The real bottleneck in AI writing isn’t the model

I’ve spent the past few years helping teams use AI to write faster, clearer, and at scale. Almost all of them start with the same complaint: the outputs feel off. AI has made writing almost free, yet somehow more expensive in time and frustration. Instead of freeing us from drudgery, it’s flooded us with endless “AI drafts” nobody wants to read.

The problem isn’t the model.

The problem is our collective inability to explain what good writing actually means.

From instinct to standard

Most teams still rely on intuition: I’ll know a good piece of content when I see it. That logic collapses with AI. A model can’t intuit your taste; it can only follow instructions. And when those instructions are vague, the output multiplies that vagueness.

Thanks for reading Cowritten.ai! Subscribe for free to receive new posts and support my work.

AI forces tacit knowledge - the instincts of good communicators - to become explicit standards. That transition is uncomfortable. I’ve seen teams hesitate because naming what “good” looks like feels pedantic or slow. Yet it’s the only way forward.

You need clarity criteria, tone expectations, information hierarchy, and even examples of what failure looks like. Each has to be concrete enough to define, test, and verify.

If that feels like too much work, ask whether the alternative (aka endless clean-up of half-useful drafts) is really cheaper.

The specification bottleneck

The new constraint is clarity, not speed. Every uncertainty in a brief or prompt expands during generation. Ambiguity compounds faster than you think.

The organisations producing the best AI-assisted writing aren’t always the ones with the best writers. They’re the ones that can define quality precisely. I often tell teams to treat writing like product design: create requirements, test cases, and acceptance criteria. Writing becomes a system of intent rather than intuition.

Evaluation as leverage

The second bottleneck is evaluation. Everyone can now generate, but few can judge. No one has time to read every draft, which means weak writing spreads faster than sound reasoning.

To fix this, evaluation must scale too. That doesn’t mean replacing editors with AI; it means giving AI a role in quality control. You can build small “critics” that perform a first-pass review against defined standards. The human eye still matters, but the machine helps separate signal from noise.

In practice, this shift makes editing less about catching typos and more about teaching judgement. The outcome isn’t faster writing but faster learning.

Information architecture is business logic

Most business writing hides weak thinking behind tidy prose. AI strips away that disguise. When you give a model a template without context (no goal, no decision logic, no sense of the audience) it fills the blanks without understanding their purpose. The result feels polished but hollow.

Templates only work when they carry reasoning as well as formatting. That reasoning is the real architecture of a business.

When I work with teams, I ask a simple question: what outcome should this piece of writing enable? If they can’t answer, that’s the problem, not the prompt.

Learning from failure

One of the simplest ways to improve AI writing is to define what bad looks like.

Collect examples of over-specified technical docs, vague strategy memos, and hype-heavy press releases. Annotate what went wrong. These “failure sets” reveal far more about an organisation’s blind spots than success stories ever could.

Good prompting begins with knowing what you no longer want to tolerate. It’s a kind of editorial therapy; naming your past mistakes so the system doesn’t repeat them.

The problem of convergence

As AI writing becomes routine, voices start to blend. Most models default to a safe, polite tone that removes both conviction and vulnerability. It reads smoothly but says very little.

I’ve seen this happen in minutes: a room full of distinct writers feeding prompts into the same model, all producing the same neutral voice. It’s efficient, but it erases perspective. Good writing has range. It distinguishes between confidence and uncertainty, between a decision and a question. Preserving that range is now an act of design.

Intent as the new cost centre

When the cost of words falls to zero, intent becomes the only real expense.

The organisations that thrive won’t be those producing the most content, but those defining the clearest purpose behind it. They’ll specify, test, and evaluate meaning as deliberately as they once managed budgets or code.

Machines will keep getting better at writing.

The real question is whether we’ll get better at deciding what’s worth writing.

Because the challenge of this new writing economy isn’t generation. It’s intent. And intent is the one thing no model can automate.

If you’re exploring how intent can guide AI in your work, reach out through cowritten.ai.