When Anthropic needed to understand why developers were abandoning Claude Code during its research preview, Jane Justice Leibrock, the company's head of user experience research, ran the study on Listen. The finding was specific: people did not want to keep switching between their code editor and the terminal. Anthropic shipped a VS Code extension. Drop-off went down.1

Listen's case study frames this as speed: "100 studies for the price, at least in terms of time, of five or six," in Leibrock's words.2 The more interesting fact is the other one. A company that trains some of the world's most capable models trusted an outside AI research product enough to make a product decision based on it. That trust is what Listen gives, and speed is one of the mechanisms.

A note on timing. On September 9, TechCrunch reported that Listen walked away from a signed $125 million Series C term sheet to pursue acquisition talks with Salesforce at roughly $2 billion.3 The talks are ongoing, and nothing below depends on how they end. If anything, they sharpen the question the reading asks: what does a buyer understand Listen to be?

Two gifts

When Listen works, it gives a company two things. The first is standing: every insight traces back to the words a customer said; every AI summary can be checked against the quotes beneath it; every finding can be traced back to its source. What comes out is evidence a research-serious company can defend. The second is depth: every interview is a conversation, with follow-ups, and the words stay usable afterward.

Listen's own positioning is two-part: technology at scale plus research expertise, and the homepage, rebuilt since July, now has a section called "Research experts, on your team."4 Standing and depth are those two parts seen from the buyer's side. In July, the first was on the site as reassurance, and the second was on the site as speed.

Standing, ready to lead

The homepage now says "Every claim traces back to a real interview."5 It sits where reassurance belongs, next to the stats about respondents and languages. Standing is bigger than reassurance. It answers the one question that matters in the room where research meets money: how sure are we? It is the difference between "the AI said so" and "here is the customer, in her words, saying it." Click a claim in a report and see the moment in the interview it came from. Click a pattern in the archive and see the participants who form it.

That is a categorical break from the version of "AI-generated" that means "possibly hallucinated," and it is the version of AI research a company can ship from. Listen's own August essay on what AI owes researchers names three demands: real people, follow-up probing, verifiable work.6 The opportunity is to lead with standing, because it is why AI research is credible at all, and because a competitor without built-in traceability can't follow.

Depth, told as what it is

The case-study taglines carry the depth story in units of time. "100 Studies in the Time of 5." "10x Faster Results." "10x Faster Insight Discovery."7 All true. Depth is the thing underneath them.

Listen’s words “100 Studies in the Time of 5.” “10x Faster Results.”

Human-first approach “Our customers have the answers, and listening to them is really the key to unlocking the next phases of growth and innovation for the company.” (Jonathan Neman, Sweetgreen)

Depth is what happens inside the interview. Listen's moderator asks the question, hears the answer, and probes. It catches the say-do gap on a shared screen: you said you filter for price, but your screen shows you filtering for color. The homepage now has a whole section built on that gap, with a 42-point difference between what participants said they would pay for and what they chose.8 A survey tool is structured. A research platform schedules a human interviewer and tidies the transcript afterward. Listen runs a moderator that keeps the depth only a human used to reach, at a scale a human cannot.

The consequence is qualitative research at quantitative reach: a hundred conversations rather than a hundred survey responses. Sweetgreen's CEO, Jonathan Neman, put it in one sentence in Listen's own case study: "Our customers have the answers, and listening to them is really the key to unlocking the next phases of growth and innovation for the company."9 That is the depth story prioritized, with customers being heard with enough specificity that the next decision has somewhere to go.

Until July, that line was on the homepage; it has since been replaced with a quote from Neman about doing ten times more research for the same budget.10 The reversal speaks to the CFO, while the old one told the CEO what the company was buying. The swap turned depth into speed, in one edit.

What the two gifts become

Standing and depth compound. Without standing, an archive of AI-moderated interviews is a folder of summaries. Without depth, it is a database of survey responses. With both, it is institutional customer memory a company can build a decision on months and years later.

Listen has named this. The founders' letter for the Series B says: "Over time, Listen becomes a living archive of what your customers think and feel."11 The idea has begun to reach the product: the homepage's "Compound your learnings" pillar is the archive in feature language,12 and Pulse, the always-on conversational tracker launched in August, is the archive as a product, "a moving picture of what your customers think and why."13 The framing exists, and the product now proves it. What is left is to put the sentence at homepage weight, where the two gifts meet.

