Even if AI creates a draft in minutes, publishing can still be slow. Track the route your content actually takes for two weeks, and the work you can hand to AI—and the bottlenecks you need to remove first—become much clearer.
Measure your team’s leakage, not the 60–70% figure
The claim that “60–70% of content team time is spent on non-creative work” should not be treated as an industry-wide benchmark. The seed article says that work requiring less creative judgment—such as consolidating research, formatting outlines, generating metadata, suggesting internal links, and creating title variations—accounts for 60–70% of team time, but it does not provide a research sample or measurement method. So this number is closer to a hypothesis to test in your own team than a reason to buy.
Still, broader research on knowledge work suggests the overall direction is hard to ignore. Based on its Anatomy of Work Index survey of more than 10,000 knowledge workers, Asana says that “work about work,” such as work communication, information searching, switching tools, prioritization, and status checking, takes up 60% of working hours. In a content-industry survey, 41% of B2B marketers cited workflow and approval issues, while 39% cited difficulty accessing internal subject-matter experts. Another 31% said they did not have a structured content creation process.
First distinction: work time versus waiting time
If an editor reviews a draft for 40 minutes, that is work time. If it sits in the review queue for two days, that is waiting time. Even if AI cuts 20 minutes from the review draft, publishing speed will barely change if the two-day approval wait remains.
That is why the unit of an audit should not be “the time people were busy,” but the full time it takes one piece of content to move from idea to publication. A value stream map shows both the actual work time at each step and the waiting time between steps, revealing rework, unnecessary approvals, and duplicated work. With this perspective, you may find that the impression that “writing is slow” actually comes from missing briefs, waiting for expert responses, or unavailable approvers.
AI candidates are not “work that takes a long time,” but “repeatable work that is easy to reverse”
Do not hand work to AI just because it takes a lot of time. It is safer to start with steps that happen frequently, have clear inputs and acceptance criteria, and are easy to undo when something goes wrong. By contrast, people should make the final call on steps that require context and accountability, such as brand perspective, strategic angle, fact-checking, and legal or regulatory judgment, even if AI helps with a draft.
| Work type | Good for a first experiment | Keep human approval |
|---|---|---|
| Research | Collecting source candidates, sorting materials, removing duplicates | Checking original sources, verifying figures and causal claims |
| Production | Formatting outlines, creating title variations, adapting formats for each channel | Core claims, brand perspective, choosing examples |
| Editing | Checking spelling, prohibited terms, and missing links or metadata | Nuance, audience fit, final wording |
| Review | Checking against a checklist, confirming citation format | Assessing factual, copyright, legal, and reputational risk |
| Publishing | Entering approved copy into the CMS and scheduling it | Approval to publish, sensitive edits, crisis response |
This boundary matters because generative AI can confidently produce plausible but incorrect facts or citations. NIST defines this as “confabulation” and explains that factual errors or internal inconsistencies can arise especially in long-form, open-ended tasks and fields that require specialized context. In other words, “the writing sounds natural” cannot be a review pass criterion.
A real comparative study also shows why speed and quality need to be assessed separately. In a rapid literature review case published by the UK government, the AI-assisted approach completed the overall work 23% faster and cut the analysis and synthesis stage for selected studies by 56%. But the initial draft was less fluent than the human-only version and needed more revisions. The researchers also stated that it was a single case that could not be generalized and that manual verification was necessary.
This is also why you should select small units within writing rather than delegate all writing at once. In Anthropic usage data, work related to copywriters and editors showed a particularly high share of “task iteration,” where people and AI exchange outputs and refine them. Translation-related work, by contrast, more often ended with minimal intervention after an instruction. Even within the same content role, the automation boundary differs by step.
Filter out approval bottlenecks that AI cannot fix first
The step with the longest waiting time is not necessarily the step where you should add AI. If a draft sits in expert review for three days, first separate whether that is because the review itself is difficult or because the owner saw the request late. In the latter case, what you need is a clear owner, deadline, notification, and backup approver—not a generative model.
