It was a Wednesday when a design colleague of Jason Zhou—who was leading Product at Relevance AI at the time—came to him after realising something strange: he hadn’t opened Figma in three days. Not because he wasn’t working; he’d been shipping features all week. He’d just been doing it differently. A quick sketch on paper, a screenshot fed into Cursor, and voilà—the feature was in the codebase and the company’s design system was intact. No mockups, no handoffs, no waiting for an engineer to implement it at 90% fidelity. 

At face value, it’s a pretty unremarkable moment, but it's the perfect example of how roles are morphing and evolving in the AI era.

To understand what’s happening, Jackie Vullinghs, Partner at Airtree, sat down with three operators to discuss this emerging pattern of designers shipping code, sales reps building agents and marketers creating entire workflows. 

The great role collapse

The story of how Jason’s colleague came to be shipping code didn’t happen by accident. Jason had locked the design team in a room for a day as an experiment to train them up as frontend engineers. He got them to create GitHub and Cursor accounts, taught them how to get a local server up and running and started to see their workflows change significantly. 

“Engineers never ship 100% of what designers want; there’s always the last 10% quality gap,” says Jason. “I started to see my designers jump into the codebase and fix all those things the frontend engineers never want to do.”

Designers aren’t the only ones leveraging code to change how they work. Over at Tracksuit, a brand tracking platform, the sales team is building and iterating on agents to replicate parts of their go-to-market (GTM) motion. 

“About a year and a half ago, our Head of Sales was talking about how they’d crafted a pre-call research agent for account executives,” says Saahil Bijlani, who sits in Tracksuit’s GTM Ops team. “It had 160 different nodes in n8n to get to the quality they wanted. Now, we’ve rebuilt that entire process and it’s 3 nodes and a number of different MCP servers.”

AI’s ability to augment roles comes with a caveat: it has to meet a high enough quality bar, which, until recently, it didn’t. 

The quality inflection point

When Saahil rolled out a pre-call brand research agent to his sales team, it was a big learning curve: in an age where you can produce almost anything, quality becomes the most important variable. The agent was 70% of the way there, but the sales reps started poking holes in the remaining 30%: it either pulled data that was too old or that didn't meet the defined constraints.

“It’s the little things that can completely diminish a sales rep’s trust in an agent,” says Saahil. “Just because you can spit out an agent quickly and ship it to the entire team doesn’t mean you should do that straight away.”

Quality is obviously a consideration for designers because they love things to be pixel-perfect. Jason, now the founder of Superdesign, had been trying to build a design agent to empower non-technical people to deliver design output as good as a designer's, but was struggling to reach the quality bar a good designer would expect. That all changed with the launch of Gemini 3 Pro.

“That’s the part that’s really exciting; every few weeks there’s a new model that changes the paradigm,” says Jason.

If you can build things quickly and the output is high quality, the next question is: why are you still buying software?

Build vs buy in the AI era

For many startups, there’s a legitimate question of whether they should bother buying software tools anymore when they can just build it themselves. 

“I spend most of my time thinking about this,” says Saahil. “I think the more low-value, low-lift things will have to be built because often you get a product that fits 90% of what you want, but there’s always that 10% where you think, ‘I wish it did this.’”

For Saahil, there are a few key considerations. When you total up the cost of software, is it easier to hire one person to build it and how much time would it take? If you build it internally, who will maintain it in the long term? Does building instead of buying shift focus away from what your core product and core USP?

Over at Chronicle, an AI-powered presentation tool, Renee Zhang sees a future where they’ll build tooling for internal workflows and ops.

“I can’t see us building a headless CMS for our website and blog, because that’s a lot to build,” says Renee. I think it depends on the scope of the tool; if it’s extremely personalised, it makes more sense to build internally because you can be in control of that workflow.” 

In practice, the build vs buy decision spills over into organisational design, forcing founders to rethink team size, roles, and ownership.

The team structure revolution

Jason has a specific number in mind for his company size: 10 people.

“Ideally, we have a very small team,” says Jason. “I care more about revenue per headcount. Each individual must be AI native and automate a lot of work.”

As a team of one at Chronicle, AI enables Renee to own multiple workflows—brand, content, socials, and SEO—with no plans to grow headcount anytime soon.