The window

The AI-research category is compressing, and this summer it took a specific turn. Simile raised $200 million at a $2 billion valuation on July 30, five months after its Series A, and defined the frontier as simulated persona modeling.14 The competitor set named in the Salesforce coverage (Simile, Outset, Keplar, Aaru) is mostly building in that direction.15 Listen sits at a different frontier: real people, deeply interviewed, whose words can be interrogated back to source and re-interviewed rather than extrapolated. Listen is hiring a founding research scientist in "human simulation,"16 so it will have a twin story of its own. The sentence that separates "grounded in real people" from "generated by a model" is worth more before that launch than after it.

Whatever Listen becomes, a Salesforce product line or an independent company at a new valuation, the story it will need most is the one the founders' letter already wrote. The homepage is ready for it.

About this series

Narrative Readings take an AI-native company at its word, on its own surfaces, and ask whether the story does the work the product needs. I call this framework Storytelling for Adoption: people only adopt what they understand, and companies only adopt what they trust. This summer I walked the diagnosis and the recommendations through with each company. The recommendations stay with them. What is published is the diagnosis, with sources in the footnotes, and what the company did next.

I'm Bruna Talarico. I built the content strategy function at Amazon Music's Global Content Organization and the go-to-market narrative for Audible's B2B unit. If you're building narrative for an AI-native product, brunatalarico.com is where to find me.

Companion readings

"I made that." (Lovable)

Earn the Search

Teach the Win

Lovable's careers page has the only sentence on its website that tells readers who she is, and it is aimed at recruits. "You are a driver. Not a passenger." Below it, "Candor over comfort."1 When I first read the customer-facing pages in May, none of them had borrowed the framing yet.

I read Lovable's surfaces the way I read any company's: what it says about itself, not what it sells. I covered the homepage, the founders page, the marketers page, the enterprise page, the security page, the careers page, and the April 22 response to a data-exposure incident.2 What follows is that reading, held up against what Lovable has built since.

One product, two gifts

When Lovable works, it gives a person two things. The first is agency: the power to make the thing yourself, today, without waiting on someone else's roadmap. Lovable turns "I have an idea" into "I made that." In November 2025 the company reported 100,000 new projects a day;3 by September 2026 the homepage counts 1.2 million a week and 60 million in total.4

The second gift is reliance: the confidence that what you built will hold, that it is yours, that it is safe, that you can put your actual business on it. Agency turns a weekend into a product. Reliance turns the product into a company.

In May, one of those was being told a great deal more than the other.

Gift one: the voice already exists

The marketers page reads today as it did in May.5 Its headline speaks to a budget line. The human version of the same promise speaks to the marketer with a campaign waiting in an engineering queue.

Lovable’s words “Ship faster. Save $50,000 by using Lovable for landing pages.”

Human-first approach “That campaign you’ve been waiting on engineering to build? Build it yourself. Today.”

Both are true, but only one changes how someone sees themselves. That is the one a new customer comes on board for.

The founders page reaches for the human version ("Lovable is your AI cofounder and development team"),6 and in May the enterprise and marketers pages led with the same two words: ship faster. A solo founder, a marketing lead, and a Fortune 500 buyer all met the identical headline. They are not the same people. The founder is deciding whether the idea has legs. The marketer is deciding whether to wait on a ticket. The buyer is deciding whether to sign. Each recognizes herself faster in a sentence about that decision.

Which is why the careers page matters. "You are a driver. Not a passenger" is five words, addressed to one person but applicable to many use cases. The move is to point that voice outward, at the people about to say, "I made that."

Gift two: the story a company can tell on purpose

In May, reliance lived on the security page and in the April response. The April post was clear, accountable, and written by a human; it showed the company could tell this story well.7 The next step was to tell it before anyone asked, so that "I made that" arrives with its second half already attached: and I can rely on what I made.

For a product holding millions of people's source code and credentials, that second half goes beyond a compliance detail and becomes the difference between a product and a company.

What Lovable built next

Between May and September, Lovable told the second half.

The homepage now says "Your code and data stay yours. Always." next to "Enterprise-grade infrastructure" and a line about automatic scans and audit logs, with a Security link in the top navigation.8 The business page, "for work," opens with "Safe and secure, as standard," names SOC 2 Type II and ISO 27001, and draws the line the reliance gift is about: "Most AI tools get you a prototype and stop."9

On July 22, Lovable announced it was the first coding-agent platform to earn AIUC-1, a third-party security, safety, and reliability standard for AI agents, and emphasized "independently verified, not self-attested."10 On August 5, it shipped a security page for every app published on the platform, at the app's own address, showing which controls are live.11 On August 12, it raised $400 million at a $13.3 billion valuation and listed security, reliability, permissions, and governance among the uses of the money.12

A company growing this fast, with employees at nearly two-thirds of the Fortune 500 using it,13 arrives at the reliance story on its own, because enterprise buyers ask for it in the room. The diagnosis held, and the company moved the reliance story onto the pages where a buyer decides.

