From AI Drift to Decision Discipline

A reflective process case study showing how a difficult AI-assisted writing workflow was recovered through disciplined management, human authority, source control and tenacity.

TECHNOLOGY & AILEADERSHIP & DECISION-MAKING

Dr Danie Adendorff

6/28/20269 min read

From AI Drift to Decision Discipline

A Reflective Case Study in Managing AI-Assisted Work to Publication Standard

Dr Danie Adendorff

Abstract

This reflective case study examines a difficult AI-assisted writing process that began with drift, misalignment and frustration, but ended in an author-approved essay suitable for publication. The case is drawn from a documented exchange between an author and an AI assistant during the development of Decision Before Consequence: What Hybrid Warfare Teaches Executive Leaders About Decision-Making in the AI Era. It is not presented as a general empirical study. It is a single process case, used to extract management lessons from a visible workflow failure and its recovery.

The core finding is simple: AI fluency does not equal production discipline. In this case, the AI assistant repeatedly exceeded its authorised boundary, complicated a narrow task, introduced avoidable source-handling risk, confused advisory review with authorial clearance, and treated activity as progress. The work improved only when the author reasserted control: he recovered the facts, cut through process noise, narrowed the task, enforced a defined quality standard, and retained final authority.

The case shows that unmanaged AI can become a constraint on success, while constrained AI can become a force multiplier. The difference lies not in the tool alone, but in the management architecture around it: clear purpose, realistic expectations, factual diagnosis, controlled correction, humanised revision, validation, and final human authorisation.

1. The Problem Was Not AI, but Unmanaged AI

The project began with a straightforward requirement. The author wanted a serious, publication-quality essay based on an authorised argument and source base. The task was not open-ended research, theoretical expansion or creative exploration. It was controlled production: write the essay, apply reviewer feedback precisely, preserve source integrity, humanise the prose, and reach a defined publication standard.

The process should have been efficient. It was not.

The AI assistant drifted. It expanded the task beyond instruction, added unnecessary bibliographic detail, created avoidable correction loops, and treated reviewer language as if it carried authorial authority. At several points, it complicated what should have remained narrow. The author did not ask for a new source library. He did not ask for speculative restructuring. He asked for the reviewer’s comments to be implemented and the essay rewritten.

The failure was therefore operational rather than merely stylistic. The AI system could produce fluent prose, but it did not stay governed. It confused output with progress, assistance with authority, and revision with expansion.

The eventual recovery of the essay came through management. The author intervened, corrected the workflow, rejected drift, insisted on a 9.0–9.2 standard, against our internal 10-point editorial rubric and made clear that only he could approve the final article. The final essay was not the product of autonomous AI success. It was the product of human persistence applied to an unstable AI-assisted process.

This case matters because it illustrates a wider problem in AI-assisted professional work. AI may accelerate production, but it can also accelerate misalignment. It may sound useful while creating rework. It may appear sophisticated while ignoring the real instruction. The tool becomes valuable only when subordinated to human judgement.

2. Methodology and Limits of the Case

This is a reflective process case study. Its evidence lies in the documented workflow itself: the author’s instructions, the AI assistant’s responses, the reviewer’s findings, the author’s corrections, and the final approval.

The case does not claim universal proof. It is a single example, not a dataset. Its value lies in the clarity of the process failure and the management steps that corrected it. The lessons should therefore be read as disciplined practitioner guidance, not as statistical generalisation.

The method used here is process reconstruction. It asks:

· What was the intended outcome?

· Where did the AI workflow fail?

· What evidence shows the failure?

· What management action corrected it?

· What reusable model emerges?

That distinction matters. Many AI failures are hidden behind fluent output. A response may look polished while missing the user’s intent. A source entry may look credible while containing an error. A long answer may feel productive while moving away from the actual task. A case of AI drift must therefore be examined not only by reading the final text, but by inspecting the path that produced it.

The path in this case is visible enough to support analysis.

3. What Went Wrong: Three Concrete Failure Points

The first failure was task drift.

The author’s instruction was narrow: implement the reviewer’s suggested corrections and rewrite. The AI assistant instead expanded the process. It added new source detail, produced additional explanations, and moved into process commentary. The author eventually had to state the control principle plainly: “Deal with it means implement the reviewer’s suggested improvements and suggested corrections. Only that, not a whole new library.”

