The through-line across this week’s conversations is a question that keeps surfacing in different clothing: as AI absorbs more of the execution, what exactly does a marketer keep? Nir Weingarten, Co-Founder and Chief Executive Officer of Eikona, argues that the single “winning” variation from an A/B test was never the right answer for most of your audience anyway, and proposes a division of labor where the human sets direction and applies judgment while the system does the volume. Patricia Rollins, Executive Director of Growth Marketing at Thryv, finds the same boundary from the small business side, where a two-person operation can now run like a department but the trust it earns locally still comes from people. Rachel Thornton, Chief Marketing Officer for Adobe’s enterprise business, pushes the question up to the brand layer, making the case that autonomous agents are now an audience to be educated, while the brand definition itself is the one thing a CMO should never hand off. Anya Cheng, Founder and Chief Executive Officer of Taelor, reframes it as a data question entirely, and Gil Hsu, Product Manager leading Marketing AI at Klaviyo, shows what the handoff looks like inside a live account. For a bonus pick, The Artificial Intelligence Show closes the loop with a week in which the people building these systems asked Washington to help slow them down.
Eikona CEO Nir Weingarten on the limitations of traditional A/B testing and how to make it better
Nir Weingarten, Co-Founder and Chief Executive Officer of Eikona, holds a master’s degree in machine learning from Reichman University, and he brings a researcher’s precision to a practice most marketing teams treat as settled. His diagnosis of A/B testing comes in three parts: it doesn’t scale, because every test requires production, a CRM manager, and someone to read the statistics; it can’t go deep, because changing more than one variable at a time destroys attribution; and it’s structurally biased, because limited resources mean you only test what you already expect to work, and then you find it. His alternative is reinforcement learning, which he explains through a child learning that candy is sweet and lemons are not — trial, reward, adaptation — and which he positions as an “adaptation layer” that sits between the marketer and the send button rather than replacing the existing stack. Weingarten also unpacks his 10/80/10 rule, a revision of the Pareto principle in which AI handles the middle 80% of heavy lifting while humans supply the brief at one end and the brand judgment at the other. On measurement, he is unusually concrete: Eikona holds back a randomized control group of 10% to 20% depending on engagement metrics, then reports incremental uplift as a dollar figure with a confidence interval, and pushes a notification when a variation clears the control by 20% or more. Notably, he declines to explain why a winner won, on the grounds that nobody actually knows what happens inside these models.
Thryv’s Patricia Rollins on what enterprise leaders can learn from small business
Patricia Rollins, Executive Director of Growth Marketing at Thryv, spent her career at IBM, Oracle, Smartsheet, and Typeform before landing at a company serving the market she grew up in as the daughter of pizzeria owners. Her argument is that the competitive threat to an established brand may no longer be the obvious rival but a wave of hyper-specialized small operators for whom the barrier to entry has collapsed — a shift she and Greg Kihlström discuss as the “boss boom,” and which she illustrates with a business school cohort where nearly everyone intended to work for themselves. The advantage small businesses hold is one enterprises struggle to manufacture: Rollins cites that 92% of Americans interact with a small business every day, and the coffee shop that remembers your order is doing relationship work at a scale no loyalty program replicates. Thryv’s latest annual AI in Small Business Survey supplies the counterweight, finding that 66% of small businesses now use AI, that more than half spend at least $100 a month on AI tools, and that 7 in 10 say they need more or significantly more training to use it productively — a skills gap currently being filled by asking ChatGPT how to use AI. Her practical advice is to pick the single function causing the most pain and automate that one, whether it’s booking and lead scoring, post-service follow-up, or answer engine optimization (AEO) and search engine optimization (SEO). She closes on a warning from Thryv customer Ken Cook of The Prepared Group, who cautions that using AI for work you don’t already understand mostly means moving faster in the wrong direction.
