Sales and marketing
AI for marketing agencies: where it saves hours, where it hurts
For owners of three to fifteen person shops. Where the hours disappear, where AI costs you work, and what to do about pricing when production time collapses.
If you're here to hire an AI marketing agency, this isn't that page. Nothing on it will help you pick one.
This is for the person running a four to fifteen person shop. Your production time has collapsed and your pricing hasn't caught up.
So the useful version of this article is short on tool names and long on the two conversations nobody is having with you: what to check before it leaves the building, and what to charge now.
Where the hours actually go
Three places, and they're narrower than the tool lists make out.
First drafts
Blog posts, email sequences, ad variants, the long explanation section of a deck. Work your team was always going to rewrite.
Starting from a draft you dislike is faster than starting from nothing. That sounds like a slogan and it holds up every time, because disliking something specific is a much easier job than inventing something general.
Reporting preparation
Not the insight. The assembly.
The useful version of this is boring and exact. Export the month from each platform. Paste the numbers into one place. Then the ask is "put this on one page, flag anything that moved more than a tenth either way, and write the two paragraph summary a client would actually read."
Your strategist still writes the insight. What just disappeared is the ninety minutes of assembly that used to happen before the insight could start.
Research and briefing
Reading twelve competitor pages and coming back with what they all say.
The instruction that earns its keep is the negative one: "across these twelve pages, which claims show up on more than half of them?" Because what you actually want is the list of things nobody is saying, and you cannot see that until you can see the consensus.
That's the honest boundary of where this helps. Everything past it costs you something, and the price is not obvious until it has already been paid.
Where it actually costs you
Three ways, and the third one is the one worth losing sleep over.
Your client's voice, not yours
A tool that has read everything writes like the average of everything. Average is the thing your client pays you to avoid.
So capture the voice properly, once per client, and keep it in a file somebody owns. Six pieces they signed off without edits. Two they sent back. The reason they sent those two back, in your account lead's own words.
Then the ask that does the most work: "list the twelve words or constructions this client never uses." That list is worth more than the six good pieces, because it is the thing a tool will reach for by default.
Accuracy in the reporting
A confident wrong number in a monthly report is worse than a late report. The number is the one thing a client checks.
So the figure check is not a proofread, it is a trace. Open the source, read the number, read the number in the draft, move on. A person who knows the account does it in about ten minutes for a monthly report, and it is the single cheapest insurance on this page.
Sameness
Brief the same tool the same way as every other shop in your category and you get what they get, faster.
So put the decision earlier in the week. Before anything gets generated, somebody writes one sentence: this piece exists to make the reader believe X. One sentence, in the brief, with a name against it.
Then the sentence is the thing the work gets judged against, rather than the draft being the thing that quietly becomes the decision.
Figure 1 / Boundary
AI earns its place only where the output was always going to be reworked and a mistake gets caught before a client sees it.
Where it helps and where it hurts
Where hours go
First drafts
Work your team was always going to rewrite anyway.
Reporting preparation
Not the insight. The assembly.
Research and briefing
The reading was never the job.
What it costs
Voice
A tool that has read everything writes like the average of everything.
Accuracy in reporting
A confident wrong number is worse than a late report.
Sameness
Brief it like every other shop and you get what they get, faster.
The lineOnly where somebody reworks it anyway before a client sees it.
The delivery check
Four gates per deliverable, and a name against each one. This is the part to print.
- Facts and numbers against source. Every figure traced to where it came from. No exceptions and nothing on the grounds that it looked about right.
- Client voice. Read by somebody who knows the client. They are listening for the two or three words the client would never use.
- The generic sweep. One pass, one question. Could this have been written for any of our other clients? If yes, it isn't finished.
- The human addition. Name the one thing in here a tool couldn't have produced. If nobody can name it, you are selling what everyone else sells with your logo on it.
The objection to this is that a six-person shop does not have four reviewers. Correct, and it does not need four. It needs three names, and it needs them written in the same place every time rather than decided per project at five on a Thursday.
Gates one and three take about ten minutes between them. Gate two is a read-through by somebody who was going to read it anyway. Gate four is a sentence.
Want this as a procedure your team follows rather than four bullets in an article? The prompt and the five-step check are free and take an afternoon.
Figure 2 / Flow
Four gates in order, each with a name attached, and a deliverable leaves only after the fourth.
The delivery check
Needs a person
Facts and numbers
Every figure traced to where it came from.
Needs a person
Client voice
Read by someone who knows the client.
