From: redpointai
The role of AI agents is set to significantly shape the future of productivity tools, particularly in email and other workplace applications [00:00:33]. This shift is anticipated to lead towards an “agentic AI future” where intelligent agents operate autonomously to perform tasks on a user’s behalf [00:03:02].
The Agentic AI Future
The vision for AI agents is that individuals will likely have multiple agents, including specialized email agents [00:08:00]. These agents are expected to be mostly autonomous, receiving goals rather than specific tasks [00:08:08]. Their capabilities will include:
- Planning and breaking down goals: Decomposing complex goals into sequential or parallel subtasks [00:08:12].
- Handling ambiguity: Resolving uncertainties by asking clarifying questions or querying other systems, such as internal APIs or CRM [00:08:18].
- Interacting with other agents: Agents will be able to communicate and collaborate with each other [00:08:26].
This AI agent future aims to free users to be “more creative, strategic, and impactful” [00:08:32].
Superhuman’s AI Feature Phasing
Superhuman, an email tool, has adopted a phased approach to integrating AI features, which serves as a blueprint for other applications [00:04:06].
On-Demand AI Features
The initial phase involved “on-demand” AI features that users explicitly trigger [00:06:26].
- “Write with AI”: Launched in mid-2023, this feature transforms a few words into a full email, matching the user’s voice and tone [00:04:21]. It also allows for shortening, lengthening, improving, fixing, and translating text [00:04:32]. These features are generally easier and cheaper to build and run, serving as a way to test user reception and technological capabilities [00:06:33].
Always-On AI Features
The second phase introduces “always-on” features that work continuously in the background [00:06:56]. These are more ambitious, complex, and expensive to operate due to their constant processing needs [00:07:05].
- Auto Summarize: Launched in late 2023, this feature provides a one-line summary of an entire email conversation, updating instantly with new messages [00:04:42]. Users can expand it for a bullet-point summary. This changes workflow as users often only need the summary, allowing them to quickly mark emails as done or snooze them [00:05:00].
- Instant Reply: Released in early 2024, this feature generates draft replies for incoming emails, allowing users to edit and send them, often with minimal or no changes [00:05:30]. This has significantly reduced email writing and sending times, doubling user speed [00:05:57]. It also inspires replies for emails that might otherwise be ignored due to high activation energy [00:09:12].
Agentic AI Future
The third phase, which the “always-on” features lay the groundwork for, is the agentic AI future [00:07:50]. This implies a seamless integration of autonomous AI agents into the user’s workflow [00:07:57].
Design Philosophy for AI Features
Designing effective AI features requires careful consideration of user experience. Key principles include:
- “When you want it and out of the way when you don’t”: This principle aims for a minimal interface that only presents AI suggestions when relevant, such as Instant Reply options appearing only when a user starts composing a reply [00:19:34]. The inspiration to reply is a key value, even if not immediately obvious [00:19:21].
- Matching interaction impedance: For fast workflows like instant replies, the generated content should be short and snappy (1-2 sentences) to maintain user speed and avoid requiring extensive proofreading [00:21:07].
- Interface over prompt engineering: While prompt engineering is crucial, the interaction design (e.g., using Tab and Enter instead of arrow keys for speed) significantly impacts usability and speed [00:22:13].
- Personalized tone: AI should adapt to the user’s voice and tone based on their past communications, making the output sound authentic and human-like rather than generic “AI-written” [00:23:50]. This is achieved through extensive prompt engineering and multi-shot learning (analyzing previous emails), rather than fine-tuning models, to ensure adaptability with model updates [00:25:26].
- Robust evaluation: Due to frequent model updates breaking prompts, a robust regression testing and evaluation framework is essential [00:26:09]. LLMs can even be used to assess their own responses against criteria like addressing the right topic or person [00:27:06].
Economic and Strategic Considerations
The high cost of running “always-on” AI features is offset by shifting revenue streams, such as reducing human onboarding costs [00:33:59]. The expectation is that AI model costs will continue to decrease, potentially running on-device in a few years [00:30:05].
For startups, a key strategy to take on incumbents in the age of AI is to focus on a niche or “underserved yet economically powerful sub-segment” of a market [00:48:42]. Incumbents struggle with:
- One-size-fits-all solutions: Large companies like Google and Microsoft must cater to billions of users, limiting their ability to build specialized, highly optimized products [00:48:17].
- Product speed: It is difficult for established companies with legacy codebases (e.g., 20-year-old client-server applications) to re-architect for modern speed and responsiveness [00:49:15].
- Design: Organizational structures often dictate product design, leading to non-intuitive user experiences (e.g., Google Workspace’s inconsistent navigation) [00:50:06].
Impact on Work and Society
AI’s impact on jobs is seen as both underhyped and overhyped [00:55:41].
- Underhyped: The speed at which AI will automate entry-level roles (e.g., customer support, sales) is underestimated [00:56:02]. If AI can perform a task 80% as well at 10% of the cost, companies will readily adopt it [00:56:13].
- Overhyped: The notion that AI will only affect these entry-level jobs is overhyped [00:56:40]. Eventually, AI is expected to perform most high-level professional tasks, including CEO functions, more effectively [00:57:03].
This transformation is likened to how current jobs would appear as “play” to people from a thousand years ago [00:57:10]. In the future, much of what we perceive as “work” might be seen as “leisure,” though humans will likely still find ways to feel productive and competitive [00:57:47]. The ultimate mission of productivity tools with AI is to help professionals feel happier, more productive, and closer to their potential [01:03:43].
The AI Agent Ecosystem and its Evolution
The precise structure of the AI agent ecosystem remains a “trillion-dollar question” [00:43:37]. It’s unclear whether all agents will reside within a single application or if they will be federated across different services [00:44:21].
- Centralized view: Products like Superhuman, which users spend three hours a day on (similar to WhatsApp, iMessage, and Slack), have a strategic advantage as a “pane of glass” for all work [00:44:51].
- Distributed ownership: It’s unlikely that one company will own every single agent. Instead, each company may build agents relevant to their domain (e.g., an email agent, a Rippling agent for HR) [00:45:45]. These agents will need authentication and cryptographic handshakes to interact securely [00:43:43].
- Future of workplace communication: The goal is to move beyond fragmented communication tools like Slack, which often lead to information overload [01:00:00]. A future workplace experience would combine the best of email (ordered conversations, subject lines, archiving, snoozing) with chat, offering modality choice and allowing users to reach “inbox zero” for company operations [01:01:18].
- Personal executive agents: The ultimate vision is a personal “executive AI” that understands user preferences and priorities, coordinating with other agents (e.g., an HR agent) to manage schedules, tasks, and even complex personal planning (like health insurance for family planning) instantly and autonomously [00:43:24]. This would eliminate anxiety over what to do next by organizing work based on user preferences and calendar availability [01:02:42].