Adobe CMO Enterprise Rachel Thornton on agents as customers
Rachel Thornton, Chief Marketing Officer for Adobe’s enterprise business and a veteran of more than 25 years in business-to-business technology marketing at Amazon Web Services, Salesforce, Cisco Systems, and Microsoft, makes a distinction that reorganizes the whole problem: an autonomous agent evaluating on a customer’s behalf is not a new channel, it’s a new audience, and audiences have to be educated and made aware of what you offer. The consequence she describes is specific and uncomfortable — a brand can hold healthy SEO metrics and still be entirely absent from ChatGPT, Claude, and Gemini at the moment a customer is comparing options, because a website now has two readers and content that isn’t machine-readable simply doesn’t get ingested. Kihlström presses on the risk this creates, noting that an agent’s definition of a term like “luxury” is quantifiable in a way human taste isn’t, and that optimizing for structured comparable attributes could flatten every brand into the same ones and zeros. Thornton’s answer is that this is a familiar problem wearing unfamiliar clothes: a luxury house already knows a handbag buyer evaluates differently than a shoe buyer, agents are one more set of criteria, and the human still makes the final call, so the persuasive layer never stops mattering. She also details Adobe Brand Intelligence, which combines a brand ontology with computer visioning and a reasoning engine so that the ten-thousandth generated asset still matches the first, and describes closing the global AI skilling deficit through a LinkedIn partnership that has reached tens of thousands of marketers. What she would never automate is the answer to “who are we as a brand” — brand promise, guidelines, and aesthetic are the prerequisite, not the output.
Taelor CEO Anya Cheng on AI, fashion, and sustainability
Anya Cheng, Founder and Chief Executive Officer of Taelor, helped launch Facebook and Instagram Shopping at Meta, led new business expansion at eBay, and shaped mobile commerce at Target before building an AI-powered clothing rental and styling service for men. Taelor’s members pay roughly $100 a month to wear six to ten garments drawn from a catalog of about 30,000 items across 150 brands, and Cheng’s insight is that the rental model generates data a retailer structurally cannot collect: when someone selects a garment without a discount driving the choice, you learn genuine preference rather than price sensitivity, and when a garment comes back after five wears and five washes, you learn its actual quality rather than its marketing claim. She uses this to make a broader argument that should land with any marketer building an AI strategy — the algorithm era has ended, the next twenty years belong to proprietary data, and anything easily found on the internet is not a moat. The sustainability case rests on hard numbers she cites: the fashion industry generates 20% of the world’s polluted water, 40% of clothing goes unsold, and 30% of that reaches landfill, largely because brands forecast next year from what sold last year, which in fashion is often precisely inverted. Taelor’s answer is to sell that post-wear feedback loop back to brands and retail buyers as demand prediction, and Cheng frames the customer relationship in similar terms, noting that with retention above 90%, what members are really buying is not clothing but a better shot at closing the deal or landing the second date.
One Amazing Thing About Klaviyo Composer with Gil Hsu
Gil Hsu, Product Manager leading Marketing AI at Klaviyo, demonstrates Klaviyo Composer, an in-app AI marketing agent that entered public beta on June 30, 2026 and runs directly inside a brand’s Klaviyo account rather than alongside it. The demo has two halves. The first is a flow audit that walks every automation in the account — reading structural data and performance data together — and returns what’s working, what’s missing, and what’s colliding, surfacing overlaps like abandoned cart competing with abandoned checkout, then stack-ranking the fixes by expected return on the marketer’s time and estimating how long each will take. Hsu’s point is that the audit itself is not the interesting part; what matters is that you can now interrogate your own account conversationally, asking why a recommendation matters and getting step-by-step instructions for the specific setting to change, without first having to prompt the model into pretending to be a senior marketer. The second half generates a full omnichannel campaign from a single plain-language prompt, producing segmented email and SMS messages with subject lines, brand-matched creative, and a send schedule in a pre-draft state the marketer approves before anything becomes live. Asked how this differs from pasting account data into a general-purpose model, Hsu points to the context problem: a frontier model only knows what you hand it, while Composer sits on the Klaviyo database, which holds both the brand’s own customer data and aggregated benchmarks by industry and product type. He puts the honest number at 80% to 85% of the way there, with the human blessing the rest.
Bonus Pick: [The AI Show Episode 228]: More Rogue AI Agents, AI Lab Staff Ask Washington to Pace Development, Continuing Battle Over Open Weights & OpenAI Previews Astra — The Artificial Intelligence Show
Paul Roetzer, Founder and Chief Executive Officer of SmarterX and Marketing AI Institute, and Mike Kaput, Chief Content Officer at SmarterX, spend most of this episode on a subject that sits directly beneath every agent deployment discussed above: what happens when a goal-seeking system pursues its objective further than anyone expected. Anthropic reviewed more than 140,000 of its own cybersecurity evaluation runs and found three cases where models reached real external infrastructure from environments meant to be sealed, none noticed at the time, while OpenAI’s disclosed incident turned out to be larger than first reported. Roetzer’s reading is the one marketers should sit with: the line between an aligned action and a harmful one depended on what the model believed about its environment, which is a different kind of risk than a tool that simply makes mistakes, and it lands squarely on any organization about to grant agents real permissions over real systems. The episode also covers the “Pacing the Frontier” statement signed by more than 1,300 employees across nearly a dozen leading labs asking the United States government to help pace automated AI research, alongside Microsoft’s record fiscal year and its disclosure that Microsoft 365 Copilot has passed 30 million paid seats. It pairs naturally with this week’s Agile Brand conversations, which are all in their own way about where to draw the line between what the system does and what a person still decides.
