What happens to your north star metric when your best users never open your app? Not because they churned – because they delegated. A growing share of interactions with digital products in 2026 aren’t initiated by humans at all.
They’re initiated by AI agents acting on human intent. And most product teams are measuring none of it.
Daily active users. Session length. Engagement rate. These were never neutral measurements – they were built on a specific assumption: that value requires a human, on a screen, spending time.
Agentic AI breaks that assumption at the foundation. And if your product strategy hasn’t accounted for it, you’re optimizing for a world that is quietly disappearing.
Arjun Barli
From operator to delegator
In traditional digital interactions, users were operators. They navigated interfaces, filtered results, and manually executed every action inside a product.
That model shaped a decade of design thinking – every microinteraction, every onboarding carousel, every engagement loop was built for a human with a thumb on a screen.
That model is changing. Users are increasingly becoming delegators, expressing intent in natural language and expecting autonomous agents to fulfill it on their behalf.
Take a frequent traveler today.
Instead of opening multiple airline apps, comparing tabs, and entering payment details, they tell their agent:
“Find me a direct flight to Tokyo for early April, business class, under $2,000.”
The agent queries systems, checks constraints, and returns a confirmed booking. The traveler never touched an interface.
In that interaction, the entity your product served wasn’t a human – it was software acting on a human’s behalf. If your product isn’t designed for that reality, it didn’t just underperform. It was invisible.
Andrew Lovell
Why engagement metrics are losing their meaning
Traditional UX was built around human cognitive strengths: visual hierarchy, progressive disclosure, interfaces that reward exploration. These remain valuable. But they create serious obstacles for autonomous systems that need structured, unambiguous, machine-reproducible actions.
An agent can’t appreciate a beautiful interface. It needs an endpoint.
When a product delivers value solely through visual interaction, an agent must resort to screen scraping and emulation – brittle workarounds that are unscalable and prone to failure.
It’s also important to be precise about where this applies. For entertainment, social, and discovery-driven products, engagement remains the right measure, a user lingering on Spotify isn’t failing to fulfill an intent; lingering is the intent.
But for task-completion products – travel, finance, logistics, professional SaaS, healthcare, e-commerce – time spent is friction, not value.
The handoff nobody has designed
Even within task-completion products, there are two distinct modes that most teams have never explicitly separated. In discovery mode, the user is browsing, comparing, exploring, intent hasn’t yet crystallized, and engagement here is intentional and valuable.
In execution mode, intent has crystallized, the user knows what they want, and every additional step is friction. Agentic AI doesn’t eliminate discovery, instead it collapses execution.
The human stays in the loop for inspiration and preference-setting; the agent takes over the moment intent crystallizes.
That boundary, the handoff from discovery to execution, is the most important design decision in your product right now. Most teams haven’t drawn it deliberately, which means someone else will redesign it for them.
Andrew Lovell
A new metric: Return on intent
If session length was the north star of the attention economy, the age of autonomous agents demands a different compass: Return on Intent (RoInt).
RoInt asks a deceptively simple question: when a user, or an agent acting on their behalf, initiated an intent, did your product deliver the right outcome, within the right constraints, without requiring human correction?
While 88% of organizations have implemented AI to some degree, only 23% have scaled agent systems into core business functions, according to McKinsey. That gap between experimentation and execution – is precisely what RoInt is designed to close.
It gives product teams a metric that reflects what agents actually do, not what dashboards were built to measure. McKinsey estimates that by 2030, agent-based systems could generate $450–$650 billion in annual revenue in mature industries.
But the failure rate is real:
It will all be down to whether they redesigned their products and their metrics around intent.
Is it already happening?
Some companies are already operating by this logic, even if they haven’t named it.
Stripe’s Agentic Commerce Protocol, launched in late 2025, is a direct embodiment of the principle: rather than forcing AI agents to navigate a visual checkout interface, Stripe built an open standard that lets agents transact programmatically, treating fulfillment as the product, not the UI surrounding it.
Klarna offers the outcome data. After redesigning its customer service around agent-driven fulfillment, Klarna reported that resolution times dropped from eleven minutes to two, repeat inquiries fell by 25%, and customer satisfaction held steady.
Their journey also illustrates the limits, agent fulfillment works for well-defined intents, but breaks down when context is ambiguous or stakes are high.
The discovery layer and the human override still matter. Neither company measures success by how long users spend in their product. They measure whether the intent was fulfilled, and their RoInt is the number that tells them if it was.
So, what changes?
- Ask a different question in your next sprint review. Stop asking “how many users are engaged with this feature?” Start asking “how many intents did this feature successfully fulfill, and how many required a human to intervene?” You don’t need new infrastructure to start. You need a new question on the whiteboard.
- Draw the handoff line deliberately. Map your product’s discovery layer and execution layer explicitly. Where does browsing end and intent begin? That boundary is where agentic AI will enter your product first. If you haven’t designed it, you haven’t designed for what’s coming.
- Treat your API as a product surface, not plumbing. If a well-instructed AI agent tried to complete your product’s core task today, without screen scraping, without emulation, without a human in the loop, could it? If the answer is no, your product is invisible to the agents increasingly acting on users’ behalf. Stripe didn’t build its Agentic Commerce Protocol as a developer convenience. It built it because the interface layer is becoming optional.
- Make RoInt visible in your analytics. Add one new dimension to your dashboard: agent-initiated interactions. Track completion rate, constraint adherence, and intervention frequency separately from human sessions. A product with rising RoInt and falling session time isn’t losing users, it’s serving them better. That distinction matters enormously when presenting to stakeholders still anchored to engagement benchmarks.
The question worth sitting with
Here is the uncomfortable question every product leader should sit with: if an AI agent could fulfill your product’s core value proposition without a single human ever opening your app, would that be a failure, or the highest possible expression of what you set out to build?
The instinct is to say failure. No sessions means no data, no upsell surface, no engagement loop. But that instinct is the attention economy talking, and it’s increasingly out of step with what users actually want.
The products that will define the next decade won’t be the ones users love to use. They’ll be the ones users trust to act. That’s a fundamentally different design brief and a fundamentally different metric, and most product teams haven’t written either one yet.
Rohan Mitra, Product Manager at PhonePe | Building in the SaaS Space.
