Uber Marketing Manager Ishaan Singh told attendees at Business of Apps London 2026 that AI agents now handle roughly 60 to 70% of the manual work his team once did, and he laid out a 90-day plan for other marketers to start shifting from scheduled campaigns to autonomous decision-making systems. His core argument: the traditional playbook of segmented campaigns and weeks-long A/B tests is too slow for markets where user behavior and competitive options change in real time.
Why the old playbook stopped working
Singh identified three pressures that make manual lifecycle marketing unsustainable. Personalized messaging has become background noise-inboxes are stuffed with tailored emails that few people read. Testing a campaign still takes two to four weeks, a full business cycle, while market conditions shift faster. And users now routinely choose among four or more apps for a single task, so loyalty cannot be assumed.
"The old playbook is dead. Right from scheduling campaigns, segmenting audiences, A/B testing, everything has now become redundant," Singh said. His replacement is not a new tool layered onto existing processes. "With AI agents, it's not about a new tool. This is all about a new operating system. We can't use the same tech stack, the same tools, and just have an AI layer on top of it. We need to reinvent the playbook."
What agents do across the customer lifecycle
Singh described the agent's core loop: sensing user behavior, estimating churn or conversion likelihood, deciding on an action, executing it, and learning from the outcome. He was direct about the current limits. "It's not 100% accurate always. It hallucinates a couple of times as well. It's not maybe 90% accuracy, but that's still good enough because it gives the whole marketing team a lot of runway to focus on other things."
Three examples showed the range. Duolingo's agent analyzes streak length and return probability to send a behavior-specific notification rather than a generic reminder-driving a 20% rise in daily active users. Netflix's agent builds a taste profile from a user's first three plays, then monitors viewing patterns, time of day, and binge behavior to time trailer recommendations for the exact moment a user is most likely to watch. Calm's agent identifies when a user is most anxious and likely to engage, then introduces paid content at that moment instead of through a generic upsell.
"Traditionally it would take weeks, manual rules, compliance checks, to get these kinds of tests done. With agents, you can do it in real time," Singh said. He acknowledged that recommendations sometimes miss, copy underperforms, and campaigns occasionally misfire enough to trigger unsubscribes. He treats this as an early-stage cost. "I think within six months, half of the problems would be solved. If we don't think about it now, we cannot act in the next six months."
Adding a brain layer, not rebuilding the stack
A recurring assumption is that adopting agents means overhauling marketing infrastructure. Singh's experience was the opposite. "The best part is we don't have to redesign our entire company's tech stack. It's just one additional layer that sits on top of everything."
The orchestration layer-tools like Customer.io and Iterable-stays in place. Channels and data remain where they are. What gets added is a decision-making layer, an API connection to Claude, Gemini, Grok, or ChatGPT, that takes over analysis and decisions a human previously handled. This approach to AI Agents & Automation means the technology functions as a brain layer rather than a replacement for existing systems. "This brain layer could actually do the work of 100 humans together," Singh said.
A 90-day plan for getting started
Singh advised against scaling immediately. Start with a single metric-activation, retention, engagement, or churn prevention-and build from there. "Pick one metric. Try to add more data to it. What channel or communication is more likely to save your user from churning? What's more likely to activate your user?"
The process: feed historical usage data into an AI tool, get recommendations, connect those to the existing CRM platform, and run experiments. Review weekly without intervening. "Do not change anything. Just let AI think, change and evolve. If the performance is not at par, try to do that. Within three to four weeks, you exactly know what worked and what did not work."
One obstacle he flagged is deciding what data an agent can access-a real constraint in regulated industries like fintech where compliance limits what can be automated. For marketers working in these environments, AI for Marketing training resources often address how to set appropriate guardrails without stalling adoption.
What's left for the human
If agents handle 60 to 70% of previous execution work, the marketer's role changes shape rather than disappearing. "It's not just about our roles being disappeared. It's just about being evolved," Singh said.
What remains: setting guardrails for what an agent can and cannot do, supervising output and catching signals an agent might miss, designing customer journeys that agents then execute, and handling subjective judgment calls. Those include recognizing when a single bad experience turned a loyal user against the brand for reasons no dataset captures. "The agent does the doing, you do the thinking, and then you evolve the agents. They start becoming like your direct reports."
Why this matters for marketing professionals
Singh's 90-day framework-pick one metric, feed it data, connect to CRM, and let the system run without interference for three to four weeks-offers a low-risk path to test whether agents can handle decisions you currently make manually. The goal is not to replace the marketing team overnight but to reclaim the time currently spent on campaign execution so you can focus on strategy, guardrails, and the judgment calls that data alone cannot make.
Your membership also unlocks:
