AI-Assisted Documentation Pipeline

The problem

Writers were spending their best hours on the lowest-judgment work in the job. Long, unstructured source material required manual drafting, restructuring, and formatting before it could ship as review-ready documentation, and output quality varied run to run, so consistency depended on individual discipline rather than on the system. Source material arrived in many forms:

• Training decks.
• Transcripts.
• SME notes.
• Existing articles.
• Image-only files.
• Table-heavy PDFs.
• Code samples.

Downstream, writers and SMEs spent review time correcting structure and format, and filling gaps where a draft missed or misread source content, before they could even validate technical accuracy. That review time is the team’s scarcest resource, spent on rework instead of judgment.

Business impact and value delivered

The pipeline's value is measured drafting capacity, and a quality level the system enforces rather than hopes for.

• Capacity without headcount. Team drafting reached 93 drafts submitted for SME review in its peak week, against a 40-per-week baseline, with the same four writers. Peak week, drafting stage only, from tracker data.
• Solo leverage, separate provenance. In the Claude Cowork experiment, one operator turned 22 source documents into 63 QA-validated articles in one week. A personal experiment, not team production, and never merged with the team figure above.
• Review time protected. SME review time is the scarcest resource in the pipeline. Reviewers now receive pre-validated drafts and review for technical accuracy only, and inline SME flags mark exactly what needs verification, so reviewer attention lands where doubt actually exists instead of being spread across the whole document.
• Consistency by system, not discipline. A fixed output contract and deterministic rendering make every draft arrive in the same structure. Draft quality no longer varies with who ran the pipeline or how careful they were that day.
• Traceability as risk reduction. Audit logs trace any output to the stage and inputs that produced it, so a defect is diagnosable in minutes rather than reconstructed from memory.

Claim limit: The throughput figure is a peak week at the drafting stage. Nothing downstream of drafting, and no sustained rate, is claimed.

Constraints

• Variable source. Source material varied wildly in structure, cleanliness, and format.
• Fixed template. Output had to follow a fixed internal template, not free-form generation.
• No silent invention. Missing, conflicting, or uncertain source information had to surface as an SME flag, never get silently invented.
• Two-stage QA. QA needed draft QA and render/DOCX QA, not one pass at the end.
• License ceiling. The Copilot workspace ran on a limited license (initially free-tier, no Copilot Studio access), which ruled out a fully automated agent-to-agent chain regardless of design preference.
• Traceable runs. Every run needed a traceable log.

The approach

Copilot pipeline: Two agents, one human checkpoint

KM Drafter handles the drafting stage:

• Takes in source material.
• Identifies distinct user goals.
• Proposes a parent/child document structure.
• Converts long source into task-focused Markdown drafts against a fixed output contract.
• Flags gaps and conflicts inline rather than inventing content.

A human reviews that draft. KM Format Editor then takes the approved Markdown and generates the DOCX.

The Format Editor applies the template, heading levels, procedures, tables, fonts, and styles against named styles, not chosen by the model.

(Chat output can’t carry visual formatting and can’t touch a Word document directly. That’s the whole reason the Format Editor exists as its own stage.)

Claude Cowork pipeline: Automated run, human gates preserved

A 13-stage file-based pipeline (intake validation → source inventory → extraction → draft planning → parent/child decomposition → draft generation → draft QA → revision → DOCX generation → render QA → logging → packaging → archiving) runs with exception-based review.

Two things stop the run for a human decision:

• Two consecutive failures on one file.
• A source spanning multiple topics.

Author review and SME validation still gate publication outside the run. Automating the draft doesn’t touch those checkpoints.

Key decisions

Results

Design intent: The system was built for writers, to move their attention off transcription and formatting and onto judgment. Throughput is the measurable outcome, not the goal that shaped the design. SME flagging was added on top of that, marking what actually needs verification so neither the writer nor the reviewer carries the whole document at the same level of suspicion.

Not claimed: Reduced backlog age, faster onboarding, improved review turnaround, or “publication-ready” status for the Claude pipeline’s output (QA-validated is the accurate term).

What changed after testing

• Split the overloaded agent. KM Drafter and KM Format Editor replaced it after the stochastic rule-dropping failure.
• Moved static lookup data out of instructions. Label sets, templates, and table specs now live in reference material the agent consults.
• Removed styling from the agent. Visual styling left the agent’s job entirely, and code now applies it deterministically in both pipelines.
• Kept the human checkpoint. The handoff between drafting and formatting stayed manual on Copilot rather than automated, given the license constraint and the undetected-error pilot.
• Built exception-based stops. Repeated failure and multi-topic source halt the Claude Cowork run, instead of a checkpoint at every stage.

Limitations and next step

Limits on the Claude Cowork result

• A sprint, not a rate. It is a personal, one-week build-and-run sprint, not a steady-state weekly rate.
• No matched comparison. It hasn’t been compared against the Copilot pipeline under matched conditions.
• No formal cost figure. Operating cost (token use) for the Claude pipeline wasn’t formally measured.

Next step

A controlled team trial of the Claude Cowork pipeline, measuring editing time, SME correction rate, cost, and throughput against the Copilot baseline.

Supporting artifacts