OLIX Just Raised $312M to Kill the HBM Bottleneck. Here's Why Photonic AI Inference Is the Most Important Chip Bet Since CUDA.

Mixpanel's 2026 State of Digital Analytics — drawn from 22 billion user actions across eight industries and nearly 290.8 billion AI-specific events — contains a finding that should make every AI product manager examine their metric framework: engagement in North American AI products dropped 38% year-over-year, even as device adoption grew 26%.

If this were a traditional SaaS product, a 38% engagement drop would trigger a product emergency. A board meeting. A root cause analysis deck. Layoffs on the growth team.

For AI products in 2026, Mixpanel's analysis suggests it might signal the opposite: that automation is working.

The Efficiency Paradox in AI Product Data

Traditional product analytics was built on a simple model: more engagement equals more value. Users who open the app more often, click through more screens, and complete more actions are users who find the product valuable. The entire discipline of product analytics — DAU/MAU ratios, session depth, event funnels, feature adoption rates — was constructed on this foundation.

AI products break this model at the architecture level.

When an AI product successfully automates a task, the user interaction count goes to zero or near zero for that task. The user who previously spent 45 minutes drafting a weekly report, clicking through five screens and editing three drafts, now opens the AI product once, initiates the task, and receives a completed draft. The engagement event count for that workflow fell by roughly 90%. The value delivered increased substantially.

As AI products mature, this pattern compounds. Tasks that required multiple prompts in early 2025 now complete in single runs. Tasks that required active oversight in late 2025 now operate in background without user prompts. The engagement count tracks the friction left in the product, not the value delivered through it.

Mixpanel's 2026 AI benchmark data makes this explicit: "Falling engagement can signal efficiency, not abandonment." This is not a consolation for bad metrics — it is a structural observation about how AI products differ from the products that metric frameworks were designed to measure.

What 290 Billion AI Events Actually Show

Mixpanel's AI benchmark covers 290.8 billion AI events across 2.61 billion devices. At that scale, idiosyncratic patterns wash out and structural dynamics emerge.

The aggregate picture:

Device adoption is up, engagement is down. North American AI product device adoption grew 26% year-over-year. Engagement per device fell 38%. The divergence means more users are relying on AI products for more tasks — but each individual reliance event is shorter and requires less active user participation.

Stickiness is lowest where adoption is highest. North American AI products have a DAU/MAU ratio of 21%, the lowest of any global region despite having the highest raw user volume. LATAM, with the smallest user base, posts the highest stickiness at 37%.

AI bot traffic is growing 8× faster than human traffic. AI agent traffic to digital products surged 187% year-over-year, growing eight times faster than human-initiated traffic. This is partly compositional — more agentic AI products initiating automated processes — and partly behavioral: users are increasingly delegating browsing, research, and task initiation to AI agents rather than doing it themselves.

The regional stickiness divergence is particularly instructive. LATAM's 37% DAU/MAU is not a sign that LATAM AI products are simpler or less automated — it is a sign that LATAM's early adopter cohort is integrating AI tools more deeply into daily workflows, returning to the product every day because it is embedded in processes they perform daily. North America's 21% DAU/MAU reflects both the efficiency paradox (automation reducing touch points) and a broader pattern of exploratory AI tool adoption where users try many products without embedding any of them deeply into daily practice.

The DAU/MAU Trap

The DAU/MAU ratio became the dominant product health metric in the consumer social media era because it perfectly captured the health of products designed to maximize session frequency. Facebook, Instagram, TikTok, and Twitter derive value from daily return visits; a high DAU/MAU ratio in those contexts genuinely signals product health.

AI products have a different value delivery model. Consider the categories of AI product interaction:

The product with the highest DAU/MAU in this table is the exploratory use case — the product that users return to frequently but do not rely on deeply. The product with the lowest DAU/MAU is the asynchronous background agent — the product that users have integrated so deeply they no longer need to touch it at all.

Standard DAU/MAU measurement ranks these exactly backwards. It rewards the product that users are still figuring out and punishes the product that is working perfectly.

Activation rate is a more durable north star than raw engagement — and even activation frameworks need recalibration for AI products where the "aha moment" may be the first time a user does not have to do something because the AI did it for them.

