The Personal Assistant in Your Pocket: What It Takes

Apple's Siri promised us a personal assistant in our pockets 15 years ago. Now in the age of agentic AI, that promise could finally come true, but only with the right infra and awareness of risks.

A Glimpse into the Future

You wake gently to the sound of cooing doves as Somnium wakes you at the perfect moment in your sleep cycle. Vesta has already picked out an outfit for you to wear based on your agenda, and as you enjoy your breakfast, Culina replenishes your supply of your favorite bread and coffee beans. Ordo runs through your day, seamlessly rescheduling your double-booked afternoon meeting and making your children’s biannual dentist appointments. And good news! Via just found a great deal on a flight to visit your parents. It snags tickets for the whole family and makes sure everyone can sit together. You lean back in your chair as you luxuriate in the ritual of the morning—the warm mug between your palms, the gentle steam against your face, the bird on the windowsill, the sound of children playing—all without so much as a glance at your phone.

This is you, assisted by a team of AI agents. All of life’s miscellany taken care of before lunch; all of life’s joys before you for the taking.

While this future might seem far-fetched, we believe that it’s not just on the way—it’s already arriving. Agentic AI promises a new wave of consumer experiences—and, in turn, a new wave of infrastructure to underpin those experiences.

So What Do We Need?

In order to realize this future, we need infrastructural renovations as well as innovations to set the foundation.

Consider what happens under the hood when “Culina” orders your groceries: it needs access to the grocery delivery platform and the underlying payment rails to place the order; memory and privacy guardrails to recall the kind of bread that you like; inter-agent coordination to schedule your delivery at a convenient time and to spend your money; security to enable the merchant to identify and verify its “identity”; permission to make decisions and take actions; and feedback mechanisms to validate its decision-making performance.

We can already see the necessary recasting of several layers of the infrastructural stack. While many of these evolutions are well underway, particularly in enterprise, consumer use cases come with different needs and constraints, and we have not yet seen a similar profusion of consumer AI agents. (OpenAI’s Operator, a browser agent, was launched at the start of 2025 and remains only a research preview with limitations.)

This need spells opportunity for those that can seize it.

Let’s look first at the overall infrastructure that will be needed to power consumer agentic AI. A few of these layers present striking opportunities for consumer AI startups, which we’ll get to in the next section.

Each of these layers represents a major thematic shift:

Agent frameworks and development platforms: How agents are built and deployed

→ The shift: From individual chatbots to teams of agents

Access: How agents interact with the web and other tools

→ The shift: From user interface to agent interface, or user experience (UX) to agent experience (AX)

Reasoning and memory: How agents make good decisions

→ The shifts: From relevant answers to optimal decisions, and from stateless chatbots to stateful agents

Experience: How we experience and interact with agents

→ The shift: from Siri and the search bar to human-like voice and new UX patterns

Payments: How agents can transact

→ The shift: From Apple Pay to agent pay

Observability: How agents are monitored and improved

→ The shift: From objective throughput to subjective outcomes

---

The Consumer Agentic AI Infra Market Map

---

Where We See Opportunity

Successful entrepreneurs know that the biggest challenges present the greatest opportunities. And within the rapidly evolving field of agentic software, we see four big challenges. Each of these is an opportunity—or many opportunities—for savvy startups to establish competitive advantages.

Solving agent access problems:

This is the first big hurdle on the way to making agents truly valuable. In order for agents to be useful they need to be able to do things efficiently. They will need the ability to navigate, retrieve, populate, transact, send, receive, and so on. And they’ll need to do those things easily, quickly, and accurately.

Giving agents meaningful, stateful memory:

Today’s AI models are making great progress on reasoning, but memory remains a big problem—so big that we believe it’s the second major hurdle on the way to truly useful agents. In order for the “personal assistant in your pocket” promise to be fulfilled, agents need to act appropriately, in accordance with your unique, contextual, and ever-changing needs, conditions, habits, preferences, tastes, and so on. Agent memory needs to be long-term, layered, and adaptive.

Delivering terrific user experiences:

In order for agents to win widespread adoption, they will need to free us from our devices in a meaningful way. That’s why the first challenge in agent adoption will be the design of agentic user experiences. Most of the productivity tools today have made us quantitatively and qualitatively more productive—but always at our desks, or on our phones. The future of productivity is in the background—finding you an in-network doctor while you’re driving to work, doing your taxes while you’re in the bath. We believe voice is part of this future experience, both being able to talk to your agent and being able to have your agent talk to others. But we’re interested in other UI patterns. What’s the next search bar?

Ensuring agents are worthy of trust:

Finally—and this may be the biggest challenge of all—consumers will need to know that they can trust the agents they’re considering using. Without trust, nothing will stick. In the enterprise world there are many kinds of compliance frameworks that generate trust. Consumers, for now, are more on their own. And while consumers have readily adopted generative AI chatbots, agents will require another whole level of trust. In order for consumers to adopt AI agents, they will need to trust them with sensitive data: personal, legal, financial, health information, and more. These concerns touch every part of the agentic infrastructure, including what the agent can access, store, and do on your behalf, as well as how developers monitor their agents’ behaviors and ensure that they are performing well and appropriately. Building agents that people can trust will be key.

Where We See Risk

Of course, there are very real dangers with the shift to agentic technologies. Entrepreneurs who want to succeed in this field will need to navigate risks in at least three major areas:

Privacy and security:

Agents will have access to enormous amounts of sensitive personal data. Securing that data and ensuring that it’s neither vulnerable to hacking nor accidentally shared with other agents or organizations is essential not only to building trust, but also to avoiding catastrophic problems. Companies deploying consumer agents will need to pay very close attention to the governance and security of those systems.

Alignment:

This is the AI world’s terminology for ensuring that agents are “aligned” with humans’ best interests. In concrete terms, it means ensuring that agents are good fiduciaries of their users and act in their long-term best interests.

Defensibility:

How can a company build agentic solutions in this space when the giants like OpenAI are building here already—e.g. OpenAI Operator? It is possible to build solutions that are additive to what the tech giants are doing. But ensuring that you’re riding along with them and not getting crushed by them requires careful business planning and strategizing. We wrote about where consumer AI startups can still compete earlier in the year here.

Get in Touch

If you’re building or supporting companies in this area, we’d love to hear from you. Please reach out!