I sat down with Matt Swulinski for a conversation about what happens to a growth team when agents take on real work. Which roles stop making sense, which ones get more valuable, and why the agentic system he built at Wispr Flow took a new approach to scale to his team.
If you haven't seen it yet, here's the full session on YouTube:
About our guest
Matt was the first marketing hire at Wispr Flow, where he built one of the most advanced AI-powered growth systems in SaaS and owned every KPI, from eyeball to app download, with paid spend peaking around $4 to $5 million per month. He's now Head of Growth at Viktor, an AI employee you hire into Slack or Microsoft Teams. He's a month into the job after the company raised a $75 million Series A, the largest ever raised by a Polish-founded company.
He has built the system, hired against it, and watched where it breaks. Here's what stuck with me from our conversation.
The system gets built out of self-defense, not strategy
Matt built an agentic growth system because the alternative was drowning.
At Wispr Flow, he was, for a long stretch, an army of one: every ad, every experiment, and every channel went through him. The system came out of necessity.
I had to become agentic, otherwise I would have lost my mind.
What got automated first was the menial work — creative analysis, moving ads from testing to scaling. Right up to his transition, he was still running all of Google Ads himself through the CLI.
The lesson isn't "build a platform." It's that a forcing function produces better systems than a roadmap does, because you only automate what is actually costing you.
Moving to a multi-player system
The break came in January, when Wispr Flow started hiring new marketing roles: a media buyer, a creative strategist, someone for influencer partnerships. Matt was handing off finished copy and asking them to ship it, which is a handoff, not a workflow.
I realized, okay, this is a single-player system. I have to be able to make this multiplayer.
The fix took about thirty-six hours with Claude Code: a git-controlled marketing OS with a context layer for the things the whole team depends on, like contracts, live campaigns, dates.
Once the system was multiplayer, several team members onboarded and started building their own workflows on top of it. But the other half never adopted it at all.
The blocker was systems thinking. Matt is direct about not being an engineer himself, but he learned to think in systems because he had to. Things like markdown files, folder structures, git, PRs clicked. But for some people, it's a different side of the brain, and more training doesn't move the needle.
No matter how much hands-on training I did, there were just some people that couldn't get it. That's not the way they were hired for.
The cost of that split shows up in output. The half of the team that didn't adopt kept asking Claude one-off questions and pulling data, and never compounded any of it.
Two people with the same title, on the same team, with the same tools, ended up an order of magnitude apart. That gap is a hiring problem, which is where the conversation went next.
Hire the best in the world (who is also AI native)
Viktor's hiring philosophy is one line: hire the best in the world that is also AI native. Both halves are load-bearing.
Matt's argument is that the unicorn candidate who isn't AI native loses in the long-term.
If I only looked for the best media buyer, that spent the most money and scaled the most companies, but is not AI native, I bet that within six months, someone that is top 10—plus agentic and curious—will immensely outperform that person.
The model he describes for a department is one human owner with agents underneath: "they're the board of directors of their department, and then all of the employees are really AI." The overseer, the brain, and the strategy stay human.
On the other end of this, one Viktor customer, has a 6-7 person agency that has stopped hiring entirely. They went deep on removing the reporting, the follow-ups, the drafted emails, and the alerts that tell someone an ad needs action. The traditional analyst and junior media buyer roles came off the hiring roadmap and haven't come back.
Deploy agents first, hire when needed
Agents get you from 0 to 1 and establish what works. But humans are needed once you want to scale the thing that's working.
His job, as he defines it, is hunting top talent. When he finds a real gravity well — someone who can bring the kind of unlock that moves the business — an agent will lose that battle every time.
Throwing an agent at it is kind of lazy. It will upkeep only the scope and the setup that you've given the agent. It is not its own self-improving thing.
The example he gives: an agent optimizing Meta ads inside a defined scope will hold that scope, and it will never be the thing that proposes a drastically different strategy. So the sequence runs the other way from how most teams do it. Spin up the basic version of everything you should be doing and put an agent on each one. When one of them gets big enough to matter, hire a person to work alongside the workflow that already exists, and let them scale it with a hundred agents behind them.
Prioritize the feedback loop
We closed on advice for people trying to level up how they work with AI, and Matt's answer was the least sophisticated thing he said all session, on purpose.
The most important part is the feedback loop, not the prompt. Don't edit the agent output in a doc and move on. Feed changes back to the agent, ask what changed and why, and turn the answer into a skill, so the corrections persist and the agent improves over time.
When you're a junior employee, you're not doing things the right way either. Treat the models the same way and give them the time and the attention.
No matter how sophisticated a system, it's not going to work if you're not bringing some sort of improvement loop.
Agent-led customer engagement
Thanks to Matt for a great conversation. If there's one thing to take back to your own team, it's that the constraint on agentic growth isn't the agents. It's whether the humans around them can think in systems, and whether you've hired for that.
The pattern Matt describes—agents establish the surface area, humans own the thing that's working, and every correction gets persisted so it compounds—is the same pattern the Knock agent uses to automate and improve customer engagement work over time.
Knock gives growth teams agentic messaging infrastructure to build lifecycle messaging, onboarding flows, and product notifications across email, in-app, push, SMS, and Slack, with guardrails to review before anything goes out.
Create your free Knock account to get started, or talk to our team to learn how Knock can power agentic customer engagement at scale.