B2B Marketing ROI: AI Bazooka or Slot Machine?

Picture of Mark Donnigan

Mark Donnigan

|

July 7, 2026

AI is not coming for marketers’ jobs. It is handing some of you a bazooka and handing the rest of you a slot machine, and the difference has nothing to do with which tool you bought.

On a recent episode of the Marketing Execution Podcast, Tim and I picked apart the idea of AI as a creative bazooka for marketers.

I realized that AI becomes a weapon aimed at the work that matters once everything else gets out of your way. That framing cuts against the panic most marketing teams are sitting with right now. The fear is that AI writes the copy, builds the sequences, and makes the marketer optional.

What is actually happening is narrower. AI is absorbing the mundane layer of marketing work, such as admin, first drafts and the repetitive prospecting that used to eat entire weeks. That time only pays off if you have something worth thinking with once you get it back.

This article walks through where AI earns the bazooka label, where it turns into a slot machine instead, and what to build first if you are running a five-person team rather than a 350-person organization.

Why AI Replacing Marketers Is the Wrong Fear

I know a Series A B2B SaaS founder who spent weeks crafting outbound sequences by hand. AI now drafts the first version in minutes. The idea behind the sequence still comes from a person. So does the positioning, and the read on what will actually resonate with a specific buyer.

AI does not replace the thinking behind a campaign. It removes the busywork standing between you and that thinking, and the removal only pays off if you spend the freed-up time thinking harder.

That is the difference between the two machines in the title. A slot machine pays out at random. You pull the lever, get a draft, pull again, and keep whichever output looks best, with no reliable way to judge whether any of them is right for your buyer. A bazooka only does damage where it is aimed, and the aim comes from what you already know about your market, your customers and the reasons your deals close or stall. Same tool, same subscription, very different return.

AI taking the job was never the real risk. The risk that actually plays out is founders spending the extra hours on more of the same, instead of the strategic work nobody else on a small team has time for.

Every section that follows comes back to one test. Did a person’s domain expertise shape the outcome, or did the tool run unsupervised? That answer decides whether the time AI freed up actually paid off.

What Klarna’s AI Bet Shows About Domain Expertise

Klarna gave the clearest public example of AI running at volume. In February 2024, the company announced that its OpenAI-powered assistant had handled two-thirds of its customer service chats in its first month live. By Klarna’s own figures, the assistant did work equivalent to roughly 700 full-time agents and resolved issues in under two minutes versus eleven before.

Fifteen months later, the story looked different. In May 2025, Forbes reported that Klarna was hiring human agents again after service quality slipped, and that CEO Sebastian Siemiatkowski had acknowledged an overemphasis on cost led to lower quality. Klarna disputed the “reversal” framing. It told Forbes it still invests heavily in AI and that the assistant now does the work of more than 800 roles. It also said it never eliminated human support and that the new hiring is a small pilot combining AI with human agents.

Read both sides and the lesson holds. The AI handled volume well. What it could not supply on its own was judgment on complicated cases, the ones that need empathy or a read on what a specific customer actually needs. That is the territory domain expertise covers, and it is why the human layer moved back into the design.

I’ve watched the same pattern play out at a smaller scale. A company deployed AI agents across its entire sales process and watched it fail completely, and not because the prompts were bad. The agents had a clear outcome to hit but no access to the why behind each decision, such as why path A over path B, or why an objection gets a different answer depending on account type. That context lives in people’s heads.

What no model can graft on

Kieran Flanagan, HubSpot’s SVP of Marketing, AI and GTM, has argued that context grafting is replacing complex prompts, and I think it becomes the defining AI conversation of the next year. You can graft documents, transcripts and data onto a model. Some of what domain expertise gives you does not transfer that way:

    • Pattern recognition built from years of seeing what actually works in your market.
    • Gut calls made under real uncertainty, with incomplete information.
    • The unwritten reasons behind past decisions that nobody on the team ever typed up.

There is early evidence that handing over the thinking itself carries a cost. In a 2025 MIT Media Lab study, 54 participants wrote essays using an AI assistant, a search engine or no tools while researchers measured brain activity with EEG. The AI-assisted group showed the weakest brain connectivity and, over four months, consistently underperformed at neural, linguistic and behavioral levels. It is a small preprint on essay writing, not a verdict on all knowledge work, but it points the same way as Klarna. Outsource the judgment and you get volume without the read that makes volume useful.

The One AI Marketing Workflow to Build This Week

A founder running a three-person team and an SVP running 350 people are solving the identical problem. Each has to know when to trust the model and when to trust their own read on the company. Tool budget does not change that question.

