By

Sherry Jiang

there’s a lot happening in consumer ai right now.

openai just launched chatgpt health – letting users connect medical records, wellness apps, and wearables to get personalized health guidance. google dropped personal intelligence, turning gemini into an agent that can reason across your gmail, photos, youtube, and search history.

we’re entering a world where ai doesn’t just answer questions, but it also takes action on your behalf.

but here’s what’s missing

2025 was the year of doing. 2026 is the year of knowing.

2025 was undeniably the year ai agents went mainstream. ppl got used to chat interfaces, where they could ask questions and get answers.

then agents started doing things:

• amazon released “buy for me” to shop retailers without leaving their app
• openai embedded checkout directly into chatgpt
• coding agents like cursor and claude code went from autocomplete to writing entire features autonomously

this progression felt natural – chat → fetch → automate → execute.

however, there is now a ceiling that we’re reaching with agentic ai – agents can act, but they still need you to tell them what to do (i.e., the whole world of “context engineering”).

it requires a level of user steering that 95% of consumers are not yet equip to fully extract value from – knowing what questions to ask, prompting the right way, providing the right context. and many simply, don’t know what they don’t know, in the absence of a more proactive ai.

the 10x unlock for consumer ai isn’t just better execution – it’s understanding what you actually want before you ask for it.

the intent capture problem

google’s personal intelligence looks genuinely impressive. it can reason across your gmail, photos, and youtube history to surface insights you didn’t ask for – like recommending all-weather tires because it noticed road trip photos and cross-referenced your vehicle from an old email.

but there’s a fundamental gap between what you did and why you did it.

google can see you ordered food delivery 12 times this month. what it can’t see is whether that’s stress eating, a work sprint, new parent survival mode, or a genuine love of trying new restaurants. each implies a completely different “helpful” response. but from the data alone, they look identical.

the deeper issue: your emails, photos, and watch history are byproducts of living your life – not deliberate signals about who you are. and some context simply doesn’t exist anywhere online. you never wrote down “i go to starbucks because it’s my escape from my parents’ house.” it lives in your head, and no amount of cross-referencing across google will surface it.

you could argue: who cares? google doesn’t need to know why you ordered delivery to recommend good restaurants.

but the quality of the intervention depends entirely on the intent behind the behavior.

this is the gap between action and understanding.

this isn’t a google problem – openai’s health product, claude cowork, apple’s siri upgrade, they all face the same structural gap.

ai that acts is reactive – waiting for commands, executing tasks, optimizing for the average case.

ai that understands can be proactive. it knows when to surface something, when to stay quiet, when to push back. instead of waiting for you to figure out the right question, ai that understands can ask you the right questions. and eventually, once it knows you well enough, it doesn’t need to ask at all.

why conversation is the $1T unlock for intent

there’s a reason humans get to know each other through conversation, not just through stalking their online presence and calling it a day.

the model companies are building the infrastructure – better models, longer context windows, improved reasoning. consumer companies can build on top of that: specialized context graphs about user intent and behavior that enrich what the models can do.

but this only works if you have something the labs don’t. that means obsessing over a specific domain – hundreds of user interviews, deep understanding of behavioral patterns, product intuition that comes from living inside one problem.

consumer companies earn the right to exist by asking good questions: the kind that create context that doesn’t exist anywhere online.

the information extraction problem is actually a design problem:

• what’s the minimum number of questions to collapse uncertainty about someone’s intent?
• how do you design inquiry that feels like genuine curiosity, not a survey?

the playbook

if i were building a consumer ai that captures intent, here’s what matters:

1. start with a segment where data inference gets you 80% there. you need initial signal to surface the right questions. spending data, health metrics, productivity patterns – something behavioral that reveals what someone did, so you can ask about why. the best domains have rich passive data that’s still ambiguous without context.

2. be proactive, not reactive. don’t wait for users to ask questions in a chat box. surface insights just in time- like showing a gen z user their black friday shopping patterns the week after, when it’s fresh and relevant.

3. collapse the decision space. don’t overwhelm users with 30 dimensions to fill out. figure out the minimum questions that collapse the most uncertainty.

4. make sharing context feel rewarding. this is where most apps fail. asking for information feels like work unless you give something back. we’ve found that turning what the ai learns into visible artifacts – like collectibles that represent your habits, your rituals, your identity – creates a loop where people want to keep sharing. they’re building something for themselves, not just feeding a system.

5. build memory that compounds. every interaction should make the next one smarter. when someone tells you “i grab coffee with coworkers after standup,” you’re not just logging a transaction – you’re learning: starbucks → latte, companions → coworkers, intent → social ritual.

where this is going

2025 was the year ai learned to act. 2026 is the year it learns to know you. the companies that figure out intent capture – how to ask the right questions, how to build context that compounds, how to turn passive data into genuine understanding – will build something the general-purpose platforms can’t replicate.

that’s where the value lives.