Two VPs of Growth. Two Unicorns. One Conversation.
No frameworks for beginners. No safe takes. No "it depends" without the answer that follows it.
There is a version of a GTM interview where a journalist asks a practitioner to explain their job to an audience that has never done it.
This is not that.
George Bonaci is VP of Growth at Ramp. He thinks in systems, moves in first principles, and has a specific kind of clarity about where GTM is going that you only develop by actually being in it at scale.
I am Josh Grant. Most recently VP of Growth at Webflow. Before that, Affirm. I have spent the last year obsessing over what happens to growth and demand generation when AI collapses the buying journey and the funnel stops being measurable in the ways we were trained to measure it.
We are both operators. We have both built growth engines inside companies with billions of dollars of pressure on them. We have both had to make bets with incomplete data, defend positions in QBRs with imperfect attribution, and figure out what modern GTM actually means in practice rather than in a LinkedIn carousel.
This interview is what happens when two people who have actually done the job stop being polite about it.
No frameworks for beginners. No safe takes. No “it depends” without the answer that follows it.
Just two growth leaders at the frontier, trying to figure out what comes next.
In a recent post I wrote about what GTM operators are unlearning in 2026, you dropped this: “The real skill is knowing which nodes on your org chart should be humans and which should be agents.” That line stopped a lot of people mid-scroll. So let’s start there. How are you actually making that call at Ramp right now, and what does the decision framework look like in practice?
Most people hear “agents” and think “faster employees.” That’s the boring version. The interesting question isn’t “what can AI do?” It’s “what should a human never have to do again?”
When I look at any role across the 8 growth teams, I ask one question: if this role disappeared tomorrow, would anyone notice the *decisions* were missing, or just the *output*? If the answer is output, that’s an agent. If the answer is decisions, that’s a human…for now at least.
I apply the same logic to the growth org chart. We’re literally hiring for a role we call the “Agentic Operator” - and the name is intentional. It’s not “AI Marketing Manager.” It’s someone who thinks about which workflows should be autonomous systems and which require human judgment.
And when thinking through the details I have a more in depth framework. The framework has three questions. First: is the decision computable or interpretable? If you can express the decision rule in a way that’s predictably measurable - bid optimization, send-time optimization, audience expansion based on historical lookalikes - that’s an agent node. If the decision requires pattern recognition that depends on taste and intuition, that’s a human node.
Second: what’s the cost of a wrong answer? If a wrong answer wastes a small test budget, fine. If it poisons a relationship with a CFO evaluating a $2M spend management contract, that’s a human that should be on the hook.
Third - and this one most people miss: is the task load-bearing for someone’s growth as a professional? I won’t completely automate a task that is the primary way someone on my team develops judgment. Otherwise you are trading efficiency for an organizational lobotomy. You’re saving the company $80K while destroying a future Director of Growth.
Having said all that, here is what might be uncomfortable: the ratio of output to decisions is roughly 80/20. Eighty percent of what knowledge workers do is production, not decision-making. The actual decision surface in most marketing roles is shockingly thin. Most companies have just built organizations that disguise production as strategy because headcount is how old fashioned leaders measure their importance.
So at Ramp, we are deliberately working to shrink the team’s production load to near-zero and concentrating humans entirely on taste and novel strategy. Taste meaning: does this feel right for our buyer? Would a CFO trust this? Novel strategy meaning: what bet should we make that no playbook covers?
The punchline is we’re not replacing headcount with agents. We’re replacing task fragmentation with agents so the humans who remain are doing the work that makes them more dangerous next quarter.
Ramp operates in a category where the buying decision is increasingly happening before anyone talks to sales. Buyers are researching, comparing, and shortlisting inside AI before they ever fill out a form. How are you thinking about owning that answer layer, and is it showing up in your pipeline data yet?
Here’s what’s actually happening: the buyer’s journey is inverting. It used to be awareness, consideration, decision, action. Now a CFO opens ChatGPT or Claude and says “what’s the best AP automation tool for a 500-person company” and gets 2 or 3 options in eight seconds. The consideration phase moved before awareness. You’re being evaluated before you even know you’re in the room.