What is left

"Ship faster" still greets the two audiences most likely to say "I made that" and least likely, so far, to hear it from the brand.1415 The reliance story now lives on the homepage and the business page. It has not yet reached the moments where a single maker decides: the pricing page, onboarding, and the first time real data gets connected.

The category is compressing. Independent comparisons this year describe the app-building step as commoditized across tools, with pricing models and integrations as the remaining differences.16 Every competitor can say "ship faster" now. What most of them cannot copy on the same timeline is a system-level story: one voice that names the change in a person's life, told the same way on every surface, backed by the sixty million things people have already made.

Lovable wrote the second half of the sentence in April because it had to, and again in July because it chose to. The place it hasn't written yet is where a maker reads it alone when deciding whether the thing she just built is a company.

About this series

Narrative Readings take an AI-native company at its word, on its own surfaces, and ask whether the story does the work the product needs. I call this framework Storytelling for Adoption: people only adopt what they understand, and companies only adopt what they trust. This summer I walked the diagnosis and the recommendations through with each company. The recommendations stay with them. What is published is the diagnosis, with sources in the footnotes, and what the company did next.

I'm Bruna Talarico. I built the content strategy function at Amazon Music's Global Content Organization and the go-to-market narrative for Audible's B2B unit. If you're building narrative for an AI-native product, brunatalarico.com is where to find me.

Companion readings

"We've been listening." (Listen Labs)

Earn the Search

Teach the Win

I had never been to Berkeley, and I was surprised to feel that it was way warmer than SF. WAY warmer. Also, and this is important, no wind. This is perfect, I thought, as I walked into the university campus alongside five thousand other people, fifteen hundred organizations represented, for Berkeley RDI‘s Agentic AI Summit last weekend.

In the two days it took place, there were more stages and talks than any one person could reasonably keep up with, but that felt like a good parallel to technology itself. We can’t cognitively process it, live, at this pace. Yet, by all accounts, panels and keynotes alike, this is the best time to be alive.

Most of the news on AI in late July covered price: Kimi (Moonshot AI)’s open-weight K3 reset the cost floor, and Anthropic answered within days with a flagship-class model at half its own price. On stage at Berkeley, however, almost nobody talked about price. Instead, they talked about how critical it is that the people building complex technologies can write down what good looks like, and in a precise-enough manner that a machine can be held to it. The name for it is evals, and the contrast with the month’s headlines was hard to miss.

The scarcest skill in AI right now is not building agents. Session after session, from different companies attacking different problems, the same claim kept resurfacing: the models are already more capable than what we manage to extract from them, and the work that closes the gap is not building. Instead, it is specifying, measuring, and judging.

Ryan Lopopolo, who just left OpenAI for Google Cloud, described teams where humans have not written implementation code in over a year; what remains of the job is decomposing the task, supplying context, and judging what comes back. Michele Catasta of Replit walked through a system that clusters millions of daily production traces to surface the rare failures no benchmark would catch, feeding them back into evaluation so the product improves continuously. The eval suite is an engine in itself. It is no longer a gate you pass once before shipping.

On the enterprise panel, HubSpot’s Duncan Lennox named writing good evals as the new, underestimated enterprise skill, and Mercor’s Adarsh Hiremath. described the failure loop of companies that skip it: guess the use case, speculate, deploy to a sandbox, improve anecdotally.

Andy Konwinski, a co-founder of Databricks and Perplexity, has started treating benchmarks as living software: think versioned, patched, saturated tasks deleted and harder ones added, because a measurement that stops evolving stops measuring. His contamination-free coding prize scored around 10% on problems where the saturated public benchmarks read 90%. The measurement was the news, and the measurement was humbling.

Every one of those functions already has a name. Newsrooms have enforced them for a century and called it editorial judgment. A style guide is a rubric. A kill criterion is a test case. A masthead’s standards are a versioned eval suite, patched every time the world produces a failure the old rules didn’t anticipate.