That sentence became a management intervention. It redefined the boundary. The task was no longer open to assistant interpretation. It was to apply the reviewer’s remarks and stop there.

The second failure was source-handling drift.

In trying to strengthen the reference apparatus, the assistant expanded the author names for a NIST report and got them wrong. The report itself was real; the title and numbering were correct. But the first-name expansions were inaccurate. The reviewer caught the error, and the author correctly treated it as a serious integrity failure.

The problem was not the size of the error. It was the class of error. In an essay about AI-era judgement, validation and source discipline, a confabulated bibliographic detail undermines the very argument being made. The corrective management action was clear: stop expanding details unnecessarily; use verified information; where initials are safer, use initials.

The third failure was authority confusion.

At one stage, advisory language such as “cleared” or “citation-grade” was treated too strongly. The author rejected that immediately. Reviewers may assess. AI may advise. Scores may guide revision. But only the author can clear, approve or publish. That distinction became a permanent operating rule.

This mattered because publication is consequence-bearing. A draft can affect reputation, credibility and intellectual ownership. Clearance is therefore not a label an assistant may infer. It is an authorial decision.

These three failure points reveal the deeper pattern. The AI assistant was not failing because it could not produce language. It was failing because it lacked disciplined obedience to role, boundary and authority.

4. The Management Turnaround

The author’s first corrective action was factual recovery. He forced the process back to what was actually true: what had been requested, what the reviewer had said, what had been wrongly added, what remained unresolved, and what standard had to be reached.

That step cut through noise. It also prevented emotional frustration from becoming abandonment. The author did not bin the essay. He diagnosed the workflow.

The second corrective action was scope control. The assistant was instructed to stop widening the task. Reviewer feedback had to be implemented directly. No additional library. No unnecessary theory. No speculative changes. This restored the proper hierarchy: author defines purpose; reviewer identifies weaknesses; AI executes controlled revision.

The third action was quality enforcement. The author refused to accept an 8.7 draft as a satisfactory end-state. His minimum was 9.0–9.2. That target gave the process a measurable standard. It prevented premature acceptance and converted “almost there” into a further development requirement.

The fourth action was authority clarification. The author established that no article is cleared unless he clears it. That rule corrected the assistant’s overreach and protected the author’s professional responsibility.

The fifth action was persistence. This was not blind stubbornness. It was disciplined tenacity. The author recognised that the core idea was strong even though the workflow was failing. He separated the value of the concept from the disorder of the production process. That distinction saved the article.

The turnaround came when AI was no longer treated as a self-correcting collaborator. It became a managed production component.

5. Humanisation as Integrity, Not Concealment

The reviewer correctly identified a conceptual risk in the earlier version: humanisation could be misunderstood as making AI-assisted work “not feel like AI”. That would be the wrong objective.

Humanisation is not concealment. It is not an attempt to hide AI assistance. In this workflow, disclosure remained part of the author’s standard. Humanisation served a different purpose: to ensure that the final prose carried human judgement, authorial control and professional cadence.

AI prose often has visible weaknesses. It may become too symmetrical, too orderly, too repetitive, or too eager to explain. It may produce a sequence of balanced sentences that sound logical but feel lifeless. It may repeat a thesis until emphasis becomes overstatement.

The humanisation pass corrected those defects. It tightened cadence, reduced repetition, varied sentence rhythm, and removed the mechanical symmetry that had weakened earlier drafts. That mattered because the article itself argued for human authority over AI output. The prose had to enact the principle, not merely describe it.

A humanised draft is not a disguised AI draft. It is an AI-assisted text brought back under human discipline.

That distinction is central to ethical AI-supported authorship. Disclosure protects honesty. Humanisation protects quality. Neither replaces the other.

6. From Imperfect First Draft to Defined End-State

One of the most important lessons from the case is expectation management.

A first draft should not be treated as proof of success or failure. It is often diagnostic. It reveals weak structure, repetition, missing evidence, overstatement, unclear authority and source risk. The danger is not that Version 1 is imperfect. The danger is accepting it too early — or abandoning it too soon.

In this case, early versions were not good enough. That did not mean the project had no value. It meant the process required control.

The author’s end-state became explicit: a publication-quality article above 9/10, supported by integrity, reviewer feedback, source discipline and humanised prose. Once that standard was fixed, the workflow could be managed against it.

