The AI draft is not the article

After months away from proper article writing, I used three language models on the same live assignment. They disagreed with one another, I disliked the first draft, and the unfinished article became more useful than any of the finished-looking versions. What emerged was a method for getting past the blank page without giving away the work that makes writing worth reading.

I am writing an article again for the first time in months. Not a LinkedIn post or a working note assembled from an idea I already understand, but a commissioned article with a brief, an audience, a word count and a publication standard to meet.

The subject sounds simple: how should a freelancer ask a friend to pay an overdue invoice without damaging the friendship? It is the sort of useful, apparently narrow question that can expose whether an article has any editorial intelligence behind it. A generic answer is easy. Send a polite reminder. Be clear. Set boundaries. Escalate if necessary. Every search result and every language model can produce some version of that list.

The assignment was for Sage Advice, aimed specifically at creative freelancers. It needed to acknowledge a real tension: the same informality that makes working with a friend feel easy can make asking for payment unusually difficult. It also needed practical language readers could use, reliable guidance on what they could do next, and a reason to keep reading beyond the headline.

I had a detailed brief. I had source material about the hidden administrative hours created by late payment. I also had an upcoming SMB Advisory Board conversation that could supply the lived experience and exact language the article still lacked.

So I did the obvious thing. I gave the brief to Microsoft Copilot and asked it to write the article.

I did not like what came back.

The brief worked. The article did not

The Copilot draft was not a disaster. That was almost the problem. It followed the structure, covered the requested points and sounded broadly appropriate for a business-advice publication. If the test was whether it had processed the brief, it passed.

If the test was whether somebody would choose to keep reading, it was much weaker.

The prose was competent but generic. The advice moved predictably from friendly reminder to firmer follow-up. The creative-freelancer audience was named more often than it was understood. There was little sense of the specific emotional calculation involved when the unpaid client is also somebody whose birthday you remember, whose messages normally sit outside your working day and whose friendship you do not want to turn into a collections process.

The piece fulfilled the document without finding much of an article inside it.

That distinction matters. A detailed brief can make AI output look more finished because the model has more boxes to tick. It can also conceal the absence of an editorial proposition. Structure, tone and coverage are not the same as tension, evidence or point of view.

My first reaction was that the draft had failed. My more useful reaction was to ask what kind of object it actually was.

It was not an article. It was a structured surface I could think against.

A bad final draft can be a good scaffold

The Copilot version made the assignment visible. Instead of facing a blank page, I could see where the obvious advice accumulated, where two sections repeated one another and where the brief’s promise of creative-freelancer specificity had not reached the prose.

It also gave me something concrete to dislike. That sounds flippant, but editorial judgement often becomes clearer in opposition. “This is too generic” is only the start. The useful questions are more exact: Which paragraph could apply to any late invoice? Where has politeness been confused with vagueness? What truth about friendship and money is the article avoiding? Which claim needs reporting rather than smoother phrasing?

Once I treated the generated copy as scaffolding, its limitations stopped being embarrassing. They became diagnostic.

The core idea began to sharpen: when money is overdue, clarity can be kinder than softness. A vague message intended to protect the friendship can prolong uncertainty for both people. A clear message can separate the commercial problem from the personal relationship.

That was more useful than another list of reminder templates because it could govern the whole article. It affected the opening, the sequence of advice, the suggested language and the definition of a good outcome. The article was no longer merely about chasing payment. It was about using clarity to stop an unpaid invoice from quietly reorganising a friendship.

The model had not found that article for me. Its draft had helped me see what was missing.

Then the models disagreed

I asked Claude to act as an editorial reviewer, not as another first-draft machine. It challenged the generic passages, the lack of creative-freelancer texture, the repetitive rhythm and the places where the article stated its thesis without sustaining it. It then produced a rewrite that was livelier, more specific and more committed to the idea that clarity can be kinder than softness.

That version was better. It was not simply right.

When I asked Codex to review Claude’s rewrite, it agreed with some of those decisions and pushed back on others. The stronger voice had occasionally become overstatement. A direct answer required by the brief had moved too far down the article. One practical question had disappeared. Some of the planned openings for Advisory Board evidence had been flattened or lost.