Needs a person
The generic sweep
Could this have been written for any of our other clients?
Affirmed
The human addition
Name the one thing a tool couldn't have produced. Then send.
- Needs a person
- Affirmed
The conversation nobody writes about
Here's what the tool lists skip, and it is the thing that decides whether your year is good or bad.
- 10 hrswhat the deliverable used to take
- 3 hrswhat it takes now, at the same quality or better
- 70%of your own revenue on that line, gone, if you bill by the hour
The work is as good or better. The client is as happy or happier. And you have quietly cut your own invoice by most of itself for doing the job well.
Three ways through it, and they are not equally good.
| The move | What it costs you | What it costs the relationship | How long it lasts |
|---|---|---|---|
| Keep billing the old hours | Nothing today | Everything, the month somebody asks how long it took | Until that question |
| Price the outcome or the retainer | A harder sales conversation | Nothing, if you do it on new business first | Indefinitely |
| Sell more volume at the lower price | A price fight you didn't pick | Delivery quality, when you overcommit | As long as your team holds |
The thing that stops owners doing the second one is not the pricing model. It is the fear of the conversation with a client who already has a rate.
So don't have that conversation. Not yet.
- Price the next new client on the outcomeOne client, one scope, a number for the thing rather than a number for the week. You now have a live comparison instead of a theory.
- Work out what the old work was really worthTake three finished projects off the last year. What would you have charged for the result if the hours had never been mentioned? Write all three numbers down before you look at what you actually billed.
- Move existing clients at renewal, not mid-flightRenewal is the only moment a price change is a normal business conversation rather than a surprise. Nobody gets repriced halfway through something they already approved.
- Say the quiet part out loud, once"We've moved to pricing the work rather than the hours, because the hours are our problem to solve and not yours to pay for." That sentence is the whole conversation. Don't add to it.
The objection you will get is "so my price should be going down, then." It is a fair question and it has a real answer, which is that the client is buying the thing at the end, the judgment that shaped it, and the fact that it is right the first time. None of those three got cheaper. One input did.
What to tell clients
Disclose at the process level rather than asset by asset. One paragraph, in your onboarding and on your site, so everybody in your shop says it the same way.
One thing first, in a single line: if you have signed something that governs who gets to touch a client's material, that document outranks anything on this page, and the person who reads it is your lawyer rather than you or us. That's the whole of the legal part and we have nothing useful to add to it.
Here's the shape of the paragraph. Change it to be true of you, because the only version that helps is the accurate one. "We use AI tools in parts of our production, mostly first drafts, research and the assembly of reporting. Every deliverable is reviewed by a named person on your account before it reaches you, and we check every figure against its source. We don't put your unreleased work, your numbers or your client list into any tool. If you'd rather we didn't use these tools on your account at all, say so and we won't."
Say it first and it is your position. Get asked halfway through a project and hedge, and it is a problem you handed yourself.
Figure 3 / Sequence
Disclosure is a position you state once, and anything you have already signed comes before it.
A position you state once
Anything you signed comes first
If a document governs who touches a client's material, your lawyer reads it, not you and not us.
Write one paragraph about how your studio works
Disclose at the process level.
Say it in onboarding and on your site
Write it once and have everybody say it the same way.
Client data, the short version
Three things end agency relationships. Unreleased client work, client numbers, and the client list.
Write one page of rules for your team covering what goes into a tool and what never does. It takes ten minutes with the AI policy template, and the judgment calls live in what your team can safely put into ChatGPT.
The line that catches agencies specifically is a pitch deck for a client who hasn't signed. It is unreleased work, it is somebody else's numbers, and it does not feel like either of those things at eleven at night.
Can you name what AI touched last week?
Here's the question that sorts a studio using these tools on purpose from one where it is just happening.
For last week's deliverables, can you say which ones a tool drafted, and who read each one before it went to the client? Not roughly. By name.
If that is hard, the problem isn't your tools. It is the kind of gap that hides inside a small shop full of good people, because everybody assumes somebody checked.
The free AI readiness check is twenty questions, one tap each, about four minutes. Your finding and all five areas land on screen when you finish. For an agency the interesting two are checking and standardization, because that is where the margin goes missing.
How we run ours
We run our own delivery with AI, including the four gates above. So this is how we work rather than a theory about how you should.
Three routes if you want help. We train your team on it. We build the check into your delivery so it stops depending on who is in that day. Or we run parts of production for you.
Which one fits depends on whether you are short on skill, on process, or on hands. The honest version of that question covers the case where the answer is to hire nobody.