The real cost of an AI article starts after the first draft

The token bill is the easiest part of an AI article to count. That doesn't make it the cost of the article.

The draft took four minutes. Excellent. How long did someone spend checking its sources, fixing the product claims, removing the filler, and getting the page published?

And what happened to the drafts that never made it?

AI can make the whole process better. But counting generation while leaving out the work around it is a very convenient way to prove that it did.

The useful number is the cost of an article you're willing to stand behind. Here's a worksheet for finding it without borrowing a vendor's savings claim.

Decide what counts as finished

A generated document, an accepted draft, and a published article are different outputs. Pick one before dividing anything.

For this worksheet, “finished” means an article that meets the agreed editorial standard, reaches its intended public destination, and passes a live-page check. Keep the approval and URL with the record.

The FinOps Foundation's unit-economics guidance distinguishes technical-resource costs from business outcomes. Cost per token tells you about a model bill. Cost per accepted publication tells you about a content operation.

You can track both. Just don't swap them halfway through the pitch.

Hold the editorial bar still

Write down what the article needs to do: answer a useful question, support material claims, use current product information, sound like your company, and give the reader something worth their time.

Google's helpful-content guidance includes sourcing, expertise, and factual reliability among its assessment questions. Those are useful checks, not a promise of rankings.

Publishing also takes work. Titles, descriptions, images, links, and body formatting need to survive the trip into the site. Google's generative-AI content guidance includes metadata in its quality considerations.

Our content QA architecture and pre-publish gate cover those checks. Count the effort. A workflow that stops before them hasn't finished the same job.

Follow a group of attempts all the way through

For a first calculation, use a closed group of comparable article attempts. Follow each until publication or rejection. That avoids some of the confusion created when this month's effort produces next month's pages.

The underlying calculation resembles GOV.UK's cost-per-transaction measure: total service costs divided by completed transactions.

Here, rejected drafts stay in the cost. They don't enter the completed-publication count. Throwing away the draft doesn't refund the review time.

A calendar-month view is useful too, but it answers a different question. Include that month's costs, count that month's publications, and show the work pending at the start and end. Don't quietly combine a cohort's costs with a different month's output.

If nothing gets published, cost per publication is undefined. Report the hours, money, and unfinished work. Zero dollars per article would be a spectacularly misleading reward for producing nothing.

The worksheet: ten attempts, eight publications

This is an invented example in US dollars, not a Deadwater result or an industry benchmark. Ten comparable attempts move through an illustrative 30-day period. Eight are accepted and published; two are rejected. Nothing is pending at either end.

All ten attempts' effort is included below. Staff time is valued at an illustrative $75 an hour.

The team previously spent $1,200 setting up the workflow. It allocates that across three cohorts at $400 each. That's a planning choice for this example, not an accounting prescription or a claim about the system's useful life.

Cost component	Assumption	Included value
Earlier setup allocation	$1,200 across three cohorts	$400
Research, input preparation, and run setup	8 hours × $75	$600
Metered model and tool usage	Distinct usage charges	$40
Workflow subscription allocation	This cohort's share	$60
First-pass editorial review	6 hours × $75	$450
Corrective rework and rechecks	4 hours × $75	$300
Publishing and live-page checks	2 hours × $75	$150
Maintenance completed within the period	2 hours × $75	$150
Total allocated workflow cost	22 staff hours plus tools and setup	$2,150

The arithmetic is simple:

Staff-time value = (8 + 6 + 4 + 2 + 2) × $75 = $1,650
Allocated workflow cost = $1,650 + $40 + $60 + $400 = $2,150
Cost per accepted, published article = $2,150 ÷ 8 = $268.75

The $40 generation line is real. So are the 22 staff hours around it.

Count each piece once

There isn't an extra row for rejected drafts because their effort is already in research, review, and rework. Tag it by article if you want to investigate rejection costs. Don't add it twice.

The same applies to corrective rechecks: they're in rework here, so they don't also go into first-pass review. Human prompting and run management are in preparation.

Your categories can differ. They need to cover the actual work without overlapping. If a designer or specialist participates, their time doesn't disappear because the worksheet started with a model call.