The three models had very different opinions of the same piece. That was useful precisely because no model was entitled to win.

This was not a scientific comparison of Copilot, Claude and Codex. They did not receive identical prompts or perform identical roles. Copilot was asked to write directly from the brief. Claude was asked to challenge and rewrite that result. Codex was asked to review the rewrite against the brief and the reporting plan. Any claim that one model had objectively beaten another would tell us more about the test design than the models.

What I got instead was editorial triangulation. Each response made a different set of choices visible. Copilot exposed what compliance without distinction looks like. Claude demonstrated how much a clearer thesis and sharper texture could improve the reading experience. Codex showed where a persuasive rewrite had traded away accuracy, scope or planned evidence.

I have argued before that a second model should not be used as a proofreader. This experience made the point more concrete. Disagreement is not noise to eliminate. It is a way of surfacing decisions that a fluent first draft can hide.

The human job is not to average the answers. It is to decide.

Search the evidence you already have

My first version of the method moved directly from challenge to new sourcing. That skipped an important layer. An organisation may already possess customer stories, interview transcripts, approved case studies or examples that could answer the evidence gap before another interview is commissioned.

At Sage, Microsoft Copilot can provide a discovery pass across the internal material it is permitted to access. The useful question is not simply “find me a customer story”. It is more exact: is there an approved customer example that demonstrates this behaviour, complicates this claim or gives the reader language grounded in experience?

That makes customer evidence a gate. A strong, relevant example can be traced to its original source and taken through any required approval. A promising lead can be followed up with its owner. If nothing adequate exists, the gap remains visible for the Advisory Board or a new interview. If customer evidence would add decoration rather than understanding, the article should proceed without forcing it in.

Copilot can surface a lead; its summary is not the evidence. The original record, context, accuracy, recency, attribution and permission still have to be checked. This gate is where the customer intelligence workflow enters the writing workflow: search organisational memory first, then report what is genuinely missing.

The gaps are part of the design

There was another reason not to treat any generated version as finished: the most valuable material did not exist yet.

The article needed a specific opening experience, a real example of the language somebody had used with a friend, and an approved quotation or prevention insight from the SMB Advisory Board. Those were not decorative additions to be pasted into polished copy. They were evidence that might change the emphasis, the examples or even the thesis.

So the working draft kept explicit spaces for them.

This is more than leaving a note that says “add quote”. It is evidence architecture. Each placeholder has a job. One must earn the opening. One must demonstrate how people actually speak when friendship and payment collide. Another must test whether the proposed advice survives contact with experience. A final contribution should help readers prevent the same ambiguity in their next project.

Those spaces also impose useful discipline on AI. A language model is exceptionally good at making an empty section look occupied. It can invent a plausible freelancer, a plausible message and a plausible emotional resolution before anybody notices that plausibility is not evidence.

Leaving the gap visible protects the reporting. It says: this part of the article has not been earned yet.

The source conversation should then be allowed to alter the draft. If Tuesday’s discussion produces a sharper account of why friends delay payment, the opening may need to change. If a contributor’s actual message is blunter or warmer than the template, the practical advice should respond. If the board challenges the “clarity is kinder” proposition, the article should not hide the complication merely because the current structure is elegant.

Evidence is not colour applied after the thinking. It is one of the things that does the thinking.

Writing over the scaffold

This is the stage I do not want to automate away.

Once the reporting is in, I will not ask a model to “humanise” the draft. I will write over it. Some sections may survive in outline; individual sentences probably should not assume they have tenure. I will choose the detail that deserves the opening, decide how long to stay with the discomfort, cut advice that has become redundant and make the rhythm sound like somebody thinking rather than a system completing sections.

That difference can be easy to miss. Line-editing an AI draft leaves its underlying decisions largely intact. Writing over a scaffold means the author can change the hierarchy, argument and emotional movement of the piece. The generated draft provides resistance, memory and coverage. It does not provide authority.

For me, this is particularly valuable after time away from long-form writing. The blank page carries an exaggerated demand: recover the argument, structure, voice and confidence at once. A scaffold lowers the cost of beginning. It gives me something to interrogate while leaving the consequential choices open.