The B2B Performance Gap

Mixpanel's 2026 B2B benchmark report — covering 577 billion events across 3.8 billion devices — identifies a structural gap between B2B software products that are widening their lead and those losing ground, independent of market size or product category.

The winning pattern has three components:

Deep workflow embedding. Products integrated into processes users perform every day — not accessed discretionally but triggered automatically by work events — show substantially better retention than products accessed on an as-needed basis. The product that appears in the user's workflow, rather than requiring the user to initiate a visit to the product, is capturing habitual reliance.

Compressed time to meaningful value. The B2B products outperforming benchmarks in 2026 are those where users reach a clear value moment in their first session — not after an onboarding journey, not after a week of data setup, but within the first 30-60 minutes of use. The benchmark for PLG 2.0 has shifted from "under ten minutes to value" to "value before account creation." AI-native B2B products delivering value in the initial interaction are compressing the time-to-retention that traditional SaaS products take weeks to achieve.

Habit conversion within two weeks. The leading predictor of long-term B2B retention in Mixpanel's data is whether a user returns to the product within the first two weeks of adoption at a regular frequency. Users who access a B2B product at least three times in the first 14 days show retention curves that flatten at a significantly higher rate than those who do not reach that threshold. The implication for product teams is that the first two weeks of a user's experience are disproportionately important relative to any subsequent intervention.

The regional performance data underscores this. APAC leads all regions in B2B one-week retention (8.3%) and weekly retention (77.9%). North America records the lowest one-week retention (5.0%) and lowest weekly retention (44.6%). This is not a product quality differential — North American B2B products are not structurally worse than APAC equivalents. It reflects different adoption patterns: North American B2B users have higher average product trial rates but lower deep-integration rates, spreading usage across more products without embedding any of them deeply enough to drive habitual return.

What Product Managers Should Actually Measure

The Mixpanel data creates a practical problem: if engagement metrics mislead, what should AI product managers track instead?

The framework Mixpanel's 2026 report implies — and that is consistent with the behavioral patterns the data shows — has five components:

1. Outcome completion rate. Did the AI accomplish the task the user initiated, regardless of how many interactions it required? This measures product efficacy directly rather than through a proxy of engagement volume. An AI product with a 90% outcome completion rate at 2 interactions per task is performing better than one with 70% completion at 8 interactions, even though the latter generates 4× the engagement event count.

2. Automation depth. What fraction of initiated use cases complete without user intervention after the initial trigger? This is the measure of how successfully the product is eliminating friction from the AI interaction. An increasing automation depth metric — even as it drives down engagement — is a leading indicator of the retention and expansion that follows when users integrate automated processes into their workflows.

3. Value frequency. How often does the user receive an outcome they consider valuable from the product, regardless of whether they actively touched it? This separates value delivery from user interaction, which is the core mismatch that makes engagement metrics mislead for AI products. For an asynchronous agent that sends a daily summary email, value frequency is daily even if the product's DAU count is near zero.

4. Two-week habit formation rate. What fraction of new users return to the product at least three times in their first 14 days? Mixpanel's data shows this as the leading B2B retention predictor. For AI products specifically, the early habit formation window is the period during which users are deciding whether to integrate the product into daily workflows or discard it. Tracking the habit formation rate and testing onboarding interventions against it is a higher-leverage activity than optimizing any downstream engagement metric.

5. Net outcome expansion. At the account level, are users initiating more task types and higher-value task types over time? This is the B2B expansion metric that maps to net dollar retention — and it captures the pattern that distinguishes products that are embedding deeply (expanding use cases) from products that are stagnating (same use case at same volume). The time-to-value benchmark data suggests that products expanding use case depth within 90 days have substantially better 12-month retention curves.

The Diagnostic Question

DAU/MAU contamination from AI agents is a separate but related problem: as AI agents initiate sessions on behalf of users, the denominator in engagement calculations inflates with non-human traffic. Mixpanel's finding that AI bot traffic grew 187% year-over-year and is growing 8× faster than human traffic suggests that many AI products' engagement metrics are already significantly distorted by agent-initiated sessions that do not reflect human user reliance at all.

