When Your AI Has Agency

Personal AI assistants may be ushering in the biggest behavior change we’ve seen since smartphones. As amazing as they seem to be, their value grows as they gain relevant context and keep it current. That can mean giving a single service access to your conversations, relationships, purchases, and plans. But we have a history of valuing convenience over privacy. Let’s explore.

My two favorite personal agents are OpenAI’s dots and Meta’s Muse. OpenAI describes dots as always-on agents with their own cloud computers, connected apps, and the ability to learn from feedback. Meta’s Muse combines ongoing memory with tasks such as shopping, travel planning, and reminders. These capabilities make continuity part of the product. It is the critical factor.

The Competition

My personal preference for these two products aside, the competition is already substantial. Grok Bot offers persistent cloud computing, memory, and scheduled routines. Microsoft’s Scout, now called Autopilot, brings an agent with its own computer, identity, and memory into the Microsoft work environment; it remains in private preview. Google’s newly announced universal Gemini agent extends this competition across Workspace.

OpenClaw gives people a self-hosted route to a personal assistant. NVIDIA’s NemoClaw supplies deployment and security controls for OpenClaw and other agent frameworks. It remains early-preview software. Nous Research’s Hermes Agent offers another open-source approach built around memory, reusable skills, and scheduled work.

On the other end of the spectrum, Anthropic is bringing Cowork and chat together in Claude, while Amazon’s Alexa+ approaches personal assistance through household coordination, shopping, and reservations. These products differ in audience, deployment, permissions, and maturity. Their shared competitive question is: Which assistant will you trust with enough context and authority to become useful throughout your day?

How Search and Shopping Change

Persistent memory is central to the answer. More of us will ask an assistant to resolve a need, with our previous decisions and preferences already available. My prediction: agent-initiated searches will grow as consumers conduct fewer searches themselves. This will dramatically change what it means to have a presence on the web. I expect more purchases to begin with a standing instruction and end with either an approval request or a receipt, depending on the authority we grant.

Unless some new, marketer-centric tools evolve, this is going to be a very high hill for marketers to climb. A marketer may never learn that an assistant considered and rejected its product, or which private preference drove that decision.

Private memory needs a precise definition here. OpenAI says advertisers do not receive ChatGPT conversations or memories. On eligible Free and Go experiences, ChatGPT may use past chats and memory to personalize ads when the relevant settings are enabled. Meta says Muse conversations and virtual-machine data are not shared with its advertising systems, while warning that activity on merchant websites can still influence advertising. Policies, product tiers, and the route to purchase all matter.

My prediction is that brands will have to compete harder for inclusion in a privately informed shortlist. Accurate product information, reliable inventory, clear return policies, and credible evidence will become more consequential. An assistant comparing a requirement against a specification needs information it can verify. A smaller company with a genuinely better fit may get an opportunity it could never afford to buy through mass-market reach.

Saying “brands will still matter” tells marketers very little about what to do. Back in the day, some Alexa custom skills required consumers to remember a service’s name and invocation phrase. That didn’t work as well as Amazon hoped because it required consumers to remember unfamiliar syntax. Share of prompt is going to present a similar problem.

Agents can carry old service failures into new purchase decisions. A good experience can reinforce a brand preference. A bad experience can persist, too. Companies should consider what happens when the customer’s assistant remembers the last three service failures.

Delegating Everyday Life

The broader change is in the doing of life. Planning a trip involves schedules, transportation, reservations, budgets, and coordination. Preparing for a meeting involves finding material, reviewing earlier conversations, and deciding what needs attention. As agents become more dependable, I expect people to delegate larger portions of this preparation and spend more of their time reviewing exceptions and making consequential decisions.

That requires boundaries. Remembering something does not confer permission to act on it. A preference can become outdated. An inference can be wrong. Consumers need ways to inspect and correct what an assistant believes, understand what it has done, and limit what it can do next. Current products provide different degrees of control. Meta says Muse users can inspect, edit, and download memory files. OpenAI’s documentation says individual dot memories cannot currently be directly viewed, modified, or deleted. These details deserve attention as we decide how much to entrust to an assistant.

Memory and Switching Costs

Memory will also create switching costs. A competing assistant may have a better model, yet know very little about you. The accumulated context in your existing assistant has value. I expect memory portability to become a serious consumer issue as people discover how much they have invested in teaching a particular service.

I love what Muse and dots are making possible. My experience with Hal has already raised my expectations for every other piece of software I use. I increasingly expect it to understand the assignment, carry relevant context forward, and help me finish.

Which leaves me with one, big, important marketing question: What happens when your customer’s agentic assistant understands their buying decision better than you do?

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.