The risk, of course, is that “scaffold” becomes a flattering name for copy I barely change. The test is not whether I touched every sentence. It is whether the final piece contains reporting, judgement and language that could not have appeared before I did the work.

The method that emerged

I have started calling this the Authored Scaffold. The name is provisional, but the sequence is clear enough to test again.

1. Brief. Define the reader, problem, promise, constraints and evidence requirement. A brief should not only specify sections. It should state what the article must help somebody understand or do.

2. Scaffold. Generate a complete working version quickly. Its purpose is to expose the obvious structure, not to impersonate finished authorship. Mark unsupported claims and absent evidence rather than allowing fluent filler to conceal them.

3. Challenge. Give a model, editor or colleague a specific adversarial role. Ask what is generic, what is missing, where the thesis fails, what the audience would distrust and what the draft has misunderstood. A second model is optional; a second editorial position is not.

4. Customer Evidence. Search the organisation’s existing customer knowledge for an example that could materially strengthen the article. Treat Copilot as a retrieval tool, not a source: trace every lead to the original material and check relevance, permission and approval. If no suitable evidence exists, preserve the gap for new reporting.

5. Source. Report into the gaps that existing evidence could not fill. Gather lived experience, exact language, evidence and counterexamples. Let the new material change the argument instead of treating it as decoration.

6. Author. Write over the scaffold. Make the decisions about emphasis, sequence, voice, omission and meaning. This is where a serviceable content object becomes a piece somebody might recognise as yours.

7. Verify. Check the final article against the brief, sources and publication requirements. Confirm quotations, legal or financial claims, links, scope, promised questions and the accuracy of every practical step.

These are not seven steps for the sake of having a seven-step process. They are editorial value gates. The Brief protects relevance; the Scaffold protects coverage without pretending that coverage is quality; Challenge protects distinction; Customer Evidence protects organisational memory from being overlooked; Source protects original reporting; Author protects judgement and voice; Verify protects trust.

This is why the method feels close to my customer intelligence workflow. In both cases, I have put enough stages between input and output to stop information being merely collected and summarised. Signals must be interpreted, challenged and applied. AI can compress production, but the workflow must preserve the points at which editorial value is created.

The sequence is not revolutionary. Editors have always separated commissioning, drafting, reporting, rewriting and fact-checking. The change is that generative AI can compress the structural work so aggressively that it becomes tempting to mistake the compressed output for the whole job.

Authored Scaffold is an attempt to keep the speed while restoring the separations that protect quality.

More process is not automatically better

There is an obvious objection. One article has now involved a brief, three language models, source research, a planned Advisory Board session and a named seven-stage method. It could look like process theatre: a great deal of machinery for telling somebody to send a clear payment reminder.

That objection is worth taking seriously. Not every paragraph needs a panel of artificial critics. A short, low-risk piece may only need a brief, one scaffold and a careful human edit. A strong writer with command of the subject may be slowed down by producing a generated version first. Multiple models can create the illusion of rigour while multiplying similar, derivative suggestions.

The method should therefore be proportional to the work. Its value is highest when a piece must balance several things at once: a detailed commission, a specific audience, factual or legal boundaries, original evidence and a distinctive editorial argument. If the workflow adds more ceremony than judgement, it has failed.

AI does, however, change how much of an evidence-led process is practical. It does not generate the evidence, but it can reduce the work around it: exposing unsupported areas, turning gaps into focused reporting questions, searching material the organisation already holds, comparing contributions, mapping a transcript back to particular claims and preserving unused evidence for a later version.

I am now testing that across four commissioned articles rather than one. Each scaffold exposes a different evidence gap, but the reporting does not have to restart from zero each time. A targeted question to a practitioner can answer a defined uncertainty. One Advisory Board conversation may strengthen several pieces. AI can help route the transcript back to the relevant drafts, while tax guidance still goes through separate technical review. The process only remains efficient if those boundaries stop every article becoming its own research programme.

