Automation in adtech can often be a double-edged sword. Marketers patently acknowledge that automating rote actions will drive demonstrably better performance. But automation can also lead to adverse edge cases if left unguarded.

Many of us at Gigi have seen this play out first hand. At our last company, Perpetua, an early retail media execution platform, we leaned heavily into machine learning models to automate manual tasks for Amazon Sponsored Ads: real-time bidding strategies, keyword harvesting, keyword negation, and budget allocation. It worked. We quickly established product-market fit and scaled our customer base across mid-market challenger brands, third-party sellers, and growing independent endemic Amazon agencies. After a certain level of sophistication and scale, though, customers would "graduate" from Perpetua once they discovered that unconstrained automation could be detrimental to their business. They gravitated toward products that allowed for an implicit level of control of inputs while tacitly acknowledging the negative performance tradeoffs of manually controlled workflows. Our journey at Perpetua is a canonical embodiment of customer sentiment toward automation in the last generation of adtech.

We learned from this experience. Knowing that Gigi would be our customers' earliest interaction with agentic AI, we put a concerted emphasis on explainability and control. To do this, we intentionally built the first version so that Gigi couldn't take any action without human direction. The workflow was textbook assisted AI: human assigned work, AI performed work, human reviewed and accepted.

We then asked our customers to rewrite their best practices for operating the Amazon DSP in a way unconstrained by human labor, knowing that Gigi is always on and always executing. There was an unanticipated effect: the human became a bottleneck. Our customers became paralyzed by being asked to author work they had long since graduated from, or were never doing in the first place. Something needed to change.

Two months ago, we launched autonomous tasks at the request of our customers. At the time, we had the following to say:

Our customers essentially promoted Gigi. She no longer needs approval to submit work and could move on to more challenging tasks. Our customers have built a sufficiently high amount of trust with Gigi to automate tasks. More importantly, though, these steps towards automation allow Gigi to provide more value to our customers because when AI is merely assistive, there are limits to its effectiveness.

Since then, we've been amazed by the rapid adoption of automated tasks within Gigi. In the last 7 days, over 75% of Gigi's actions have been automated. Interestingly, our customer profile is exactly the type of customer who would have "graduated from Perpetua": large agencies representing enterprise brands. So why have they adopted automation so quickly? With agentic AI, there are some key differences from legacy SaaS:

1. Entire workflows are being automated, not just rote actions. And those workflows aren't being prescriptively written by the software company, they're being flexibly crafted by customers. Our customers feel in control because they are the architects of Gigi's actions, not because we've opened a black box.

2. Explainability. Agentic products operate like team members. They provide clear explanations for the actions they take. Unlike legacy audit trails, which are just raw CSV exports, an agentic audit trail creates a feedback loop on why decisions were made.

It's clear that increased automation delivers a better customer experience by improving performance and eliminating the paralyzing need to review every AI action. Frankly, all low-level ad ops tasks should be automated. But the way our product was constructed wasn't conducive to increased automation. Historically, automation needed to be decided at the task level. We saw strong adoption for low-risk tasks like domain exclusions, but less for higher-impact tasks like bid and budget changes.

This week, we're introducing contextual safeguards for all of Gigi's tasks. With safeguards, customers can automate tasks while requesting human direction whenever a specific action hits a threshold, like line item minimums above "$x" or budget changes above a "% of change." Three things should follow:

1. The human bottleneck is eliminated. Fewer actions to review means our customers are only seeing what genuinely requires their judgment.

2. Perceived control increases. The human feels more in charge because they're only seeing actions they've said matter.

3. Advertising performance improves. With Gigi always on and always executing, customers and their clients benefit from near-real-time optimization fidelity at a level no human could sustain.

This marks a continued step on Gigi's journey from junior ad ops executor to autonomous strategist. The customers who win with Gigi won't be the ones who press the most buttons. They'll be the ones who define the best objectives, make the right calls when Gigi surfaces something that genuinely needs them, and trust her to handle everything else.

Note: This was originally published as part of our Cherry Picked newsletter. Subscribe here 🍒