What Would Make an AI Assistant Worth Paying For?

This discussion with David Pawlan of Assistant Bench unpacks the explosive growth of personal AI agents, exploring their capabilities, current landscape, and future implications. Topics include diverse agent use cases from daily admin to finance, the evolution of human-agent interfaces (voice, hardware), the critical role of proactivity and trust, the emergence of silent multiplayer agents, and the transformative impact on commerce. The conversation also delves into agent-to-agent interactions, economic models, and the profound shift towards a consumer-aligned internet.

September 29, 2026 7 min read

What Would Make an AI Assistant Worth Paying For? — source: https://www.youtube.com/watch?v=3T5sij3spWw

David Pawlan, creator of Assistant Bench, joins Anish Acharya to dissect the sudden explosion of personal AI agents, their current state, and future trajectory, particularly for a technical audience in Machine Learning and AI. They highlight the rapid evolution from foundational models to productized consumer agents, emphasizing the transformative potential.

The Agent Boom and Ecosystem Overview

The discussion kicks off by framing the current AI agent craze as the third major “seeing God” moment in AI, following ChatGPT’s initial impact and the rise of coding agents. This latest wave, fueled by products like Instinct, Muse, and ChatGPT’s latest capabilities, signifies an “extraordinary enthusiasm around personal agents.” The space has rapidly evolved in just weeks, with early adopters like Poke (September 2023) demonstrating novel interactions (e.g., negotiation with the agent). Subsequent releases like those from OpenAI (referred to as “Open Claw” in the transcript), Instinct, Muse, Grok Bot, and numerous long-tail agents (Caddy, Ollie, Palley) indicate a “wild wild west” phase of exploration.

Assistant Bench: A Consumer-Focused Evaluation

Pawlan introduces Assistant Bench, his platform designed to compare AI assistants on a use-case basis. It tests agents with identical prompts (e.g., “Book me a flight to Chicago”) and evaluates their performance across 16 dimensions, focusing on real-world outcomes, follow-up questions, and reply speed. Launched 16 days prior to the recording, the benchmark quickly garnered over 100,000 visitors, reflecting widespread curiosity about the industry’s direction.

Landscape of AI Agents

The current agent landscape is broadly categorized:

B2C Consumer-Facing Generalized Agents: Such as Instinct, Muse, Caddie, Ollie, Season. These aim to be horizontal, performing a wide array of tasks.
Specialized Agents: Within B2C, specific verticals emerge, like travel-specific (Sore, Miso) or email-specific agents.
B2B Workflow Assistants: Examples include Town, Catch, Vellum, which function more as executive assistants in professional contexts.

Out of 122 identified agents, 64 are in the general B2C category, reflecting a strong initial focus on broad applicability.

Key Use Cases and the Invisible Agent

While travel often generates viral attention (e.g., automatic flight check-ins), daily usage patterns reveal different priorities:

Daily Admin: Tasks like inbox management and form filing are highly valued due to their tedious nature.
Agent Orchestration: The interaction and efficiency of agents talking to each other is a significant area of interest.
Development-Focused: Agents assisting with coding on mobile devices or making small design iterations via text.
Finance: While seen as less practical for wealth generation, finance is a compelling category for cost-saving activities. This aligns with the thesis that consumers seek “free money” rather than marginal efficiency gains. Examples include HSA reimbursement filing, automatic flight credit retrieval when prices drop, and smart home automation (e.g., Grok Bot connected to sprinklers based on weather, reducing water bills by 50%). These exemplify the power of invisible agents – those that proactively deliver value without direct user prompting.

Interface and Interaction Modalities

The choice of interface significantly impacts user experience:

iMessage/Text Threads: Preferred by some (Gen Z) for its seamless integration and personal feel. It feels like a “privileged place” on the phone.
Dedicated Applications: Favored by others (Millennials, Gen X), especially for visual-driven tasks like travel planning or mood boarding.
Widgets: Emerging formats like Sky’s iPhone widget offer a different surface.
Voice/Audio: Seen as critically important, especially when users are occupied (e.g., gardening, cooking, biking). ChatGPT Voice, with its recent Gmail and calendar integrations, is highlighted as creating a “magic moment” by enabling hands-free task completion (inbox zero during a bike ride). This reinforces the “invisible assistant” paradigm.
Hardware: The Muse Charm (Meta’s pendant) is discussed. A “hot take” suggests its primary purpose isn’t direct hardware adoption but rather real-world data collection for Meta’s metaverse mapping efforts, utilizing its camera and microphones for ambient, 24/7 context capture. Audio-only glasses are also mentioned as a potential “sleeper hit” for their ability to provide connectivity without social awkwardness.

