Why AI drafts still need heavy editing, and how to need less

AI drafts need heavy editing mainly because of what goes into them, not because models write badly. Most revision time is spent restoring voice, audience specificity and accuracy that the brief never supplied. Teams reduce editing by fixing inputs: stronger briefs grounded in audience evidence, a codified voice system, curated examples and a clear edit checklist.

Is heavy editing a model problem?

In 2020, yes, partly. GPT-3 outputs wandered off topic, repeated phrases and confidently stated things that were not true. Every piece we produced that summer went through a careful human edit, and some were rewritten almost entirely. Models have improved enormously since, especially after instruction-tuned chat models became mainstream with ChatGPT in November 2022.

Yet our listening on AI integration in marketing keeps surfacing the same complaint: AI content that needs so much human revision that the time savings disappear. If models are much better, why does this persist? Because the remaining problems are mostly not model problems.

Where does the editing time actually go?

When we review editing logs with teams, edits tend to fall into four buckets:

• Accuracy: fixing facts, claims, product details.
• Voice: removing generic phrasing, restoring brand rhythm.
• Structure: reordering, cutting padding, sharpening the point.
• Audience fit: replacing abstractions with what customers actually care about.

Accuracy edits are essential and will always remain. The other three are largely self-inflicted. They come from briefs that said too little, so the model defaulted, and the editor rebuilt the piece by hand.

How do you move effort upstream?

1. Put facts in, do not ask for them

Give the model the product details, claims you can substantiate and the sources. Ask it to write from those, not to recall them. Accuracy edits shrink when the model has no need to guess.

2. Supply the voice

A codified brand voice system with examples removes much of the voice editing. The editor checks rather than rewrites.

3. Supply the audience

This is the bucket most teams miss. A brief that includes how the audience describes the problem, the tension behind the decision and the objection they raise produces a draft that already sounds relevant. SOMIN, an AI audience-research platform and our technology partner, is how we source that evidence; the SOMIN for agencies overview explains how teams fold it into briefs.

4. Define the shape

Specify length, structure and the single point the piece must make. Structure edits drop sharply when the model knows what the piece is for.

What should the human edit focus on?

Once inputs improve, editors can spend their time where humans are irreplaceable: judgment. Is this the right thing to say now? Would a customer find this respectful? Does it overclaim? Is there a sharper idea hiding in the second paragraph? That is the editing that makes content good, not merely acceptable.

The goal is not zero editing. It is editing spent on judgment rather than repair.

A worked example

Picture a team producing weekly newsletter intros, where editors spend most of their time replacing generic openings and adding customer context. They change the brief template to include three audience phrases from that week's listening, one fact sheet and two example intros. Editors can then spend their time deciding which angle to lead with, not rewriting sentences. We are not quoting a figure because yours will differ; the point is to measure your own before and after.

Checklist: the low-edit brief

• One clear point the piece must make
• Verified facts and sources, pasted in
• Voice system reference or condensed rules
• Two relevant examples
• Audience phrases and tension, from real conversations
• Format, length and channel
• Named editor and edit categories to log

Definitions

• Human in the loop: a workflow where a person reviews and approves model output before it is used.
• Edit log: a lightweight record of edits by category, used to improve inputs.
• Grounding: giving a model verified source material to write from.

Measuring what lands

Less editing is only half the result. The other half is whether content performs. Pairing edit logs with engagement data shows which inputs matter most. If you are also running paid distribution, our sister brand Trafix looks at true profitability of creative rather than vanity ROAS, a useful discipline for content teams too. The KPI Media case study is worth reading for how agencies connect insight with execution.

Heavy editing is a symptom. Treat the inputs and the symptom eases, leaving your best people free to do the work only they can do.

How do you start an edit log?

Keep it simple enough that editors actually use it. A shared sheet with five columns works: piece, time spent, main edit category, a one-line note, and whether the fix could have been supplied in the brief. Ask editors to fill it in for a month. At the end, sort by category and by the last column.

Patterns appear quickly. If voice edits dominate, your voice system is missing or too vague. If audience-fit edits dominate, briefs lack evidence of how customers talk. If accuracy edits dominate, the model is being asked to recall rather than to write from supplied facts. Each pattern points to one upstream change, and you can test that change on the next month's log.

The log also changes how editors see their own work. Instead of feeling like they are cleaning up after a machine, they become the people who improve the system, which is a more satisfying and more valuable role.

Frequently asked questions

Why do AI drafts need so much revision?

Usually because the brief lacked audience evidence, voice rules and examples. The model filled gaps with defaults, and editors then spent time putting back what the brief left out.

Should AI drafts ever be published without editing?

For brand content, no. A human should check accuracy, voice and judgment on every piece that represents the brand. Editing can become lighter, but it should not disappear.

How do you measure editing effort?

Track time per piece and classify each edit as accuracy, voice, structure or audience fit. The dominant category shows which upstream input to fix first.