I have built choice-architecture engines behind business and product decisions more than once. They all started as written, versioned answers to the question that every Berkeley speaker kept circling back to: how do we know this is good, and would we notice if it stopped being good? I absolutely love how that question has quietly become the highest-leverage question in software. Cue the confetti.

I have written before that the output is a reflection of the thinking. Agentic systems make that literal. When implementation is nearly free, the thinking has nowhere left to live except in the specification and the standard. The eval is where an organization’s judgment either exists in writing or doesn’t exist at all. Dawn Song’s keynote made the stakes concrete from the security side: even evaluation infrastructure, she noted, is now part of the attack surface. The very thing you use to define good has become valuable enough to steal (!).

The market is currently bidding up the people who can build agents. But the scarcity that lasts is on the other side of the loop: people who can articulate a standard, encode it, and keep it honest as the world moves. The industry spent three years asking who can prompt, and boy, it looks like the next three will belong to the people who know how to grade.

Subscribe now

PS: I attended this summit online last year, and the recordings alone were worth enough to convert me into an in-person attendee this year. Every talk from both days is now up, stage by stage and free, on Berkeley RDI’s summit page. I hope you enjoy it as much as I did!

For thirty years, software came to market on one rule: ship first, defend later. Facebook shipped and was fined afterward. Uber shipped and was sued afterward. The regulatory bill always arrived after launch, priced in as a cost of doing business. This year, AI stopped following that rule.

The gate

By May, all five major US frontier labs - OpenAI, Anthropic, Google DeepMind, Microsoft, and XAI - had signed agreements with the Commerce Department’s Center for AI Standards and Innovation (CAISI) to submit new models for security review before release. The agreements are voluntary on paper. In practice, launching over a documented government safety finding is a risk no general counsel signs off on, so the review is now the gate every launch has to clear first. Drugs come to market this way. Aircraft come to market this way. Consumer software never has.

The framework in public

Six weeks later, the framework ran in public. Anthropic launched Fable 5 on June 9; three days later, the Commerce Department pulled it under export controls. The model was dark for nineteen days and came back on July 1 with a cyber-safety classifier that was not in the original release. Anthropic was not the one to set that timeline. OpenAI showed GPT-5.6 to roughly twenty government-approved organizations on June 26 and received clearance for general availability on July 9 — a preview to a vetted list is not a beta, it’s a briefing. And on July 14, the White House launched Gold Eagle, a clearinghouse coordinating how AI-discovered security vulnerabilities get handled across agencies. Gold Eagle is not a launch-review body, but the direction is hard to miss: what was ad hoc in June had letterhead by mid-July.

Each of these was covered as a separate news event. Read together, they are one story: the machinery agreed to in May is now effectively running.

The exit door

Two days after Gold Eagle, Moonshot AI released Kimi K3, a Chinese model that performs in the same class as GPT-5.6 and Fable 5, and released it open-weight. The spec sheet says 2.8 trillion parameters; what that means for a regular user is simpler: a model as capable as the ones behind the American gate, free to download. Open-weight means anyone can pull the model onto their own hardware and run it there, and export controls cannot recall a file that is already on someone’s laptop. The moment the US formalized a gate, a Chinese lab shipped a model the gate cannot hold. I don’t know whether that door stays open, and neither does anyone writing about it this week. But it should temper any conclusion that this regime is settled. The escape valve appeared the same month the pressure did.

Trust now has two audiences

For thirty years, tech built trust with exactly one audience: customers. Better, faster, more useful, and the market cleared your brand. That audience still matters. But there is a second one now, and its trust behaves differently. A customer’s trust can be rebuilt with a product update, a price cut, a good keynote. A government official’s cannot. Fable 5’s nineteen days offline were not a capability failure; by every benchmark, the model was state of the art. They were a trust failure with an audience that did not exist for consumer software a year ago.

That changes the work for anyone whose job is the product story. The launch date is provisional in a way it never was, now movable by a review body most of the team has never heard of. It used to move only for an engineering slip or a competitor’s counter-programming. And the press release now has readers it was never written for: policy staff, foreign regulators, enterprise legal teams deciding whether your model is a procurement risk. A product narrative that pretends the political layer is not there does not make the layer disappear. It just means someone else (a competitor, a regulator, a columnist) supplies the political story about you, and you get to respond to their version.

Custody

I was a journalist for more than a decade before I ever wrote a product narrative — VEJA, O Globo, g1, revista piauí. Reporters learn one thing early: when the government enters a story, the story changes owners. That is what happened to the product story this summer. The teams that learn to speak policy while their competitors are still writing feature announcements are the ones that

will keep custody.