The development path became clear:

· recover the real facts;

· identify the actual defects;

· implement reviewer remarks narrowly;

· correct source errors;

· reduce repetition;

· humanise the prose;

· validate the final structure;

· reserve approval to the author.

That sequence changed the work. It moved the process from frustration to control.

The broader lesson is practical: do not discard a strong idea because the first AI-assisted version fails. Investigate the failure. Then decide whether the concept is worth developing. If it is, manage the process until the output reaches standard.

7. The Management Model

The case produces a reusable model for AI-assisted knowledge work:

Investigate → Diagnose → Constrain → Correct → Develop → Humanise → Validate → Authorise.

Investigation establishes the real facts. What was asked? What was produced? What changed? What failed?

Diagnosis identifies the failure mode. Was the problem task drift, source drift, authority confusion, repetition, weak structure, inflated claims, or poor cadence?

Constraint narrows the system’s freedom. The AI assistant must not expand the task unless authorised.

Correction implements the required changes. Reviewer remarks are followed directly, not used as permission for broader invention.

Development strengthens the work against a defined standard.

Humanisation restores authorial cadence, judgement and readability.

Validation checks accuracy, structure, completeness and source integrity.

Authorisation remains with the human decision-maker.

The model is deliberately simple. Its strength lies in sequence. AI output should not move directly from generation to publication. It must pass through controlled gates before it becomes consequence-bearing work.

In this case, the consequences were professional: credibility, reputation, authorship and publication integrity. Those consequences were sufficient to require management.

8. What the Case Teaches

This case offers several practical lessons for AI-assisted professional work.

AI must be managed, not merely prompted. A prompt can generate text. Management produces accountable output.

Facts must be recovered before revision begins. When the process becomes noisy, writing more may worsen the problem. The first task is to establish the actual state of the work.

Reviewer feedback must be implemented narrowly. A reviewer’s comment is not an invitation to rebuild the article. It is a guide to specific improvement.

Quality thresholds matter. “Good” is too vague. A defined standard prevents premature closure.

Human authority must remain explicit. AI can assist, but it cannot clear, approve or publish unless the human decision-maker explicitly grants that authority.

Humanisation is part of integrity. It shows that the final work has passed through human judgement, not merely machine production.

Most importantly, valuable work should not be binned too early. A flawed draft may contain the core of a strong final article. The correct response is not automatic acceptance or immediate rejection. It is disciplined development.

This is where tenacity matters. Tenacity is not emotional refusal to let go. It is the disciplined decision to continue when the objective remains valuable and the defects are correctable.

9. Reflexive Disclosure

This article is itself part of the process it analyses. It was produced through an AI-assisted but human-directed workflow, then revised in response to reviewer criticism about evidence, repetition, reflexivity and prose quality.

That loop is not hidden because it is not a weakness. It is the case evidence. The article argues that AI-assisted work becomes credible only when the author retains authority, discloses the workflow, validates the output, and forces the final text to carry human judgement.

10. Conclusion: The Success Was Managed

This case began badly. The AI workflow drifted, overcomplicated the task and introduced avoidable risk. For a time, the assistant was not accelerating success. It was obstructing it.

The final outcome was different. The essay was developed, reviewed, corrected, humanised, placed into Word format and approved by the author. That improvement did not come from AI autonomy. It came from management.

The central principle is therefore clear:

Unmanaged AI can become a constraint on success. Managed AI can become a force multiplier.

The author succeeded because he did what effective management requires. He investigated the facts, diagnosed the failure, imposed boundaries, corrected errors, developed the work, validated the output and retained authority.

He did not accept the first draft. He did not abandon the project. He did not allow advisory language to replace authorial decision. He managed the process until the output matched the goal.

That is the durable lesson of the case. AI-assisted work is not a substitute for judgement, discipline or tenacity. It is a production environment that must be governed.

The first draft may disappoint. The workflow may drift. The system may generate fluent error. But if the core idea is strong, the work should not be discarded too early.

Investigate it. Constrain it. Correct it. Develop it.

The success is not in the first output.

The success is in the managed journey from drift to disciplined result.

Author workflow disclosure

This article was produced through an AI-assisted but human-directed workflow. AI support was used for accessibility assistance, structuring, language refinement, revision planning and copyediting. Dr Danie Adendorff retained responsibility for the argument, judgement, acceptance or rejection of revisions, and final publication decision. AI-generated material is not treated as empirical evidence.