AI Dating and The Enterprise

What can enterprises learn from Bumble's AI dating concierges? When might you use an AI Agent versus a conventional algorithm?

AI Dating and The Enterprise
Your AI dating concierge hard at work

Bumble dating app founder Whitney Wolfe Herd recently described a future where an AI agent would do matchmaking on your behalf through interactions with other agents:

“There is a world where your dating concierge could go and date for you with other dating concierges” [...] “And then you don’t have to talk to 600 people and it will scan all of San Francisco for you and say ‘these are the three people you really ought to meet’.”

If you are able to get past the dystopian hellscape vibes then this does raise some interesting questions for when to use generative AI and agents – whether for business or pleasure.

How can GenAI help with dating?

It wasn't obvious to me how employing AI agents in this scenario would lead to better outcomes than what might be achievable through a conventionally coded algorithm. Using controlled and structured inputs such as location, age, likes and dislikes – the task appears predictable and a conventional algorithm seems appropriate.

However, what if a future Bumble app could use the context beyond your profile? your social feed? photo library? emails? bank transactions? You get the point. With more information an AI concierge may better impersonate the actual you rather than the curated, "instaworthy" you. Ideally this added context would expose your revealed preferences and not just your expressed ones. Your concierge might know you better than you know yourself through the emergent behaviours of generative AI.

Even if you're suspicious of a large language model's ability to be a "wingperson" then you can be reassured that the actual person gets the final say – including an IRL date! This person could also be consulted during the matching "conversation" – providing additional context and direction in ways that might be difficult to predict upfront.

AI Agents in the enterprise

What can we take away from this example when we think about enterprise scenarios? Here are my thoughts.

Uncertain tasks

An engineer might choose to use an AI-based agent over a simple algorithm when the task steps are unpredictable and the inputs are uncontrolled or unstructured – you want human-like general problem-solving capabilities.

Emergent behaviour

An organisation might hope for insights or solutions that can't easily be expressed in a set of pre-defined instructions. Just like the dating example, for an AI agent to bring insights it will require context. Without context a general-purpose AI model will be operating based only on its training data – which won't necessarily be in alignment with your organisation. If, for example, we're building an AI procurement solution then it may need to know:

  • What are the requirements of the product/service we're purchasing?
  • What is negotiable? what is not?
  • Considerations other than price: ESG priorities, tax, legal...
  • the list is possibly endless...

It's very difficult to know what context will be beneficial – particularly if we're hoping for unexpected insights. Instead, it may be better to reach out to a human for further input when required.

There are many different approaches to give context to a GenAI model – all having different tradeoffs. This includes model training, model refinement, retrieval augmented generation (RAG) and prompt engineering. Choices here will impact deployment costs as well as the quality and alignment of the output.

Human supervision

Just as you wouldn't want to be stuck with a dud date, you usually wouldn't want errors in your business output. AI agents will need supervision. The naming of Bumble's "concierge" and Microsoft's "Copilot" is an obvious hint to the near-term autonomy expectations for GenAI.

Full autonomy is only practical if the cost of failure is outweighed by benefits such as cost savings or entirely new income streams. In the near term I see this likely only being applicable to some high-volume, low-value transaction use cases.

Could an AI dating concierge work?

I can see AI being useful as a wingperson to help bring good candidates to the top – but only with careful supervision and ideally consultation. However I do have reservations about a general-purpose large language model's ability to have useful emergent behaviour in the area of matchmaking. Particularly if its training data includes the entire corpus of rom-coms and romance novels.

An AI dating concierge is only going to be valuable if you provide it with A LOT of personal contextual data beyond the common profile attributes and images. I would be hesitant to give this data away to a third-party. We will need solutions to allow for informed consent and control for the sharing of an individual's digital twin.

Bumble may also find it difficult to convince their current user base – many people seem to enjoy the matching process more than the dates themselves.