When to use AI vs humans to give product recommendations

People prefer AI-based recommendations for practical products (e.g. a dishwasher), and human recommendations for experiential and sensory products (e.g. a holiday).

We’re seeing more and more retailers offering online human ‘personal shoppers’ instead of automated recommendations. Why?

Note: the term ‘AI’ is used broadly here for simplicity (e.g. a simple recommender algorithm is considered AI)

Previous tip: Make your loyalty rewards a gamble, not a certainty (All tips here)

Tip type: New research (November 2020)

Use AI recommenders for functional products and humans for experiential and sensory products

Impacted metrics: Customer acquisition | Customer satisfaction
Channels: Customer experience | Sales team | Personalization | Product recommendations


Use AI to recommend practical products (e.g. a warm and resistant winter coat, a shampoo to fight dandruff, an informative documentary). People will perceive them as better.

Use a human recommendation (e.g. a personal shopper, “others also bought this”) for experiential, sensory, and enjoyable products (e.g. a fashionable coat, a relaxing scented shampoo, a stand-up comedy show).


  • Recommendations from an AI algorithm are perceived as better when someone is focused on practical value and usefulness. Human recommenders are better when the focus is on enjoyment and experience. For example:

    • When people were looking for a practical real estate investment 59.8% preferred an AI recommender. When they were looking for a fun and enjoyable property 75.7% preferred recommendations from a human agent

    • The same chocolate cake was rated healthier when people were told the ingredients were chosen by an ‘AI chocolatier’ and tastier when they were told a human chose the ingredients

  • The negative effect of AI recommenders for experiential and sensory products can be:

    • Eliminated when the recommendation is made alongside a human (e.g. “AI enhances and supports this human recommendation”). However, it doesn’t perform any better than human-only recommendations in the eyes of customers.

    • Reduced when people’s prejudices are challenged at their moment of judgment (e.g. “I know you may mistrust a robot suggesting you a recipe, but give me a chance, set aside your expectations, and I’m sure you’ll love it!”)

  • If people are looking for something that fits their unique characteristics and circumstances, they will prefer a human recommender (e.g. a human real estate agent) even if their goal is practical (e.g. finding a flat close to their workplace).

Why it works

  • We have a strong prejudice that machines (AI) don’t fully understand what we enjoy in experiential and sensory products (no matter whether that is true or not). However, we trust them in making good decisions about the pros and cons of a functional product.

New around here? Subscribe below at no cost for two evidence-based tips each week.


  • This study is based on current common perceptions of AI. As AI evolves rapidly, perceptions are likely to change with it.

  • It’s unclear how much attitudes towards AI recommenders may change for different product types.

Companies using this

  • One of the biggest appeals of offline retail is the possibility of receiving recommendations from a human employee.

  • Online, most large enough retailers use AI to give product recommendations (e.g. to logged-in users on their website, via email).

  • Since the Covid-19 pandemic, more and more retailers are realizing the importance of scaling human recommendations for online shopping. For example, Massimo Dutti Spain has recently started offering personal shoppers, reachable via Whatsapp. Amazon has started something similar in the US.

  • A few companies use AI in partnership with humans to make recommendations (e.g. Stitch Fix).

Steps to implement

  • Depending on your product, your audience, and your market, decide in which cases functional qualities or experiential and sensory qualities are more important.

  • You can then either substitute or decide to emphasize or de-emphasize AI or human recommendations. For example, Netflix could say “Our algorithm recommends this documentary” but also “Others also liked this horror movie”. TripAdvisor could highlight the AI recommender role for business travelers, but hide it for holiday-goers.

  • Taking it a step further, you could print on your product label “Recommended as the most reliable by our world-class AI” or “Rated as ‘The most romantic’ by our Experience Director”.

  • If in your situation it’s convenient to have AI recommend experiential and sensory products (e.g. online ordering of flower combinations), instead of a human, you could have a human review and ‘sign off’ the AIs choices to eliminate the negative effect.

Study type

Online, lab, and field experiments, United States and Italy


Longoni, C., & Cian, L. (November 2020). Artificial Intelligence in Utilitarian vs. Hedonic Contexts: The “Word-of-Machine” Effect. Journal of Marketing, 0022242920957347.

[Link to paper]


Questrom School of Business, Boston University and Darden School of Business, University of Virginia. United States

Remember: Because of the groundbreaking nature of this paper, it could be disproven in the future (although this is rare). It also may not be generalizable to your situation. If it’s a risky change, always test it on a small scale before rolling it out widely.

Want to share feedback or ask a question? -> Reply to the email or write to thomas@ariyh.com

Was this forwarded to you? -> Subscribe below or read previous marketing tips here