Do not index
"How much of our delivery should we actually be handing to AI?" Every agency owner is asking some version of that right now, usually while looking at a margin line. Automate the operation, never the output. The moment AI touches the judgment your client is paying for, you have converted a differentiated service into a commodity and you will find out at renewal.
Forrester's The State Of AI Inside US Marketing Agencies 2026, released with the 4As, puts numbers on how far the industry has already gone the other way. Nine in ten US agencies use generative AI. Half now use agentic AI for execution. And 61% still treat AI as a cost of business rather than a capability that changes what they can sell. "AI has fundamentally transformed marketing agencies, but the industry is at risk of mistaking efficiency for effectiveness," said Jay Pattisall, VP and principal analyst at Forrester. That single sentence describes most of the AI implementations I have looked at in the last year.
This is written for agency owners between $200k and $2M in revenue running teams of 3 to 12 people, on retainers between $5k and $30k per month, where the client is buying a specific point of view rather than a volume of assets. It applies to ghostwriting shops, positioning consultancies, and founder-content operations. Skip this if you sell content by the unit at $200 to $500 a post, because at that price your client is buying throughput and automating the output is the correct business decision. If you are still competing on turnaround time rather than on the quality of the thinking, this article will not change your model.
The Judgment Line
What I call the Judgment Line is the boundary between mechanics and judgment, drawn deliberately rather than discovered by accident. Mechanics are everything that would be identical no matter who ran it. Pulling transcripts, formatting drafts, checking a post against a style guide, scheduling, tagging, building the reporting pack, chasing approvals, moving a piece from draft to review to scheduled. That entire layer should be automated aggressively, and agentic tooling is genuinely good at it now.
Judgment is the part that changes based on who is doing it. What the client actually believes. Which of their three stories is worth telling this quarter. Which sentence in the draft is the real one and which four are throat-clearing around it. What to cut. What position to take that competitors will not take. That layer stays human, and not for sentimental reasons. It stays human because it is the only part a client cannot get anywhere else, and it is therefore the only thing holding your rate.
The Forrester barriers list tells you the industry knows something is wrong but is diagnosing it as a tooling problem. Agencies cited accuracy and bias at 63%, legal concerns at 62%, lack of AI agent expertise at 54%, and data infrastructure gaps at 51%. Every one of those is a mechanics problem. None of them is the actual risk, which is that a client reads three months of output, cannot find a single sentence that sounds like a person who has done the work, and concludes the retainer is replaceable. That churn does not announce itself. It shows up as a non-renewal with a vague reason attached.
Where the line pays for itself
The version of this that works is unglamorous. In a content operation, the agent runs intake, pulls the source material, drafts a structural skeleton, and flags anything that violates the client's known positions. A human writes the argument. A human decides what the piece is for. Then the agent handles distribution, versioning, scheduling, and the reporting. You end up automating maybe 60% of the hours and none of the value.
Draw the line in the wrong place and the economics invert. Automating the output lowers your cost per deliverable, which feels like margin until your client notices the deliverables are interchangeable and prices you accordingly. The efficiency gain gets competed away within two quarters because every other agency bought the same tooling in the same month. What you traded for it was the one thing that was not commoditized. This is the same failure mode that shows up in delivery quality generally, and a documented quality control system is what catches it before the retainer ends rather than after.
There is a straightforward way to audit where your own line sits. Take last month's deliverables and mark every step that a competitor with the same tool stack could have produced identically. If that covers more than 70% of the finished work, you are selling execution, and execution pricing is going to one place. If the marked portion is mechanics only and the argument in each piece is unmistakably yours, you have built the thing the Forrester report says most of the industry is giving away.
The strategic read is that AI has split the agency market into two businesses that look similar from the outside. One sells volume and will compete on cost until the cost approaches the tooling. The other sells judgment, uses automation to remove everything that is not judgment, and gets more expensive as the cheap version floods the market. Both are viable. They are not the same business and they cannot be run with the same cost structure. What determines which one you are running twelve months from now is not the tooling you buy this quarter. It is where you decide the line goes before the efficiency math makes the decision for you.
