Steph Smith
Content Manager
Nobody’s short of output anymore. What’s scarce is someone with the context to say whether it’s any good. Here are three operators on the bottleneck that creates, and why your best work is being judged by people running a detector they can’t fully explain.
Andrew Davies is Chief Innovation Officer at Paddle, which handles payments and billing for software companies. They’ve spent the past year getting Paddle to use AI, and it worked. 91% of the company now actively uses an agentic platform wired into the CRM, the marketing automation stack, and the core data infrastructure. Ask it something in plain English and it queries. Hand it a workflow, and it runs.
Watching for AI-generated content has become an important daily task. It’s always a direct message, worded with care: “Hey, do you stand by this? This is where I think it’s missing context. What do you think?”
Paddle saw adoption faster than expected. But it wasn’t by accident. Monthly all-hands demos, an agent of the month, weekly Slack profiles of a clever automation, functional leads translating the potential value to their own teams, hackathons built around the tools.
The person prompting the content must also be the quality control department. Andrew explains, “We can create much more content than we can validate.”
Validation scales with attention
For 20 years, the constraint in go-to-market was production. You wanted more landing pages, more sequences, more decks, more variants for more segments. But you couldn’t have them, because making things required people, and people cost money and took time. Every process, agency retainer, design queue, and request form existed to ration a scarce resource.
That resource stopped being scarce sometime in the last 18 months, and yet most teams are still running the rationing systems.
What’s in short supply now is the other half of the job. Someone has to read the content and decide whether it’s true, on message, any good, and customer-ready. And that capacity hasn’t moved at all. It’s still one person reading at the speed a person reads, and still needing enough context about the market to know when something is better than just “good enough,” or wrong.
Teams need to be wary that output scales with compute, but validation scales with attention. Every model release widens the gap rather than closing it, which means this isn’t a transitional problem that gets solved by the next version.
"We can create much more content than we can validate."
Andrew Davies
Paddle
Andrew’s situation shows how this plays out. Paddle’s growth meant they onboarded a lot of new people into a fast-moving market many of them didn’t know yet, and asked them to deliver value quickly. Then they equipped them with powerful tools. Nobody made a bad call; the output just arrived faster than the context.
Agentic workflows only heighten this. When AI drafts a paragraph, you check it. Done. When an agent pulls from your CRM, cross-references call notes, drafts a follow-up deck and files it in a shared folder, what’s under review is a chain of decisions, most of which you never see.
None of this is an argument for using less of it. It’s just that the important work has moved to the front, before anything gets triggered or generated.
You’ve lost the benefit of the doubt
Stefan Bader runs Cello, which builds referral and partner infrastructure for B2B software companies. His go-to-market operation is about as forward-leaning as they come: AI-driven research and enrichment at a scale that punches above their headcount, multi-threaded communications across channels through a sequencing tool, and automated meeting prep. Everything lands in HubSpot, and Claude reads across Slack, email, and call recordings to give anyone on the team the full picture of an account in seconds.
He’s so committed to this that MCP support has become a purchasing veto. “If you can’t hook a tool up, you can’t offload tasks to AI. That information is just siloed.” It’s all about read and write access, he adds. He wants the AI enriching deals, not just describing them.
As much as he relies on AI, Stefan still expects humans to put care into making sure the final output is high-quality: “Why should I spend time and effort engaging with content that no one has put time and effort into creating?” When he sees a message sprinkled with em dashes, for instance, it erodes his trust in the business.
And it’s spreading beyond punctuation. Even when there’s no obvious tell, readers assume AI has touched everything. The default for communication used to be human, but now it’s the reverse. This puts a strange new job on the desk of anyone doing client-facing activities. “Human-authenticated signals are becoming more important,” says Stefan. “Your work has to be good, and it has to carry visible evidence that a person made it to be perceived as valuable.”
"Human-authenticated signals are becoming more important. Your work has to carry visible evidence that a person made it to be perceived as valuable."
