AI Will Not Save Your Brief

For the last two years, I have watched a particular conversation happen, in roughly the same shape, with roughly the same people, in roughly forty different rooms.

A marketing leader sits down with their team. The team has been asked to use AI more. The team has tried. The output, the leader says, has been mixed. Some of it is good. Some of it is bad. The bad outweighs the good, on average. The leader has heard, from a vendor or a peer or a podcast, that the right model used the right way produces output indistinguishable from the work of a senior strategist. The team has not, in the leader's opinion, achieved this. Therefore the team is using AI wrong.

The leader is correct that the output is mixed. The leader is incorrect about why.

I want to make a case that has gotten lost in the last eighteen months of AI-related discourse, and the case is this. The model is not the bottleneck. It has not been the bottleneck for a year. It is unlikely to be the bottleneck again. The bottleneck is, and has been, the brief.

A brief is the document that explains what the work is supposed to do. It is what you give to a creative team before they start. It is what a copywriter or designer or strategist or, now, an LLM, reads in order to know what to make. The brief is supposed to contain, at minimum, four things. Who the work is for. What it needs to accomplish. What it needs to look or feel like. What constraints exist that cannot be ignored.

Almost no brief I have read in the last fifteen years contains all four. Most contain one and a half. Many contain zero. A surprising number contain four things, but they are not the right four things. They are, instead, the things the person writing the brief assumed everyone already knew, which is the worst possible foundation for a document whose entire purpose is to communicate.

Bad briefs predate AI. They predate the internet. They predate, I am willing to bet, the agency model. Bad briefs are not a technological problem. They are an organizational one.

Here is what a bad brief looks like.

The bad brief is a single page, typed, that contains the following information. The client's logo. A paragraph that begins with "We are a leading provider of..." A bullet list of audiences, each described in two words ("millennials," "decision-makers," "moms"). A list of "tone words" between five and fifteen items long. A budget, in a range. A deadline, often "ASAP." A line at the bottom that reads "needs to feel modern but timeless."

Every one of those elements is wrong. Each of them is wrong in a different way. Together they form a document that, when handed to a creative team or a model, produces work that looks roughly like the average of all marketing work currently in existence, which is approximately what the brief, if you read it carefully, is asking for.

This is not the model's fault. The model has done what the brief asked. The brief asked for the average. The model produced the average. The model, in fact, produced a slightly better version of the average than most humans would have, faster, for less money. Pointed, however, at a brief that asks for nothing in particular, even the best instrument will give you nothing in particular, expensively.

We built a tool that grades briefs.

I run a small agency. We have, in the last year, done a thing that I want to describe in some detail, because I think it explains the actual constraint on AI-assisted creative work better than any benchmark, white paper, or vendor demo could.

We built a tool that grades briefs.

The tool is internal. It is a wrapper around a few models, plus a rubric we wrote, plus a database of bad briefs we have collected over the years, plus a small piece of UI that lets a strategist paste a brief in and get a score back. The score is between 0 and 100. It comes with a list of the specific things the brief is missing. The most common entries on the list are these:

• No clearly named primary audience.
• No measurable objective.
• No definition of "modern."
• Tone words list contains a contradiction.
• Budget and deadline are inconsistent with each other.
• Brief assumes the reader already knows what the brand stands for.

The tool is at version 0.4. It is correct, by our reckoning, about 80% of the time. The other 20% of the time, it is angry. We are working on the angry part.

The tool's first job was to grade incoming briefs from clients. Its actual most useful job has turned out to be grading our internal briefs. The briefs we write to ourselves before we ask anyone, human or otherwise, to do work. These briefs were, on inspection, only slightly better than the bad briefs we had been complaining about for years. We had been blaming the inputs. The inputs were partly us.

Three things happened.

Once we started fixing the briefs, three things happened.

The first thing is that the work got better. This is unsurprising. A clearer instruction produces a clearer output. The work got better whether the work was done by a human, an LLM, or a team of both, which is the most useful information in this paragraph.

The second thing is that the AI output, specifically, got dramatically better. There is a lazy version of "use AI more" that means "give the same vague brief to the model that you used to give to the intern." The lazy version produces lazy output. The non-lazy version, which means writing a brief precise enough that a literal-minded reader could execute it, produces output that is, on a good day, indistinguishable from the work of a competent human, and on a bad day, still better than the average vendor pitch.

The third thing is that we found out which of our team members were good at writing briefs. This was, surprisingly, not a function of seniority. Some of our most senior strategists, it turned out, had been getting work done despite their briefs, not because of them. The more junior people, who had less authority and had to spell things out for whoever was actually doing the work, were often better at writing briefs than the people running the projects. This is not a moral judgment. It is a description of how the muscle gets developed. You write good briefs when you have to. Most senior people don't have to anymore.

I want to be careful not to make this sound like a diagnosis with a single cure. The brief is not the only thing wrong with most marketing work. The hiring is wrong. The review process is wrong. The metrics are wrong. The relationship between the agency and the client is, in many cases, structurally adversarial in a way that no document can repair. I am not saying that fixing the brief will fix the work. I am saying that not fixing the brief guarantees the work will not get fixed.

Garbage in, garbage out.

There is a phrase I encountered in software engineering that I have been thinking about a lot. The phrase is "garbage in, garbage out." It refers to the fact that no algorithm, however sophisticated, can produce a useful output from a useless input. The phrase is decades old. It is still true. AI has not changed it. AI has, if anything, made it more visible, because the speed at which the garbage cycles through has accelerated, and the garbage is, increasingly, in writing.

The vendors will tell you that the next model will fix this. The next model will not fix this. The model after that will not fix this. There is no model that fixes the inputs, because the inputs are not a model problem. The inputs are a thinking problem. The brief is the thinking. If the thinking has not been done, the brief reflects that, and the model produces work that reflects that, faster.

This is, I think, the actual lesson of the last two years, and it is much less exciting than the lesson the vendors are selling. The lesson is that AI is a magnifier. It magnifies clarity into output that is fast and good. It magnifies vagueness into output that is fast and bad. The vendors do not want to talk about this because the vagueness is not their problem. They are selling models. The vagueness is your problem. It is, in some sense, the only problem you have left.

Five practical notes.

A few practical notes, if you are a marketing leader trying to actually do something about this.

One. Take a brief you wrote in the last month and read it as if you have never met your company before. Note every place you assumed something the reader could not know.

Two. Ask the people on your team who do the work which briefs they liked, in the last six months, and which they had to fight through. The list will surprise you. The good briefs are not always from the senior people, and the bad ones are not always from the junior ones. There is an asymmetry in this data and the asymmetry is the data.

Three. Stop calling the document a "creative brief." Call it a "what we want." It is harder to write a vague "what we want" than a vague creative brief. The genre conventions of the creative brief allow for a kind of professional hand-waving that the phrase "what we want" does not. Try this. It is annoyingly effective.

Four. Build the tool. Or, if you don't want to build the tool, write the rubric. The rubric is the thing. The rubric is a list of the questions every brief must answer before it goes to the team. The rubric will probably contain seven items. You will know it is the right rubric when looking at your existing briefs against the rubric makes you wince.

Five. Stop blaming the model.

The model is fine. The model has been fine for a while. The model will be more than fine soon. The output your team is producing is the output your team's instructions deserve. If the output is bad, you have two choices. You can write better instructions. Or you can keep blaming the model and watching the budget compound.

The brief has been the problem for sixty years. AI has not made it worse. AI has only made it more obvious that it was always the problem. We can, with effort, do something about it.

Or we can buy a better model.

The model will not save us.