Where we use AI, and where we refuse to

Every owner asks us some version of the same question now: should we just use AI for our marketing and save the fee. The honest answer is that we already use it, every day, for a specific set of tasks. The problem is not the tool. The problem is owners handing it the one job it cannot do, which is deciding what the business should actually say.

We have watched a dental office publish AI-written blog posts about 'the importance of oral hygiene' that read exactly like every other dental office's AI-written post about the importance of oral hygiene. Nothing in it mentioned that the practice takes walk-ins on Saturdays, that the owner speaks Cantonese, or that the office is two blocks from the elementary school. The post was grammatically fine and completely useless, because it contained no decision about what made this practice different from the one three blocks away.

What we hand to the model

Research is the clearest case. Pulling competitor pricing, summarizing forty reviews to find a recurring complaint, or scanning a year of posts to see what topics got engagement — a model does this in minutes and does it accurately. We use it constantly to compress the boring part of research so we spend our own time on what the research means.

First drafts and variants are the second case. Once we know what a bakery wants to say about its Saturday cake pickup, generating eight headline variants to test, or rewriting one post into a shorter version for a text blast, is mechanical work a model does well. Image cleanup — removing a background, straightening a crooked photo, resizing a logo for twelve platforms — is the same category: tedious, low-judgment, high-volume, and safe to automate.

Migration work belongs here too. When we move a client's product catalog from one platform to another, or reformat three years of blog posts into a new template, that is thousands of repetitive small decisions with one correct answer each. A model does this faster and with fewer typos than a person doing it at midnight.

What we never hand to it

Positioning is a judgment call about a specific business in a specific market, and a model has no access to the information that judgment requires — what the owner is actually willing to compete on, what the last three customers said in person, what the competitor down the street is bad at. We have tested this directly: asked a model to position three different bakeries in the same neighborhood, and it produced nearly identical language for all three, because it had no reason to differentiate them beyond the words in the prompt.

Pricing decisions are the same problem in a different shape. A model can tell you what competitors charge. It cannot tell you whether raising your rate will lose the price-sensitive customers you don't want or the loyal ones you do, because that requires knowing your actual customer base, not the average customer base implied by training data.

The decision rule we use is simple: if the task has one correct answer that doesn't depend on knowing this specific business, it goes to the model. If the task requires knowing something true about this business that isn't written down anywhere, it stays with a person who has talked to the owner.

We apply the same rule to customer-facing chat and reply drafts, which is a category owners often skip. A model can draft a reply to a Google review or a DM asking about hours, and that draft is fine to use as a starting point. It is not fine to auto-send without a human reading it first, because we have seen a generated reply to a one-star review thank the customer for their five-star feedback — technically fluent, contextually wrong, and worse for the business than no reply at all.

A cost comparison worth doing once

Run the math on your own shop before deciding where the line goes. Research and first-draft work that used to take a person six hours a week — pulling competitor prices, drafting four social captions, cleaning up ten photos — now takes a model twenty minutes and a person twenty minutes to review and fix. That is real, measurable time back, and it is the honest case for using these tools at all.

Positioning and pricing work does not compress the same way, because the bottleneck was never typing speed. A boutique owner deciding whether to raise prices on a signature product needs to know that her three best repeat customers specifically mentioned the old price as a reason they trusted the brand, something no model can know unless a person tells it, and by the time a person has typed that fact in, they have already done the judgment work themselves.

So the real savings number is not 'we cut our content budget in half.' It is closer to: research and production time drops sixty to seventy percent, freeing that time for more positioning and pricing decisions actually made by someone who knows the business, which is the opposite of what most owners assume automation buys them.

The dental office, worked through

Back to that dental practice. We didn't throw out AI when we took over their content — we used it for exactly the research and drafting steps described above. We had it summarize two years of patient reviews, which surfaced a pattern the owner hadn't noticed: a third of new patients mentioned anxiety about the office specifically, and a third of those mentioned relief that the front desk explained costs before treatment started.

That's a positioning fact a model surfaced but could not act on. We wrote the actual page copy ourselves, leading with 'We tell you the cost before we touch your teeth' as the headline, because that's what real patients said mattered to them. The blog post about oral hygiene stayed generic and got ignored. The new page about cost transparency became the second-most-visited page on the site within six weeks, and new-patient calls that mentioned pricing anxiety dropped noticeably, because the page had already answered the question.

The first step

Look at whatever piece of your own marketing you wrote fastest or paid the least attention to this year — probably a bio, a service page, or a post you generated and barely edited. Read it and ask if it says anything a competitor with a different name could not say word for word. If the answer is no, that's the piece that needs a person's judgment, not more automation. If you want help sorting which parts of your content should be handed off and which need a real decision, that's what our 30-minute strategy call is for.