There is also an AI visibility reason for working this way. In AI-mediated discovery, competent generic advice is abundant and interchangeable. If an article restates the same consensus available everywhere else, an AI system has little reason to retrieve or cite it rather than another source. Specific practitioner experience, attributable evidence, concrete examples and a defensible interpretation make the work less replaceable.

That does not guarantee GEO performance. Retrieval and citation depend on authority, accessibility, technical structure and the wider information environment as well as editorial quality. The narrower claim is that GEO begins with creating knowledge worth finding. Authored Scaffold may make that level of evidence and distinctiveness achievable within an ordinary editorial assignment instead of reserving it for a major research project.

The article is becoming an input

There is another reason I think the evidence layer matters.

I recently read Eli Schwartz’s argument that SEO and AEO can create a brand problem when claims made for search visibility escape their original context and reappear in AI-generated answers. I would be cautious about the neatest version of the technical story. Training data, live retrieval, search ranking and citation selection are different mechanisms, and a published claim does not simply enter a model and become permanent truth.

But the organisational problem is real enough without that simplification.

For most of the web era, a page was mainly an endpoint. Somebody searched, clicked and read it. A claim written for an SEO page could live at some distance from the language used by PR, brand or investor relations.

Generative search makes that separation harder to rely on. Published material can become an input into systems constructing answers somewhere else.

That changes the question from:

Can we rank for this topic?

to something closer to:

What evidence are we putting into the information environment from which answers about this topic are constructed?

That is a claim-governance problem as much as a search problem.

The chain I increasingly care about looks like this:

Evidence → editorial scaffold → defensible claims → content → AI/search ecosystem

The scaffold is useful partly because it creates a control layer between evidence and publication. It gives me somewhere to ask what supports a claim, what qualification it needs, whether the organisation already has stronger evidence and where reporting still has to happen before fluent prose makes the gap disappear.

This is also why I think the most interesting effect of AI on writing is not cheaper words.

AI can make evidence work cheaper too.

A long interview can be interrogated against several drafts. Existing customer stories can be searched for relevant proof. Contributions can be compared. Unsupported claims can be surfaced. A conference transcript can be routed back to the questions it actually helps answer. Evidence gathered for one article can remain available for the next instead of disappearing into a folder.

That makes a more rigorous editorial process possible without turning every article into a research project.

And it matters for GEO because generic production is now extremely cheap. If ten companies ask language models to answer the same question from broadly the same public information, there is little reason to expect ten genuinely distinctive pieces of knowledge.

Original research, practitioner experience, customer outcomes, expert knowledge and first-hand reporting give the article something the model could not simply generate from the same common pool. That is why I am becoming less interested in the GEO question “How do we make this article visible to AI?” The more useful question may be “What have we produced here that is worth an AI system finding?”

That is a much harder standard. It is also a healthier one.

Nor does disagreement among models prove that the final human decision is good. It only widens the field of visible choices. The editor still needs taste, subject knowledge, accountability and the willingness to reject attractive language that the evidence cannot support.

Where this sits in Editorial Intelligence

I do not yet think Authored Scaffold is a framework. A framework should provide a durable model for thinking across contexts. This is currently a method: a repeatable way of moving a substantial piece of writing from commission to publication while keeping human evidence and judgement in control.

The Sage Advice article is its first documented application. It is also an AI workflow demonstration, because the point is not which interface produced the best paragraph. The point is how several tools, sources and editorial roles were arranged to produce a better decision.

The public account belongs here, among the working notes, while the method is still being tested. If it survives other subjects, formats and writers, it can become a formal method within the Workflow Library. If it only works when I already know what I want to say, that limitation needs to become part of the method too.

This distinction matters to Editorial Intelligence. The useful unit is not the prompt or the model output. It is the designed relationship between brief, evidence, challenge, judgement and memory. The article is one deliverable. The reusable process created around it may be the more durable piece of intellectual property.

But only if the process continues to produce writing worth reading.

That is the test I still have to pass. I have the bones of the article, several competing editorial readings and deliberately empty spaces waiting for real experience. Next comes the Advisory Board conversation. Then I have to sit down and write.

The AI draft is not the article. It is the point at which the article becomes visible enough for the author to begin.