Shared costs need an explanation

A subscription used by five workflows shouldn't be charged in full to all five. It also shouldn't vanish from all five because allocating it is annoying.

FinOps allocation guidance describes fixed, proportional, and proxy-based approaches. Use a basis you can explain. Record the full bill, the portion assigned here, and where the rest goes.

Check what's bundled and what's separate. For example, Claude's pricing documentation covers token charges and additional charges for certain server-side tools. Reconcile the categories relevant to your setup.

The broader workflow software buying decision includes more than the visible generation charge. A precise-looking total built from incomplete bills is still incomplete.

Capacity isn't the same as cash

The $2,150 total is an allocated operating-cost estimate. It isn't $2,150 of new cash paid this month.

It contains $1,650 of valued staff time, $100 in usage and subscription allocation, and $400 of earlier setup allocation. Don't add the full setup payment again. Don't add payroll on top of hours already valued under your chosen labor convention.

If the workflow saves an editor two hours, you may still pay the same salary. The benefit could be two hours available for interviews or a difficult article. That's useful capacity; it doesn't need to dress up as a cash saving.

The maintenance row covers work completed during this period. Future repairs create future costs. A lifetime forecast needs its own horizon and assumptions.

Repeated prompt repairs and manual rescues are part of the brittleness tax. Track them even when fixing one article also improves the shared system. Reusable work still took somebody's afternoon.

The cheap model may not be the expensive problem

Now change one assumption at a time. Every unmentioned input stays fixed in this sensitivity exercise; real-world changes may affect several inputs together.

Change to the example	Revised calculation	Cost per publication
Metered usage doubles from $40 to $80	$2,190 ÷ 8	$273.75
Review and rework take four more hours	$2,450 ÷ 8	$306.25
Only six articles reach accepted publication	$2,150 ÷ 6	$358.33
No articles reach accepted publication	$2,150 ÷ 0	Undefined

For this mix, four extra review hours matter much more than doubling usage spend. This team should probably investigate review before spending a week shopping for cheaper tokens.

Your mix may be different. Extensive research calls, image generation, or repeated failed runs can change the balance. The point of the worksheet is to find your expensive problem.

GOV.UK's service-benefits guidance recommends establishing a baseline and scrutinizing assumptions that could change the decision. Keep counts and topic difficulty beside the average, especially for small batches.

One difficult article can move the number considerably. So can postponing a release into next month. Neither automatically means the workflow got worse.

Compare the same job

The 2023 Noy and Zhang working paper studied short professional writing tasks and noted limits to generalizing beyond self-contained work without company-specific context.

That's evidence about a particular setting. It doesn't price your sourcing, approvals, product corrections, and publishing process.

Compare workflows using similar topics, source requirements, and acceptance standards. Give both comparable inputs. Count research, drafting, correction, review, and publishing on both sides.

An AI demonstration with a perfect source pack versus the team's hardest historical article is a sales comparison. It isn't a useful operating baseline.

Inspect the finished page too. Anthropic's evaluation guidance separates an agent's success claim from the outcome it produced. “Published successfully” deserves a visit to the URL.

Keep elapsed time separate from active effort. Meeting a launch deadline and reducing editor hours can both matter. They are different benefits.

Spend the next hour where the ledger points

A small record is enough to start: article, stage, person, minutes, charge, correction reason, status, and publication date. Keep shared allocations alongside it.

Look for repetition:

• The same unsupported claim keeps appearing: fix the authoritative input and its check.
• Metadata keeps needing repairs: check it before the editor sees the draft.
• Every piece waits for one unavailable approver: address the responsibility or capacity problem.
• The prose keeps coming back bland: inspect the writing process and examples, not just the prompt's adjective count.

Keep the failed example and the acceptable correction. Our guide to incorporating editorial feedback explains how that correction can become maintained context.

Then see whether the problem recurs in later, comparable work. Count the cost of maintaining the fix too. A rule that saves one edit and creates three exceptions needs another pass.

Cost per publication still won't tell you whether the article helped a buyer or supported a sale. Track those outcomes separately. Cheaper production of work nobody needs isn't much of a victory.

Bring the last few articles—and the work it took to finish them—if you want help mapping your workflow. The interesting number usually lives somewhere after “generate.”