A marketing head sits down to plan next year’s budget and hits the same fork every operator hits. Hire the work in-house, or rent it from an agency. This year a third option crashed the meeting, and its name is AI. That fork is what Steve and Peter took apart on the latest Misadventures in Marketing, the AMA San Francisco podcast, in an episode they called “Agency vs. In-House” Steve Haney, a fractional go-to-market executive in Silicon Valley, kept pushing one uncomfortable question. Peter Farago, a startup marketing executive in the San Francisco Bay Area, kept refusing the easy answer.

Does AI make marketing agencies obsolete?

Short version, from someone who builds AI for a living: not yet, and maybe not ever. Peter’s company makes AI tooling for engineers, and he lives inside these systems for hours every day. His verdict on the question is blunt. “AI does not replace full roles very easily. It replaces parts of them.”

That one line resets the whole hire-versus-rent question. The choice was never really people against software. It’s about which parts of the work can leave the building, and which parts can never leave at all. Frame it that way and the budget meeting gets a lot clearer.

Steve pressed the opposite case, and his version deserves a fair hearing. Picture the closed loop working: the AI reads your funnel, scores intent, watches where each prospect sits in the buying cycle, and answers the board’s questions on demand. “Why would you ever have to hire external resources,” he asked, “when the AI is just going to do that for you?” It’s a real question, and half the room already believes the answer is yes.

The work that already collapsed

Numbers make the case better than opinions do. The cost to build a company website used to start around 30K and climb past 50K, depending on how nice it needed to look. Then companies like Webflow dragged that toward 10K. Peter’s own CEO, a computer science PhD building on the frontier, found a LinkedIn post from a founder who went further. “I dropped Webflow and I just had Claude build my website.”

Peter has watched the same collapse inside his own workflow. His company runs highly customized sales decks, one intricate manual build per customer conversation. So he built a deck maker in Claude with a front end his sales team can actually use. A rep pastes in a transcript of the last customer call, answers a few prompts, and a branded deck comes back with the customer’s own pain points repeated in the customer’s own words. The tool does 80 to 90 percent of the work. He pins the link in a Slack channel, wires it to a company-level Claude project, and removes himself as the bottleneck entirely. “It’s an entire software workflow,” he said. “It’s insane.”

Notice what actually got automated. Layout, assembly, first drafts, the tedious plumbing between a transcript and a finished artifact. The parts a junior person used to grind through for a day now take minutes. That is a real change, and pretending otherwise insults anyone who has built a deck at midnight.

The work that won’t leave the building

Then Peter drew the line, and he drew it in a place worth memorizing. “These AI tools can’t design for you in an original way, not in a meaningful way, not at the standard you and I have.” He was direct about where he still spends money. “If I have to create company identity, I am working with real designers, period.” Taste and craft don’t export. Hand AI a finished design and it will stretch that look across fifty slides. Give it a blank page and ask for a brand, and it has nothing to offer.

Press relationships fail the export test for a different reason. A good PR agency knows which reporter at the Wall Street Journal covers which beat, gets drinks with them, and can actually get a reply. That kind of access always carries a price. Peter put a real full-service agency at a quarter of a million dollars a year, roughly 25 to 30 grand a month, versus maybe 10 grand a month for a solo veteran gone independent. The relationship is the product, and no model has one.

Paid media sits in the same bucket, though Peter called it something better. Search, SEM, bidding strategy, and the parts of SEO that still move all carry, in his words, “a black arts aspect.” The algorithms move weekly. Staying current is a full-time job, which is exactly why he hands it to an agency that runs it across many clients and keeps a cut of the media spend. You outsource what moves too fast to track yourself.

The quietest constraint might be the most important. Peter’s stack runs HubSpot for marketing automation and Attio for CRM, and the two don’t sync cleanly. He wants closed-loop reporting, from first touch to closed deal, and the pipes between his systems won’t carry it. AI does nothing to fix that gap. “Garbage in, garbage out,” he said, and “AI on just HubSpot won’t solve all my problems,” because HubSpot is only one piece of what he has to reconcile. Point a brilliant agent at broken data and you get confident, well-formatted nonsense.

The filter that survives AI

Strip away the tools and Peter’s decision rule is old and durable. Start with what you’re trying to get done, then ask what you need to own versus what you can rent. Positioning, story, and brand come from inside the company, even when an agency runs the workshop. Two things justify paying outside money: expertise that’s hard to get and genuinely valuable, plus the raw need to scale faster than a small team can.

His operating model carries a name he repeats often: learn, then design, then delegate. Before he handed Google paid search to an agency, he ran a couple of campaigns himself, just enough to guide them on keyword clusters and match types. “You can’t completely export it,” he said, because you have to know enough to manage the people you hired. That same principle holds double for AI. You can delegate the doing to a machine. What you cannot delegate is the judgment about whether the doing is any good.

And none of this ever comes for free. “Every agency you manage, every outside contractor you manage, takes part of your time,” Peter said, to keep the quality up and the plan coordinated. The same will be true of a team of agents. Somebody on your side still owns the outcome.

Steve’s engineer friend thinks the ground is about to move again. He sees the web itself getting wrapped in AI plumbing, banking and reporting handled by conversation instead of clicks. “In 10 years, kids will be asking, what’s a URL?” And the advertising-funded web, in his phrase, looks like “a comet coming out of the sky about to crash into the Yucatan.” Maybe. Even then, the in-or-out question doesn’t disappear. It sharpens into something more honest: what can only a human do, and who do you trust to do it?

Listen to the full podcast on Spotify and Apple Podcasts.

Check out summaries from other episodes:

• Founder-led marketing: who’s actually driving the brand
• The Artist Is the Planet: Marketing When Distribution Is Free
• Pitch Deck Storytelling: Your Slides Are Footnotes

Misadventures in Marketing is a weekly podcast by the AMA San Francisco chapter. Veteran Silicon Valley marketing execs Peter Farago and Steve Haney explore the messy, rewarding, and occasionally absurd world of high-tech marketing  –  especially in early-stage startups. Each episode covers real-world challenges, trends, and lessons from the front lines.