AI Assistants That Shop, Remember and Connect People

Anish Acharya welcomes Josh Elman and David Pawlan, creator of Assistant Benchmark, to discuss the fast-changing market for personal AI agents. Pawlan recently convened a meetup where more than 100 participants compared ideas and ten builders pitched products.

Pawlan says participants were unusually open to sharing feedback, even though their products overlap. The group sees the market as early enough that founders can still learn from one another about what works reliably and what people actually enjoy.

The speakers expect more than one agent to serve a person. Different assistants could help with scheduling, travel, work, personal reflection or playful social interactions, much as people turn to different friends for different needs.

One participant describes an agent that shared a friend’s power outage in a social context, creating an opportunity to reconnect. The appeal was not saving time on a task, but making an interaction possible that might not otherwise have happened.

Pawlan says he has encountered more than 147 assistants and groups them broadly into general, work, multiplayer and travel products. He has seen fewer mature finance-focused examples and suspects trust is a major reason.

In approximately 3,600 messages from groups of AI power users, Pawlan says 22 percent discussed trust as a concern. This is an observation from his particular sample, not a representative estimate of all consumers.

The group discusses a “fence” analogy: limit an agent’s access to specified data and channels so it can act within an understood boundary. They cite a reported incident in which a personal agent exposed financial information in a work Slack channel as an example of why context separation matters.

One question is whether boundaries can be expressed only as hard rules or whether an agent also needs qualitative principles and judgment. The speakers agree that agents can be capable yet miss social nuance, making permissions and monitoring important.

A useful assistant need not make every decision autonomously. It can research options and tradeoffs, then ask the user to approve the final choice, price and other conditions before acting; trust might expand gradually with experience.

The speakers distinguish stated privacy concerns from actual behavior, while noting that convenience may lead people to share sensitive information. They also argue that accidental embarrassment can be a powerful barrier when an agent speaks or acts on someone’s behalf.

One participant frames privacy and utility as a tradeoff: better service can persuade people to share more, while comparable services may be judged on privacy. The conversation treats this as a hypothesis about behavior rather than a universal rule.

Sending an email on someone’s behalf requires tone and discretion as well as technical access. One speaker says he identifies the AI model alongside his own name when using it to write, so the recipient knows AI was involved.

Pawlan’s flower test asked assistants to find the cheapest flowers that could arrive the same day. He says several found good combinations of price and delivery time, while others satisfied delivery but placed less weight on price.

The experiment illustrates how a task with multiple constraints can expose differences in judgment. A browser integration may also help an assistant navigate available options; Pawlan specifically mentions Kicker’s use of Chrome in his test.

In another informal comparison, three agents planned a luxurious but cost-efficient Los Angeles weekend. Their interpretations differed: some emphasized spending less, while another placed greater weight on maximizing time at the destination.

The hosts ask whether assistants should clarify the tension, since a traveler might accept a red-eye flight but want a hotel room for a morning nap. Pawlan says the agents asked questions, though he deliberately kept his answers vague.

He then placed the agents in a shared conversation for three rounds of itinerary proposals. Pawlan observed one taking useful details from the other proposals and incorporating them into its own, an anecdotal test of agent-to-agent influence.

The flower test deliberately measured two constraints, not the full social context of a gift. The speakers note that a real assistant might also need to consider the recipient’s preferred flowers, a note and whether the price is visible.

Working with agents can require people to make implicit preferences explicit. The group considers how an assistant might learn those preferences without forcing a user into an uncomfortable description of their own identity.

Pawlan discusses Dots and the value some users see in prior conversation context. In his survey of 354 AI power users, he says only 3 percent used it as their primary agent; that small, selected survey does not establish broader market share.

Memory also creates switching costs. Pawlan says he asked the top 15 agents in his benchmark to export what they knew about him, and all but one produced a file he could use elsewhere; he describes the exports as broadly useful, not necessarily complete.

Photos and screenshots could provide context people would not think to type. The group discusses a photo-based product’s stated emphasis on forgetting unnecessary information, while recognizing that granting access to an entire photo library is sensitive.

A participant who built a screenshot-extraction app says the information was interesting but he rarely returned to it as its value decayed. Another suggests organizing saved images into reusable mood boards for design and coding instead of treating them only as a passive archive.

For shopping, the group speculates that agents could discover products from smaller direct-to-consumer brands that struggle with marketing. They also note that being a smaller seller does not by itself mean a product is better.

The idea that agents will make consumers more rational draws disagreement. One speaker argues that an agent may develop its own shortcuts, such as favoring merchants with clean APIs and documentation, which are not the same thing as product quality.

Another speaker doubts that people want an agent to impose an idealized set of choices, such as only presenting diet-friendly meals. He expects users to prefer tools that accommodate their tastes and switch if the tool becomes overly paternalistic.

Pawlan distinguishes routine commodities, where a person may simply want a suitable low-cost option, from preference-heavy purchases that benefit from discovery and entertainment. He expects experimentation beyond general productivity assistants.

The group discusses Wabby’s separate threads for areas of life, such as a shared watchlist of books and shows. A new group conversation that begins with both people and an agent may also make it easier to establish shared expectations about what the agent can see.

Playful “gossipy” agents can ask friends how they met, notice differences in their stories and prompt a conversation. Participants describe this as a source of social delight, while earlier concerns about discretion still apply.

Pawlan says reliability is a recurring frustration: when an agent cannot finish a task, the user may find doing it manually easier. He is watching browser operation, checkout integrations and infrastructure that can make completion more dependable.

He also sees real differences in performance at scale. Some products degrade or suffer outages as users increase, suggesting that operating reliable software remains a challenge even as building prototypes becomes easier.

The speakers compare managed, plug-and-play assistants with local or self-hosted systems. Pawlan mostly tests products that work out of the box for consistency, but sees potential in giving people more control over models and infrastructure.

At the meetup, attendees could send their agents into a separate agent conversation. Pawlan says some exchanges became incoherent after many messages, illustrating both the curiosity of the experiment and the difficulty of making such interactions useful.

The more ambitious possibility is agents discovering complementary interests and asking their users whether they want an introduction. The group imagines study groups, local meetups and other encounters, while acknowledging that people’s desire to meet strangers varies.

They consider other context sources, from smart-home habits and purchase history to conversations at social events. More context could improve personalization, but recording or sharing information about other people raises permission questions.

One speaker warns that search history is not the same as enduring preference: researching a World Cup or a child’s favorite artist does not necessarily mean that interest defines the user. An assistant needs a way to distinguish curiosity from long-term taste.

Asked what may change in the next few weeks, Pawlan predicts more attention to enterprise agents. Others expect assistants to become more entertaining and conversational, with usefulness expressed through companionship as well as completed tasks.

Pawlan describes watching his technology-averse mother use an assistant, including asking it to help find a dog resembling a family pet. He sees this as a prompt to study how people outside the early-adopter community will actually use such products.

The conversation ends with a speculative idea: an assistant might eventually help people pursue aspirations or develop capabilities, not merely clear a to-do list. The speakers describe this as a possible future product direction, not a demonstrated outcome.