human can do:

1. Define selection criteria
2. Evaluate competing offers
3. Determine success

Identifying selections and criteria

There are many types of purchases a consumer currently makes.? Some have existing sources AI could use to inform them:

1. Repeat purchases - this is one of the easier categories to automate.? Items with regular re-purchase cycles like consumables, personal care items and household products are already addressed by manual subscription solutions such as Amazon’s Subscribe & Save.? The gaps to achieve full automation include programmatic identification of cyclable purchases, identification of purchase intervals, capturing the tolerance for change (i.e. substitutions) and improving the consumer experience by shifting from a static decision to an informed one (“am I out of this product?” “should I buy more early since there is a sale” etc).

1. New purchases - many items not previously purchased or without fixed intervals can still be automated.? Targeted ads have shown the accuracy of many existing ad models at identifying likely purchases.? However, as consumer purchasing is automated, the source of these models will disappear or change (glance views, search results, online research results, etc).? These models are likely to rely on new sensors / collections from consumer entertainment, social feedback or behavior indicators as traditional ecommerce evolves.

There are many other types of purchases such as situational or derivative (“if I buy this drill, I need batteries”).? These will be addressed;? driven by revenue opportunities and new solutions which emerge as APS frameworks are developed.

However, these are all examples of single purchases.? A single purchase is much easier to automate than replacing a consumer’s need to ‘buy anything’.??

Autonomous buying will need to work at multiple levels:??

• Making the best individual purchase based on price, availability, arrival, taste, quality, and past purchases and many more variables.? The weighting of these factors will vary by consumer and require AI learning to achieve quality results
• Evaluating collections of purchases to optimize for consumer top line goals and constraints such as need, budget, integration (home and product ecosystems), restrictions (government, state and local) and many more.? These decisions can apply to individual purchases and combinations
• Re-deciding prior decisions based on events.? Few consumer decisions are static.? Decisions change based on a variety of reasons:? other purchase failures, weather change,? job loss, family changes, etc.? Effective APS’ will need to continuously review and refinement of decisions until a purchase is completed.

Evaluation of offers?

Most retailers have technology moats thwarting a consumer’s ability to easily compare pricing from all sellers for the same offer.? Ironically there has been investment in pricing intelligence and monitoring for retailers.? However, consumers have few tools available for comparing offers across all sellers to make buying decisions.??

Google Shopping is one the most well known price comparison sites.? However, Google Shopping results are influenced by advertising revenue and not unbiased.? Similarly, Amazon surfaces offers solely in its ecosystem (advertised, listed of fulfilled) and uses its own weights for surfacing offers.? Nearly all platforms prevent automated inquiries by would-be aggregators (i.e. bot traffic).??

So in the near future, automation of offers can occur in private label environments but not all-in purchasing scenarios.?

For APS to become a reality, a new protocol which enables offering listings to be collected and surfaced needs to be introduced.? This may be derived initially from some of the same companies building pricing intelligence.? However, in the future, a standard is likely to emerge similar to the evolution of an advertising data consortium.??

Getting it right

While the future of autonomous smart shoppers is almost guaranteed, predicting the widespread adoption is difficult.? However, as that adoption happens, how will the results of the purchasing be evaluated?? How will our Autonomous Purchasing Test be scored?

To get an indication, we can look at a parallel purchasing evolution driven by AI.? Currently AI is being used for automation of ordering restaurants. Essentially:? converting spoken human orders into digital purchase orders.? This is something most people would expect to be a no-brainer given the plethora of existing conversational AI systems.??

However, McDonald’s recently announced terminating a partnership with IBM.? MCDonalds blamed IBM's Automated Order Taker (AOT) technology for ordering failures after a consumer backlash and the corresponding PR nightmare.??

This may be more of an indication of impatience than failure. ? Rolling out new technology requires time and human speech has nuances even for small product catalogs.? Ultimately, few believe this won’t happen despite the current state.??

While this isn’t automated buying, it is a good example of the challenges we’ll see as AI becomes a larger part of the bridge between consumer and purchasing.? Getting desired results will take time, patience and innovation.?

Up Next: In our next installment, we will look at the impact to other aspects of the retail equation driven by APS.?