We treat this as an AEO problem, not anything else. We went from sporadic AI citations to dominating answers for queries we care about. How? We reverse-engineered what makes an AI cite you. It’s not keyword density. It’s not backlinks. It’s structured proof. Named customers. Specific problems. Quantified outcomes. Implementation mechanics. When a finance leader asks an AI “who should I trust to manage $50M in spend,” the AI is looking for receipts - same as that leader would.
The thing most people aren’t seeing: this creates a winner-take-most dynamic more extreme than Google’s page one. In traditional search, position two still gets clicks. In AI answers, there’s often only one recommendation. The citation layer is a power law, and we’re in the land-grab phase right now. The alpha is now.
Is it showing up in pipeline? Yes. We’ve been seeing a growing cohort of inbound where the first touchpoint is effectively invisible in our attribution system - no ad click, no webinar, no content download. They just show up ready to buy. When we talk to them, a meaningful percentage say some version of “I asked ChatGPT.” That’s the dark funnel becoming the primary funnel.
Most growth leaders are still running toward traffic, MQLs, and pipeline volume. What is the metric you are actually optimizing for in 2026 that most of your peers are not measuring yet?
Revenue per cognitive hour.
Not revenue per headcount. Not revenue per marketing dollar. Revenue per hour of genuine human *thinking* my team invests.
Here’s why this matters. When agents start handling 60, 70, 80 percent of production, your cost structure collapses in a way that breaks every existing ratio. CAC looks absurdly low but the denominator changed, not the numerator. Pipeline volume spikes but the marginal cost of a bad lead hits near-zero, so quality signals get noisier.
Revenue per cognitive hour cuts through that. It forces you to separate thinking from execution. And when you do that honestly, you realize most of your very smart team’s week isn’t thinking. It’s execution.
When you compress all the production work into agents and look at what’s left, you find out how much strategic capacity your team actually has. At Ramp, that number surprised me. We had more raw strategic horsepower on the team than I thought. It was just buried under busywork. Freeing it up requires a new agentic first architecture.
The implication is stark. If you measured this honestly at most companies, you’d find that doubling your team’s impact doesn’t require doubling headcount. It requires cutting their non-thinking work by 80% and letting them actually think.
We both came up through growth at unicorns during a specific era. Blitzscaling, paid acquisition, funnel optimization, CAC math. What is the one thing you learned in that era that you have had to actively unlearn, and what replaced it?
I spent years worshipping funnel conversion rates. Top of funnel, middle of funnel, bottom of funnel. Every QBR was a waterfall chart. Every optimization was about finding the leak and patching it. And it worked - in a world where the buyer journey was linear and measurable.
What I had to unlearn is the core assumption underneath all of that: that the buyer’s journey is a sequential process you can instrument and optimize stage by stage. That assumption is just wrong now. It makes you optimize for local maxima at each stage while missing the system-level dynamics that actually drive growth.
What replaced the funnel for me is something I think of as a field model of demand. Instead of a funnel with stages, I think about a field with gradients. Buyers exist in a field of awareness, trust, and intent that shifts constantly based on signals from everywhere - AI recommendations, peer conversations, product experience, brand perception, content, news. Your job isn’t to move people through stages. It’s to increase the overall field strength so that when someone hits a trigger moment — their current vendor screws up, their company hits a growth threshold, a new budget cycle starts — the gradient naturally pulls them toward you.
Now, I know what you’re thinking: “If I am at a traditional company, how do you explain field theory to a board that wants a waterfall chart?” but that’s the reality of the new world we are marketing in.
There are two versions of AI adoption in GTM. One is workflow automation: move faster, do more with less. The other is structural: it changes what the team looks like, what the success metrics are, how decisions get made. Which version is Ramp actually living, and what did it take to get there?
I told my team something a few months ago that was unusual to hear. I said: I actually don’t care if we miss our numbers in the short term if it sets us up for long-term success. The goal should be that we don’t even want to hire anyone going forward because it would slow us down compared to investing in agentic capabilities.
And that wasn’t just a motivational speech but rather a resource allocation statement. And it’s the difference between workflow and structural adoption.