A documented content workflow separates request, drafting, review, approval, and handoff stages, making the owner of each stage visible. That lets a project manager see where content is getting stuck. Value stream mapping likewise recommends documenting responsible roles, information movement, handoffs, work time, and waiting time together.
You can handle audit results in three directions. Standardize templates and storage locations for steps where people repeatedly look for inputs; eliminate simple handoffs and status checks with rules-based automation; and test generative AI only for steps that summarize unstructured material or suggest multiple phrasings. If you cover an unclear approver with a prompt, or add another unnecessary review step as an AI review, the tool itself becomes a new queue.
Do not choose a tool during the audit.
If you define the problem around a specific product’s features from the first week, only the steps that product does well will look large. Log the steps, time, and errors first, then choose a tool for the one highest-scoring step after the second week ends.
Run a two-week content workflow audit like this
Day 1: Choose 10–15 representative content pieces
Select ordinary work from content types you create often, such as blog posts, newsletters, and social posts. Set the start as “brief requested” and the end as “actually published.” If you select only urgent campaigns or unusually complex projects, your baseline will be distorted. The seed article also recommends a parallel test with 10–15 representative pieces when comparing the existing approach with an AI approach.
Week 1: Log the actual flow one line at a time
Create columns in a spreadsheet for content ID, step name, owner, start and end time, actual work minutes, waiting minutes, tools used, input materials, outputs, number of revisions, and reason for delay. Do not reconstruct it from memories in a meeting; record it as work happens. Keep normally hidden work—such as time spent chasing someone on Slack and waiting for approval—in separate rows too.
Days 8–9: Classify every step into four categories
Label each step as one of “eliminate, rules-based automation, AI assistance, or keep human.” For AI candidates, score repeatability, work time, clarity of criteria, and ease of reversal from 1 to 5, then deduct points for factual, legal, and brand risk. Mark steps with long waiting time but little actual work time as operational bottlenecks, not AI candidates.
Days 10–12: Run a parallel test on only one high-scoring step
For example, hand only “extracting metadata and internal-link candidates from the source text” to AI and keep the rest of the flow unchanged. Apply the same inputs and acceptance criteria to the existing method and the AI-assisted method. Keep every prompt, model, output, and human edit so you can reproduce the result.
Days 13–14: Decide on total cost, not just time saved
Compare the full lead time for each piece of content, human work time, number of revisions, factual errors, brand edits, and approval rejections together. Adopt only steps where work time falls while quality criteria hold. If errors increase or review time consumes the savings, pause the candidate rather than getting stuck endlessly fixing the prompt.
The final output of the audit is not a massive AI transformation plan. A one-page map that assigns ownership by step—such as “AI assistance for metadata drafts,” “rules-based automation for approval reminders,” and “keep people responsible for core claims and fact-checking”—is enough. With that map, when you watch the next tool demo, you can ask not about the number of features but whether it will actually reduce your bottleneck.
If you want to dig deeper
Identifying Which Content Workflows Your AI Tool Should Actually Handle — The seed article for mapping content-production steps and designing AI-human handoffs and parallel tests. publishpoint.io
How work about work gets in the way of real work — A resource for understanding time spent outside skilled work on tasks such as status checks, information searches, and switching tools. asana.com
B2B Content and Marketing Trends: Outlook for 2024 [Research] — Provides figures on approval, expert access, and structured-process issues in content teams. contentmarketinginstitute.com
Value stream mapping with Confluence — Explains how to separate work time from waiting time and map handoffs and bottlenecks. atlassian.com
Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST material covering factual errors in generative AI and the need for validation and monitoring. nist.gov
AI-assisted vs human-only evidence review — A comparative case showing both the stages where AI assistance saved time and those that required extra revisions and manual verification. gov.uk
.png)