The diagnostic question for any AI product team looking at their engagement data is not "why did engagement fall?" but "what happened to the tasks that drove engagement in the prior period?" If engagement fell because those tasks are now automated — because users are getting the output without performing the interactions that generated the engagement — then falling engagement is the proof that the product is working. If engagement fell because users started the tasks and gave up, or stopped initiating tasks they used to initiate, that is a different problem requiring a different diagnosis.

Mixpanel's benchmarks give product teams a calibration reference: North American AI products run at 21% DAU/MAU stickiness at scale. If your product is at 35% stickiness and engagement is stable, you may have a different problem than you think — your product may be requiring more user interaction than value delivery warrants, and the 35% stickiness might reflect friction that users haven't yet abandoned rather than genuine daily reliance. If your product is at 15% stickiness and engagement is falling, the Mixpanel framework says the diagnostic question is what the 15% of daily-returning users are getting from those sessions — if it is high-value outcomes, the 85% of monthly users who are not returning daily may be getting equivalent value through automated delivery.

The Implications for AI Investor Due Diligence

The AI engagement paradox has a direct implication for how investors evaluate AI companies at the product metrics stage.

A founder presenting engagement metrics that look weaker than legacy SaaS benchmarks should not be apologizing for the data — they should be explaining the automation thesis behind it. The right question from an investor is not "your DAU/MAU is 18%, what's wrong?" but "your automation depth is rising and your DAU/MAU is falling — what does your outcome completion rate show, and is net dollar retention growing?"

The 2026 SaaS benchmarks are increasingly bifurcated: companies that have successfully automated core user workflows are showing strong NRR and expansion metrics alongside weak traditional engagement metrics, while companies in between — not automated enough to show the engagement decline but not embedded enough to show strong retention — are struggling with both metrics simultaneously. The bifurcation is the tell. Traditional SaaS metrics punish the best AI products and flatter the mediocre ones.

For investors doing product diligence on AI companies, the key metrics to request — and the questions to ask when they are not proactively provided — are outcome completion rate, automation depth trend (quarter-over-quarter), two-week habit formation rate by cohort, and net outcome expansion at the account level. DAU, MAU, and session engagement are context, not conclusion.

Rebuilding the Metric Stack

The transition from engagement-centric to outcome-centric product measurement is not purely conceptual — it requires instrumentation changes, dashboard rebuilds, and in many cases a re-education of the stakeholders who have historically received weekly engagement reports and calibrated their intuitions about product health against them.

The practical path for AI product teams:

First, audit what your current engagement metrics are actually measuring. For each major event type in your analytics stack, ask: does this event occur because the user is getting value, or because the user is doing work to get value? Events in the second category are friction proxies, not value proxies.

Second, instrument for outcome completion. Define what "task completion" means for each core use case in your product, and build instrumentation that fires when the AI delivers the completed output — not when the user performs the final action. This is a different event from "user clicked submit."

Third, instrument for automation depth. Track what fraction of use case completions required zero user intervention after the initial trigger, one user intervention, and multiple user interventions. The distribution of this metric over time tells you whether your product is becoming more automated or more demanding.

Fourth, set up cohort-level two-week habit formation tracking. This requires a funnel view that is not available in most default product analytics configurations — you need to see, for each new user cohort, the fraction that reached three product interactions within their first 14 days. Build this view before you need it, because it is the leading indicator that tells you your retention curve before the retention curve arrives.

Fifth, align stakeholder reporting to the new metric framework. Presenting automation depth and outcome completion rate alongside DAU/MAU to leadership teams that have calibrated their intuitions against engagement metrics requires explicit explanation of why the new metrics are better leading indicators. The Mixpanel data — 290 billion events, eight industries, four regions — is the best third-party evidence available for why the old framework misleads on AI products.

Takeaway: The 38% engagement decline Mixpanel found in North American AI products is not a warning sign — it is the signature of automation working. Products that surface it as a problem to solve are optimizing the wrong variable; products that understand it as evidence of deepening user reliance are positioned to see what the engagement data is actually showing: that users are getting more value per session, per week, and per dollar of product cost, exactly as the AI product thesis promised. The metric stack that was right for social media, SaaS, and consumer apps was wrong for AI products the moment users stopped touching the product to get outcomes from it. The rebuild is not optional.