Workflow automation is using AI to do the same work faster. Your writer drafts with Claude. Your analyst summarizes with a model. The org chart stays the same. The job descriptions stay the same. You get more throughput.
Structural adoption is when AI changes *what work exists*. Not faster copywriting, but content that generates itself at the point of need with no writer in the loop. Not faster analysis, but continuous autonomous monitoring that surfaces decisions you didn’t know to ask about.
Getting to structural adoption required one thing above all: permission to sacrifice short-term output for long-term capability. Most growth teams can’t do this because they’re measured quarterly and optimizing for agents is a multi-quarter bet. That’s why the org chart question from your first question matters so much. If your team is set up to produce, they’ll resist the shift to building systems. You have to change what the team *is* before you can change what it *does*.
You are sitting at the intersection of fintech and AI at a moment when both are moving fast. Where do you think most growth leaders are still underestimating the shift, and what are they going to wake up to in the next 12 months that they are not seeing yet?
The org chart is the bottleneck. Not the tech.
I wrote this on LinkedIn but most growth teams are still organized like it’s 2024. Writer writes. Designer designs. Ops person wires it together. PM manages the backlog. But the job of a growth team was never to launch ads, design A/B tests, or write emails. Those are tasks. The job is solving the problems that prevent the company from growing faster. Most teams spend 90% of their time on tasks and 10% on problem-solving.
AI can flip that ratio. But only if you let it. And most orgs won’t, because the tasks are what justify the headcount. This is the part no one says out loud: the biggest obstacle to AI adoption in marketing isn’t the technology. It’s the incentive structure. Leaders who’ve spent careers building empires of 30, 50, 100 people are not going to voluntarily shrink those empires, even when shrinking them would make the team more effective.
Two other things people are sleeping on:
The buying entry point is shifting from search engines to thinking engines. You can’t buy a PPC ad inside ChatGPT’s response (well not really at least). You can’t retarget someone on Claude. The entire concept of a paid channel that you can buy inventory in starts to dissolve when the primary research interface is conversational AI.
And marketing in general is splitting in two. Content for humans (stories, emotion, brand trust) and content for machines (structured data, citation-worthy facts, verifiable claims). Different strategy, different team, different measurement. Most companies are still focused on one and hoping it works for both. It won’t.
If you were building a growth function from scratch today, zero legacy stack, zero inherited playbook, first principles only, what does day one look like?
Day one, I don’t hire anyone. Genuinely how I’d start, not a provocation.
Day one is me, a terminal, and three questions. Who is the buyer, and what are they asking their AI assistant right now? Not in a personas-and-journey-maps way. In a literal, “I’m going to map the top 100 prompts a buyer in my category types into Claude and see what comes back” way.
Second: what proof do I have that this product delivers? Not messaging. Not positioning. Proof. Numbers. Customer outcomes. Verifiable claims. I’d build a proof library before I touched a content calendar.
Third: what are the five to ten decisions per week that, if made well, actually move this business? And which of those can an agent make autonomously?
Day two, I build three agents before I make a single hire. An answer engine monitor that tracks my brand’s presence across LLMs daily, identifies opportunities, and kicks off work to capture them. A competitive intel agent watching pricing pages, feature launches, messaging changes, finds our angle, and implements it. A content production agent that takes my proof library and generates content, pages, and whole inbound flows, optimized for both humans and agents.
Day three, first hire. One person. Not a traditional marketer. A systems thinker who can code. Someone whose instinct when they have an idea is to build it, not brief it. This person’s job is to build the system that does marketing.
The traditional growth playbook says hire a team, build a funnel, buy some ads, optimize the conversion rate. That playbook assumes humans are the unit of production. They’re not anymore. Agents are. Humans are the unit of strategy.
The playbook most growth leaders learned was built for a world where humans were the unit of production.
That world is gone.
What replaces it isn’t a new playbook. It’s a new question: not “how do I build a bigger team?” but “what decisions does my team actually need to make?”
George has already answered it for Ramp.
The ones who win the next three years will be the ones who stop defending the old answer and start building around the new one, right now.