How AI Should and Shouldn't
Be Used in Marketing

What Actually Changed

For twenty years, agency economics ran on a pyramid. A senior strategist sold the account, a few mid-level people managed it, and a large base of junior staff did the work: first drafts, keyword research, ad variations, reporting, resizing, scheduling. The client paid senior rates and received junior output, and everyone pretended not to notice.

Language models did not replace the strategist. They replaced the base of the pyramid. The first draft, the research pass, the twenty headline variants, the spreadsheet cleanup, the transcript. Work that used to take a junior person a day now takes a senior person twenty minutes of prompting and forty minutes of editing.

That is the whole story, and it has two consequences. Agencies that were built on the pyramid are now either laying off the base or pretending the base is still doing the work. And clients who understand the shift can now get senior judgment on everything, at a price that used to buy them a junior account coordinator. We wrote more about the economics of that in Senior-Led Marketing ROI.

Where AI Earns Its Keep

Research and synthesis. Reading fifty competitor pages, summarizing a hundred reviews, pulling the themes out of a year of customer emails. The model does in minutes what a person does in days, and it is better at not getting bored on page forty.

First drafts and variations. A working draft of a landing page, thirty subject lines to pick from, an ad set in five lengths. The draft is not the deliverable. It is the starting point that saves the blank page.

SEO and content analysis. Mapping what a site covers against what people search for, finding the gaps, identifying which pages are cannibalizing each other, spotting the questions a page should answer and does not. This used to be a spreadsheet exercise that nobody finished.

Data work. Cleaning exports, reconciling ad platform numbers with CRM numbers, writing the query, building the chart. The model is a competent analyst who never complains about the formatting.

Production support. Transcription, caption generation, rough-cut selects from an interview, alt text for a hundred photos. Tedious, necessary, and now nearly free.

Code. The internal tool, the integration script, the one-off migration. Custom software that used to be too expensive to justify for a mid-size business is now often a day of senior time.

Where It Fails, and Why That Matters to You

Final copy without a human edit. Model output has a texture. The rhythm is too even, the transitions are too smooth, the lists come in threes, and certain words show up constantly. Readers have learned to recognize it, and when they do, they discount it. Copy that ships unedited is copy that tells the audience nobody at the company cared enough to read it.

Facts. Models produce confident, fluent, wrong statements. A statistic that sounds right, a regulation that was repealed, a competitor feature that does not exist. Every factual claim in AI-assisted content needs a person who knows the subject to check it. In legal, healthcare, and financial marketing, this is not optional.

Strategy. A model can list what a business in your category typically does. It cannot know that the biggest competitor in Greenville just lost its founder, that the county is about to change a permitting rule, or that your best referral source is a specific surgeon who hates email. Strategy is made of local, current, human knowledge. The model has none of it.

Voice. A brand voice is a set of choices about what you would never say and how you say the things you do. A model will approximate a voice from examples and drift from it within three paragraphs. Keeping it honest is an editorial job.

Judgment under pressure. Crisis statements, sensitive client communication, anything where the wrong word has consequences. A model can draft. A person with experience decides.

How to Tell Whether Your Agency Is Using AI Well

Ask them directly. "Which tools do you use, for what, and who reviews the output?" A good answer is specific and unembarrassed. A bad answer is either "we don't use AI" (they do) or "we're an AI-first agency" with nothing behind it.

Ask who edits. If the answer is a junior coordinator, you are paying senior rates for a model plus an inexperienced reviewer, which is the worst of both arrangements.

Ask what got cheaper. If AI made the agency more efficient and your invoice did not change, the efficiency went somewhere. It should be showing up as more senior attention on your account, more output for the same retainer, or a lower price. Ask which.

Read the work. If the blog posts sound the same as every other agency's blog posts, the model wrote them and nobody rewrote them. If they sound like a specific person with opinions, someone did the job.

Ask about your data. Where does client information go when it is put into a tool? Reputable agencies use enterprise or API access with no training on your data and can tell you so in writing.

What Every Business Should Do Now

Whether or not you hire an agency, three things are worth doing this quarter.

Make your website answer questions. AI assistants and AI-generated search results pull from pages that state things plainly: what you do, where, for whom, what it costs, how it works. Pages that are all slogan and no substance do not get cited. This is the biggest shift in search in a decade and most regional businesses have not noticed it yet.

Get your facts consistent everywhere. Name, address, phone, hours, services. Across your site, your Google Business Profile, directories, and social accounts. AI systems cross-check, and inconsistency reads as unreliability.

Decide what you will not automate. Customer replies, reviews responses, anything a client reads that carries your name. Draft with a model if you like. Send only what a person has read.

How Sidestreet Uses AI

We use Claude and in-house models across the agency, every day, and we say so. Campaign research, SEO analysis, first drafts of copy, ad variation, reporting, transcription, caption work, internal tooling, and a good deal of the code behind our own applications and client integrations.

Every output that a client sees is edited by a senior person. Not reviewed. Edited. The copy on our own site went through that process. Some of it started as a model draft and almost none of it survived unchanged.

The economic effect is that we run without a junior layer. There is nobody to hand your account off to, because the work that a junior would have done is now done by the model and the judgment is done by the people you met. That is what "no junior handoffs" means in practice, and AI is why it is affordable.

Client data goes into tools under agreements that do not train on it. We will put that in writing. If you want to talk about what AI should and should not be doing in your marketing, we are glad to have that conversation without the hype.