“The past year has changed so drastically when it comes to tools available and what people can now do themselves vs have a team do,” says Renee. “You can very much be a solo builder now.”

As we see teams doing more with less, it’s important to determine who is responsible for integrating new technologies across the board.  

“If you want a sales team to go build an agent themselves, what you’re asking people to do is their full-time job and learn a new language at the same time. If everyone’s stretched already, it’s never going to work,” warns Saahil. 

At Tracksuit, they have a 5-person centralised AI ops team serving 90 GTM employees, which Saahil says is a small team relative to the impact they have.

“You don’t want to scale through headcount, you want to scale through efficiency,” adds Saahil. 

Jason predicts that more roles will start blurring as each individual can do so much more than before. 

“If you have a growth team, why do you still have the traditional setup of one product manager, one designer, and two engineers?” asks Jason. “You can have people like Renee who have an idea and push it to production code. For early-stage startups, the decision-making quality and speed are almost always better because one person holds all the context.”

Lean teams shipping high-quality work quickly; it all sounds good, but what should founders and operators be mindful of before going all in on reimagining their workflows using AI?

Where AI meets its limits 

For any founders getting started, Jason points out AI does have its gaps: “When you’re at that very early stage, the muddy ideation phase, I don’t see it as a good solution.”

AI is great at executing clear specs, but not so great when things are messy and undefined. It’s a reminder that the most valuable work happens before specs exist. 

From Renee’s point of view, it’s no longer building that’s the constraint; it’s attention. Having once thought she would need to pursue an engineering degree to launch a product, she says that now that that’s been unlocked, distribution is the real challenge—one she was excited to tackle and that led her to pivot into marketing.

“How do you grab people’s attention? How do you build content, scale that and make it go viral in order to promote the products you can now build with AI?” asks Renee. “It’s actually the distribution side which is hard to crack and I don’t think AI can solve that for you.” 

If the tools can’t solve everything, what skills matter? As Jason was interviewing an engineer, one of his colleagues asked a question that cut straight to the point: “Coding agents are getting so good; what’s the value you’re adding as a human?”

It’s a brutal question, but Jason’s answer gets at something anyone can lean into—their unique qualities. It’s about holding context that AI can’t access, like knowing why a feature matters, understanding what a customer actually needs as opposed to what they say they need and developing good taste. 

“Businesses don’t want to make a hard choice between hiring someone really good and experienced, or someone AI-native? You want to be the person in the middle.” The most valuable people combine deep domain expertise and use AI tools to execute faster. One without the other leaves you either too slow or too shallow. 

The tactical playbook

If you’re reading this and thinking, “I’m behind”, here’s what to do tomorrow. 

“Pick a personal problem. You can go to ChatGPT, take a long walk, turn on voice mode and tell it the problem you want to solve,” says Saahil. “It will give you the instructions and the tools. You might flop doing it and you’ll break a lot of things along the way, but you’ll find the limitations of what is and isn’t possible.” 

Once you’ve shipped your first product, you need to distribute it. 

“Learn how to market yourself or market the product and get comfortable with posting on LinkedIn, Twitter, or whatever channel you choose,” says Renee. “Start to put yourself out there and get over the hump of coming off as cringe. It will open more doors and opportunities.” 

To use these tools effectively in the long term, you need to understand what’s happening under the hood. Not how to build the models themselves, but how they actually work. “If you’re working in a product org, you should learn a bit more about the fundamentals of how large language models and agents work,” says Jason. “Try lots of different tools to figure out what is best practice for the role you’re doing, which will develop your taste for what’s going to work and hit the quality bar.”

Looking ahead

Everything you’ve read here will be out of date within 6 months. Gemini 3 Pro will be old news. Claude Code will have new features. A lab will ship a new model that changes everything again. 

That’s the point. 

What won’t change is the underlying shifts. Quality, not speed, is the new bottleneck. The build-versus-buy decision is moving towards build. Revenue per employee becomes the primary metric. And when anyone can build, distribution becomes the real advantage. 

We’re already seeing this play out. Cursor reached $100m ARR with fewer than 20 employees. Chronicle is running growth with a single person. Tracksuit’s 5-person ops team is servicing 90 in go-to-market. The 10-person company competing with 1,000-person incumbents isn’t a future scenario; they’re operating today.