Sources

CAISI agreements with the five major US labs, May 2026: Nextgov (https://www.nextgov.com/artificial-intelligence/2026/05/commerce-ai-center-will-evaluate-google-deepmind-microsoft-and-xai-models/413349/); Politico (https://www.politico.com/news/2026/05/05/microsoft-xai-google-caisi-safety-testing-00906529).

Fable 5 launch, export-control disruption, and July 1 redeployment: Bright Defense (https://www.brightdefense.com/news/claude-fable-5-restored-after-u-s-lifts-export-controls/).

GPT-5.6 limited preview to ~20 government-approved organizations, June 26: Nextgov (https://www.nextgov.com/artificial-intelligence/2026/06/openai-releases-new-gpt-56-model-select-partners/414474/); Axios (https://www.axios.com/2026/06/26/openai-gpt-sol-terra-luna-trump).

GPT-5.6 general-availability clearance, July 9: Reuters (https://www.reuters.com/technology/openai-gets-us-approval-broad-gpt-56-rollout-axios-reports-2026-07-08/); Engadget (https://www.engadget.com/2210308/openai-rolls-out-gpt5-6-july-9/).

Gold Eagle clearinghouse launch, July 14: Reuters (https://www.reuters.com/technology/us-launch-ai-cybersecurity-coordination-group-white-house-says-2026-07-14/).

Kimi K3 open-weights release, July 16: Model Diplomat (https://modeldiplomat.com/story/trumps-ai-gatekeeping-white-house-controls).

There were some very smart perspectives shared on one concept I have been elaborating on recently. It is what I'd call an epidemic of overconfidence - not in AI itself, but in our mastery of it.

The New York Times recently published a piece by Savannah Sobrevilla about "toxic confidence," the cultural shift from humble brags and impostor syndrome to unshakable self-assurance and self-appointed expertise. The under-qualified declare themselves authorities, and I think we are all seeing the line between conviction and performance getting thinner by the day.

This is showing up in the AI space in a way that isn't just cultural and has a tangible operational cost.

People confidently claim decades of expertise in a use case that has been broadly consumer-accessible for only over three years (OpenAI launched ChatGPT in late 2022). LinkedIn has become a place where, increasingly, you can't tell if there's a person thinking behind the post or if it's AI writing for AI to read. I am not against the use of AI to help refine the delivery of ideas, though, and this very post went through many iterations supported by Claude. The issue is not that. It is that everyone is a GenAI strategist now, and everyone has cracked the framework.

Research is starting to show that the perception that AI has made us more productive is not always accurate. A Harvard Business School study with Boston Consulting Group (BCG) found that for tasks outside AI's sweet spot, people using AI were actually 19% less likely to produce correct solutions. And a recent piece in the Harvard Business Review by researchers at University of California, Berkeley, Haas School of Business found that AI didn't reduce work: it intensified it, with employees taking on broader scope and longer hours without being asked. The efficiency created its own pressure cycle.

When you outsource the thinking, the prompting quality drops. When the prompting quality drops, the output quality drops. And then you spend more time iterating than you would have spent thinking and putting pen to paper. The output looks polished (think nice words, clean structure), but the thinking either isn't there or is barely there.

The output is not the writing. The output is a reflection of the thinking. And I feel like we can't always tell the difference anymore.

When that gets lost, and leaders start managing their teams based on what a model told them rather than what their experience tells them, we've unconsciously moved from leveraging a tool to being contained by one. Has anyone here ever had a leader walk into a room and say, "AI told me this is how we should do it, so this is what I want to see"? That moment, when a model's output overrides years of earned judgment, is the one we should be paying attention to.

The ones who can show that there is reasoning behind what they produce, that there is a question behind the answer, are the ones who will be best positioned as this technology matures.

A practical way to try and not be overwhelmed by the speed of feature launches and new tools to implement is to create a RACI framework for your goals and priorities. If you are the accountable person (i.e. the one who is ultimately answerable for the outcome and has final decision authority) on any given topic, that is where your time and mental bandwidth should go. Do use AI to leave no stone unturned. But try to narrow. Context switching is driving a real cost: research from UC Irvine shows it takes over 23 minutes to regain deep focus after an interruption, and high-frequency switching has been linked to burnout independent of workload. Though that is something to discuss at another time.

The humble brag may be dead, but honest thinking doesn't need one.