Stefan Bader
Cello
Stefan sees this moment as historic: “It’s like the early days of the internet. There were no spam filters back then. The term spam didn’t even exist. You read every email you received because you barely got any. Then people started using it as a way to acquire customers, and spam evolved, followed by spam filters. I think we’re at the pre-spam-filter age of AI right now.”
The platforms are where this is playing out. LinkedIn already has a feature to report when a post “seems like AI slop,” and they’re granting higher messaging volumes to accounts verified with government ID. If you’re verified, you get more reach. Conversely, if you automate on top of the platform, your reach is penalized.
Push that process far enough, and it changes what a sales or marketing asset is. The scarce, valuable, high-converting signal becomes proof that a human was involved, and that almost imperceptible touch is what clicks with the reader.
Judgment is a muscle, and AI skips the reps
Nick Franklin founded ChartMogul, which does subscription analytics for SaaS companies. His engineering org is deep into AI: 30 engineers working in Claude Code, product managers and designers now shipping their own small UX fixes rather than queuing them, agents reviewing every pull request for code quality and security.
His go-to-market operation looks almost untouched by comparison. No AI-written customer emails. A human answers the phone. Clients get a shared Slack channel and a reply from a person. The illustrations on the website are still drawn by hand.
And this is by design. It’s a rule Nick applies consistently: automate where you aren’t differentiated, defend by hand where you are. Site illustration is the brand. Answering a customer who pays thousands a month is the product. ChartMogul’s high-touch customer approach relies on building authentic relationships.
“People can produce pretty amazing imagery,” says Nick. “But the vast majority don’t have strong visual taste. So you still need people with those skills to choose whether the AI-generated imagery represents your brand. And ironically, the people who can do that could produce the work themselves anyway.”
The ability to judge output and the ability to produce it turn out to be the same skill. AI delivers its biggest apparent gains exactly where judgment is weakest, and because its output is so sure of itself, the people least equipped to catch the errors get the fewest signals that there are any.
“You still have to be an expert in your domain,” Nick says. “Otherwise AI will give you a lot of confident nonsense.”
There’s an uncomfortable implication here. Judgment gets built by doing, often badly for a few years until you develop an eye for what good looks like. If AI removes those reps, no amount of additional tooling will fix the validation shortage that follows.
“You still have to be an expert in your domain. Otherwise AI will give you a lot of confident nonsense.”
Nick Franklin
ChartMogul
Designing humans into agentic workflows
ChartMogul ships dozens of pull requests a day, and the help center inevitably falls behind. So they set up an agent that reads what went live, compares it against the existing documentation, identifies what’s wrong, and drafts the correction. Then it stops and asks a person to approve.
The agent does the reviewing, which is the part that scales. The human does the deciding, which is the part that doesn’t. That’s the design principle worth taking into every agentic workflow you build this year. Not “should a human review this,” which always gets answered yes and then forgotten when things get busy. Instead, ask “where in the chain is the human, what specifically are they deciding, and does the workflow physically stop until they do?”
Andrew sees the same thing playing out at a larger scale. He describes three stages of AI adoption: individuals automating parts of their own job, teams building workflows around what automation makes possible, and eventually entire cross-functional processes getting rebuilt around it. “This next stage won’t be about making an individual or a team more efficient,” he says. “It will be rewiring where our human talent is placed versus where nonhuman, agentic workflows are taking place.”
At Paddle, he puts his people where the learning is, which is why he spends so much of his time in rooms with customers and partners rather than in the systems he built. This is also his answer to where companies are missing the mark right now: “Teams often waste resources trying to keep up with everybody else’s learnings, rather than going and getting their own.”
Every team reading this will automate more this year, connecting tools into workflows that change what’s possible. The ones that come out ahead will be discerning about which parts they don’t.
This leaves one question to ask about everything your team ships: Do you stand by it?