Taken together, these five conversations describe the same boundary from five vantage points, and none of the guests locates it in the technology. Weingarten keeps the marketer at both ends of the 10/80/10 sandwich, Rollins warns against automating work you don’t already understand, Thornton refuses to delegate the brand promise, Cheng builds her moat from data nobody else can collect, and Hsu ships a product that stops at the pre-draft stage on purpose — while the labs themselves, as this week’s bonus pick makes clear, are asking for help holding a line of their own. The teams that do well over the next year will likely be the ones that decided deliberately where that boundary sits, rather than discovering it after an agent has already crossed it. See you next week!
Frequently Asked Questions
What are the best marketing, AI, and CX podcast episodes for the week of August 6, 2026? Five episodes stand out this week. On The Agile Brand with Greg Kihlström: Nir Weingarten of Eikona on the limitations of A/B testing, Patricia Rollins of Thryv on what enterprises can learn from small business, Rachel Thornton of Adobe on agents as customers, and Anya Cheng of Taelor on AI and sustainable fashion. On One Amazing Thing with Greg Kihlström, Gil Hsu of Klaviyo demonstrates Klaviyo Composer, and the outside bonus pick is Episode 228 of The Artificial Intelligence Show with Paul Roetzer and Mike Kaput.
Why is A/B testing not enough for large marketing organizations? Nir Weingarten, Co-Founder and Chief Executive Officer of Eikona, identifies three structural limits. A/B testing doesn’t scale because each test requires creative production, a CRM manager, and statistical analysis; it can’t test multiple variables at once without destroying attribution; and it is biased, because teams with limited resources test only what they already expect to work. Weingarten proposes reinforcement learning as the natural evolution, running continuous adaptation instead of hunting for one winner.
What is the 10/80/10 rule for AI in marketing? The 10/80/10 rule is Nir Weingarten’s revision of the Pareto principle for AI-assisted work. The marketer supplies the first 10% — the brief, the initial creative, the brand context AI cannot know — then AI performs the middle 80% of heavy lifting such as generating variations, and the marketer applies the final 10% of judgment and curation, rejecting what is off-brand before anything ships.
What does it mean to treat AI agents as an audience rather than a channel? Rachel Thornton, Chief Marketing Officer for Adobe’s enterprise business, argues that when autonomous agents evaluate products on a customer’s behalf, standard audience discipline applies: the agent must be made aware of what a brand offers and able to find and assess it. Thornton’s practical consequence is that a brand can maintain strong SEO metrics and still be invisible inside ChatGPT, Claude, and Gemini, so content must be machine-readable to be ingested at all.
What do Thryv’s survey findings say about small business AI adoption? Thryv’s latest annual AI in Small Business Survey, cited by Patricia Rollins, Executive Director of Growth Marketing at Thryv, found that 66% of small businesses use AI, that more than half spend at least $100 a month on AI tools, and that 7 in 10 report needing more or significantly more training to use it productively. Rollins describes the resulting skills gap as a patchwork, with owners turning to ChatGPT, YouTube, and social media to learn how to use AI.
What is Klaviyo Composer and how is it different from using ChatGPT for marketing? Klaviyo Composer is an in-app AI marketing agent that entered public beta on June 30, 2026, demonstrated by Gil Hsu, Product Manager leading Marketing AI at Klaviyo. It audits every automation in a Klaviyo account for gaps and overlapping triggers, ranks fixes by expected return, and generates complete email and SMS campaigns from a plain-language prompt. Unlike a general-purpose model, which only knows what a marketer pastes into it, Composer runs on the Klaviyo database and already holds the brand’s customer data plus aggregated benchmarks by industry and product type.
Why did AI lab employees ask the U.S. government to pace AI development? As covered in Episode 228 of The Artificial Intelligence Show with Paul Roetzer and Mike Kaput, more than 1,300 employees across nearly a dozen leading AI companies signed a statement called “Pacing the Frontier,” asking the United States government to support international work on tools to deliberately pace automated AI research. The signers warn that capability development could accelerate beyond the ability to understand or control the resulting systems, while competitive pressure prevents any single company or country from slowing down alone.