Multiplayer Agents

Integrating agents into social settings like group chats presents challenges: they can feel intrusive or diminish social capital. Three approaches are observed:

Traditional group chat addition.
Agent networks (e.g., Instinct’s agent-to-agent interactions).
Silent Listeners (e.g., Doc by Shane Mac): An agent joins a group chat silently, takes notes, and pings individuals separately for action items, minimizing disruption to group flow.

It’s noted that current LLM limitations in prose quality and “bedside manner” make direct agent contributions to social connectivity difficult; utilitarian roles (e.g., cataloging collections) are more genuinely additive.

Personality vs. Proactivity

Agent personality (e.g., Poke’s sassiness) is deemed not a defensible moat as it’s easily configurable by users. The true differentiator is proactivity – an agent’s ability to act on a user’s behalf without explicit instruction. This creates “magic” (e.g., checking into flights) but walks a “very fine line.” Overly presumptuous actions (e.g., an agent breaking up with someone) can instantly erode user trust. The “Bong Chayng” incident (an agent hallucinating a middle name during check-in) illustrates the risks involved.

Defining “Assistant” vs. “Agent”

Anish and David clarify the terminology:

Assistant: A lesser form, definitionally assisting with directed tasks.
Agent: Possesses “agency,” making good things happen proactively, akin to a peer. This implies a higher level of autonomy and decision-making.

This distinction also raises questions about social perception and comfort with non-human entities in daily life.

The Future: Agent-to-Agent Interactions and Commerce

Broader Societal Impact

Looking five years ahead, agents are envisioned to:

Relieve Bureaucratic Pressure: Optimize finances, health, and provide access to services (legal, family office) previously only available to the wealthy.
Enhance Human Experience: Reduce screen time, foster more genuine human interaction, and enable exploration of new hobbies or projects (e.g., building a virtual record shop).
Possibility Expansion: The technology isn’t just about cost reduction but about enabling entirely new possibilities that wouldn’t have existed otherwise.

Agent-Native Infrastructure

The conversation forecasts the emergence of agent-native systems – infrastructure and products designed exclusively for agent-to-agent interactions (e.g., agent emails, agent phone numbers). This will redefine service industries, create new cybersecurity paradigms, and catalyze a new wave of startups solving niche problems.

Commerce Transformation

Agents will fundamentally alter commerce:

Shopify vs. Amazon: Shopify embraced Muse, aligning with its goal of democratizing commerce. Amazon blocked Muse, likely to protect its ad revenue-driven model. When agents make purchasing decisions, human eyeballs (and thus ad impressions) are removed, threatening traditional profit structures.
Supply-Constrained Markets (e.g., Restaurants): Agents can overwhelm reservation systems (Resy blocking bots). Future scenarios include power shifting back to restaurants (prioritizing loyalty/spend) or the rise of agent-driven bidding wars.
Demand-Constrained Markets: Aggregation of demand could lead to merchants bidding for consumer interest, benefiting consumers.
Recommendation Engines: Agents will require sophisticated recommendation engines that understand nuanced user preferences beyond simple searches (e.g., “white socks”). Companies with rich user data (like Meta via Instagram/Facebook) may have an advantage. This could lead to a “truly consumer-aligned internet”, where agents prioritize product quality and user preference over marketing or brand, potentially fostering serendipity and supporting long-tail, unique suppliers (e.g., finding a specialized knitter on Etsy).

The Economics of Agents

The economic viability of agents is a key challenge:

High Costs: Running an ambitious agent can cost ~$20 per user per day, translating to hundreds of millions annually for a startup.
Pricing Models: Out of 122 agents, 65 are paid (35 fully paid, 30 freemium), while only 13 are free. This creates a difficult competitive landscape for startups when dominant players like Muse, Instinct, and potentially OpenAI offer services for free.
Future Hope: Rapid deflation in token prices and browser use costs could make free offerings more feasible. However, the ultimate goal is to build agents so compelling that consumers are willing to pay significant amounts (e.g., $1,000/month), demonstrating extreme market fit and benefiting from future cost reductions.