How agencies actually use AI: three practitioners on infrastructure, coordination and thinking

Most posts about AI in marketing are about generating content. So we asked practitioners a plainer question: what do you actually use it for, and did it cost anyone their job? Three of the answers were specific enough to publish. What stands out is that none of the three describes the writing itself as the point. They use it as production infrastructure, as a coordination layer between departments, and as a thinking partner. One of them does have the tool write the deliverables, and even there what she credits is the validation around it.

How we gathered this

We put out a call for sources on Connectively, the source-request platform run by Featured, which bought HARO from Cision in April 2025 and now operates both. The question was how businesses use AI like Claude, and whether it led to team cuts or just shifted what people work on. We received 85 responses.

Most were not specific enough to publish: generic descriptions of AI that could have been written about any tool by anyone. We shortlisted the ones that said something concrete, and among those we looked at closely, several carried the platform’s own automation flag, which is its own small irony given the question. We then verified every source we wanted to quote: that the person exists, that the company is real and operating, and that the job title matches the public record. Two of the three checked out on every point. For the third we corrected the attribution, and we say so below rather than quietly printing a title that overstates the business.

That is the whole method. Three voices, not a survey, and we are not going to dress them up as statistics.

“We run Claude as infrastructure”

Her six-person agency in Mount Pleasant, South Carolina runs the most automated setup of the three, and the most carefully bounded.

We run Claude as infrastructure. It’s the production layer the business runs on. A brief gets approved up front, and from there the system researches, writes, fact-checks, and validates the work itself. What comes out the other end is finished (a report, a forecast, a proposal) not a draft.

What she built around it matters more than the automation.

We use Claude to build and validate. The validation is the whole thing. Every deliverable has to clear a series of gates before it can move forward. It gets scored against a written standard for that type of work. It gets an editorial pass. Every factual claim is verified against its source. It clears a compliance read. Fail a gate, and it doesn’t advance.

Then the line that separates this from most “we use AI” stories:

We don’t do those edits. That’s the part people get wrong when I describe this, they picture my team cleaning up AI output. We’re not. If we’re hand-editing a deliverable, something upstream is broken. So we don’t fix the document, we fix the rule that let it out. The standard gets rewritten, the system inherits it, and everything produced after that is better.

That is a real distinction. Hand-editing AI output is a treadmill: you pay the same cost on every deliverable forever. Rewriting the standard is a fix that compounds, because everything produced after it inherits the correction. Her own joke about it lands the same point: “I always joke about human-in-the-loop…we need more AI in the loop.”

Worth noting where the time went, because this is the part that matters commercially. She says it was reallocated to three things: client-facing work, strategy, and improving the process internally.

The coordination problem

SmartSites is a digital marketing agency in Paramus, New Jersey, considerably larger than the other two operations here. Her use case is less about producing the work and more about holding a lot of accounts together.

At SmartSites, we integrate Claude into almost all of our processes. We’re a digital marketing agency with a heavy focus on SEO/AEO/GEO and web design/development. We also have a robust PPC department as well as a social media team. We manage a lot of independent clients, many of whom are SMBs but some of whom are also enterprise-level businesses. Keeping track of so many accounts requires consistency across messaging and processes. It also requires quick responses.

Claude helps streamline processes so we’re able to respond to clients quicker without sacrificing accuracy. It enables us to stay present in meetings, send thorough follow-up emails and directives, and communicate with various team members across different departments. By using the skills and connections features, we prevent miscommunication and help all teams get the information they [need] so we can deliver the best results for our clients.

“Stay present in meetings” is the phrase to notice. What she describes is a coordination tax: running many accounts across PPC, social, SEO and development at once, where the same client fact has to reach several teams without mutating on the way. That is an unglamorous problem and it rarely shows up in AI marketing copy, which is probably why it is worth reporting.

“AI is a terrible writer”

Raisor’s practice is a one-person operation in Kentucky, and he writes a Substack of essays on the side. He submitted under the title Growth Director, but since the business is a solo practice we are attributing him as the person who runs it rather than printing a title that would read as an agency executive. His answer was the most contrarian of the three, and he opened with it:

AI is a terrible writer, unless you train it extensively, or edit it heavily. However, it is a great tool to bounce ideas off of, and the more detailed you get in the ideas, the better the tool works. Not only does it provide valuable feedback and an alternative perspective, but it provokes thoughts I wouldn’t have arrived at.

What he uses it for instead is unusual enough to quote at length:

In addition, I have a very difficult time breaking ideas down into analysis. I’m very intuitive and arrive at theories and conclusions without knowing how I got there. AI is an amazing tool for diving into these details and understanding my own subconscious processes. So, typically, I’ll arrive at an idea for content, and have an extensive conversation with Claude to make sure its unique, and to understand the details of how I got there.

Reverse-engineering your own intuition into something you can explain is not a use case you see on product pages. It is also the hardest of the three to hand off, because the raw material is the practitioner’s own judgment. The tool is doing interrogation, not production.

So did it cut any jobs?

This was the second half of what we asked. None of the three reported a job lost to it, and only one of them answered the question in enough detail to publish.

Bermudez was the one who volunteered the uncomfortable version, and it is the most useful sentence in the whole set precisely because it works against her own story:

We haven’t replaced people, but the work has changed. We have allocated time in three places: in front of clients, on strategy, and on improving processes internally. We are a small team, and realistically, in a larger organization, a system like ours could lead to work being consolidated and cutting teams. I think there has to be a balance.

A six-person agency absorbs freed capacity by pointing it at clients and strategy, because there is always more of both. That arithmetic changes in a larger organization, and she says so rather than pretending it does not.

What we take from it

Three businesses of very different sizes, three jobs for the same tool, and none of them is simply generating content. Production infrastructure with gates around it. Coordination across departments. Interrogating your own thinking. The writing is not where any of them locates the value, and one of them was blunt about why.

That maps to what we see in our own work. More text is not scarce. What is scarce is the checking, the consistency and the parts of the process nobody wants to do by hand. If you want the search side of that, our take is in our AI SEO strategy guide and in the AI SEO service we run for clients.

One footnote on method, since it is relevant to the topic. We asked how people use AI, and a number of the answers we read appeared to have been produced by it, generically enough that they said nothing at all. The three above are the ones that carried something specific enough to argue with. That is the filter, and it is the same filter a reader applies.