Start with research, not content

If you are starting from zero, skip content creation as your first AI workflow. Start with research instead. It offers the best return for the lowest risk, since nothing you produce there goes public. You are gathering material faster than you could alone, and you remain the filter on what gets used.

Last week I built a small research assistant in Claude for a piece on why marketers need to break from inherited best practices, and asked it for a story that would stick. It surfaced Dick Fosbury. As a high school jumper in Oregon in the 1960s, Fosbury dropped the standard technique and started going over the bar backward, an approach made safer as foam landing pits replaced sawdust. He won Olympic gold with it in Mexico City in 1968, and by 1980 most Olympic high jump finalists were using his flop.

I would not have found that story on my own, but I had the domain expertise to know it fit. AI finds the options, and you make the call. The model widens the search, and your judgment narrows it to the one option that fits the argument and the audience.

Split one outreach prompt into four jobs

The second habit worth building now is to stop writing one giant prompt for outreach. Split it into pieces, because each piece solves a different problem:

    • The subject line earns the open through curiosity.
    • The opening shows relevance to this specific buyer.
    • The body carries the proof.
    • The close makes the next step clear.

One prompt cannot serve curiosity, relevance, proof and clarity at the same time, because each asks the model to weigh different things. Splitting the work lets you test and fix each part on its own, and it gives you an obvious place to add what you know about the buyer. The same thinking sits behind how we structure AI prompts for our own content.

HubSpot’s results point the same way. Speaking on the Ahrefs Podcast in February 2026, Flanagan said AI-personalized email workflows built on custom prospect data increased sales meetings at HubSpot by 30%. That is a self-reported result from HubSpot’s own program, not an independent study. Still, the source of the gain is telling. It came from personalization grounded in data about each contact, not from one clever prompt.

None of this requires becoming a prompt engineer. It requires staying sharp enough that when AI hands you three options, you can tell which one is right.

Know when to close the tab

Some weeks, the right move is closing the tab entirely. I had three real strategy problems last week and didn’t touch AI for any of them, because I had the context to work through them myself.

That is not an argument against AI. It is the same test applied in reverse. When a problem depends on context only you hold, the model has nothing to aim at, and asking it anyway turns the bazooka back into a slot machine.

Why Founders Agree and Still Don’t Build the AI Layer

Knowing this framework and running it are different problems. Most founders I talk to agree with everything above and still don’t have a research assistant built, prompts modularized, or data sources mapped three months later.

The belief behind the delay is usually that the founding team can build this itself, in the gaps between everything else. The gaps rarely come. Survival work fills every open hour, and nobody on a five-person team owns building the AI layer on top of the strategy, the data, and the judgment calls that make it work. That is true for most early-stage startups I work with, where the founder is still running product, sales and marketing at once.

So the gap is not tools or ideas. The person with the market judgment has no time, and the people with time are not positioned to make the judgment calls. A Marketing Pod is built for exactly that gap. A Virtual CMO who owns positioning and the marketing plan works with a writer, a designer and a digital marketer who build and run these AI-enabled workflows alongside you. The judgment that aims the tools stays connected to the people doing the work, instead of living in a slide deck about them.

Aim AI Where Your Judgment Already Lives

The return on AI in B2B marketing depends less on the tool than on whether someone with domain expertise is aiming it. Klarna’s volume gains were real, and so was the quality gap that brought more human support back into its model. HubSpot’s meeting lift came from personalization built on real prospect data, not from a single clever prompt.

For a small team, the practical move is to pick one workflow this quarter. Research is the safest first choice, because nothing goes public until you have judged it. Build it, review every output for the first few weeks, and note where your own knowledge changed the result. Those are the points where AI is aimed. Everywhere else, you are still pulling a lever.

If you want a working model for keeping your thinking in charge while AI carries the structure, start with the Second Brain Copilot framework.

Listen to the Full Episode

Hear Tim and me break down the AI workflows Kieran Flanagan built at HubSpot, and what small teams can borrow from them.

Let’s Talk

I can’t wait to make you my next success story. 
Please complete all fields so that we can explore together how best to help your company dominate the market and achieve its full potential.

By clicking the “Submit” button, you are agreeing to our Terms of Use and Privacy Policy.

DOWNLOAD NOW!

To access your free downloads, please provide the following details and hit the Submit button.

Once you have submitted the form, check your email for the download links.

[sibwp_form id=4]

By clicking the “Submit” button, you are agreeing to our Terms of Use and Privacy Policy.

LET'S TALK

I can’t wait to make you my next success story. Please complete all fields so that we can explore together how best to help your company dominate the market and achieve its full potential.

By clicking the “Submit” button, you are agreeing to our Terms of Use and Privacy Policy.