AI Content Cost Savings: 5x Output, Same Review

Most arguments about AI content cost savings happen at the wrong end of the invoice. Somebody holds the licence fee up against a writer's day rate, decides the numbers look implausible, and stops there. The licence fee is the least interesting number in the whole exercise, and AI content cost savings don't come from it.

Here's the claim, stated plainly and defended after. The drafting cost collapses, and the review cost you already had doesn't rise to meet it. That gap is the saving. It's large, it's real, and it's available to any team willing to write its standard down.

The Saving Is Real, and It Is Not the Licence Fee

The published figures cluster tightly enough to quote. Independent write-ups this year put agency production between five hundred and eight hundred dollars an article, against fifty to a hundred and fifty for AI-assisted production, with turnaround dropping by roughly an order of magnitude.

But take those at a discount, because a good share of them get published by companies selling the tooling. Halve the claim and it's still the biggest single change to editorial economics in my working life, and I've been doing this twenty-eight years.

The reason it isn't the licence fee is arithmetic. A seat costs tens of pounds a month. A writer costs tens of thousands a year. Any saving worth discussing lives in the second number, which means it lives in how many people a given volume needs, and that's the number AI actually moves.

Where the 89% Comes From, and What to Discount

A figure that round deserves a going-over before it lands in a business case.

The baseline is an agency rate, and agency rates carry account management, margin and a project manager. Holding that up against a marginal production cost isn't like-for-like, and a fair version of the number would put the agency baseline against an in-house one instead.

Do that and the saving shrinks, because an in-house writer was never costing you eight hundred dollars a page. It drops to something nearer the time figures, which are the more defensible half of the data. Two thirds of production time, or a little over four hours a piece.

The Number I Would Actually Quote

Two thirds of production time, sustained, with the review step unchanged. That's smaller than eighty-nine per cent and it's the one I can defend in a room with a finance director, which makes it worth a good deal more than the bigger number.

It also compounds in a way the per-article figure hides. Time saved doesn't get banked, it gets redeployed, and a team that puts it into volume gets the ranking effect on top of the cost one. That second effect is what clients actually notice.

What Actually Collapses

The First Draft, Almost Entirely

Drafting used to be the visible part of the job, the bit with a person sitting at it looking busy, and it ate most of a writer's week. It's now close to free and close to instant.

Most of the mechanical saving sits right here. A brief that took a competent writer most of a day to turn into a publishable first pass now takes minutes, and the pass is better than what a tired writer produces on a Friday.

Research, Structuring and the Second Draft

Less discussed and nearly as valuable. Gathering sources, building an outline, rebuilding it when the outline turns out wrong, producing three variants so an editor can pick. Hours once, minutes now.

The four hours a piece the industry figures report isn't one big saving. It's this, piling up across a dozen small steps nobody used to time.

The Review Was Always There

This is where most honest analyses of AI content cost savings go wrong, and I've made the mistake myself in print. The argument runs like this. AI generates more, so review load rises, so the saving is smaller than it looks.

It sounds rigorous and it's wrong, and the error is a simple one. The review step is not a cost AI introduced. Every content operation worth the name already had a human reading the output before it went live. That was true in 1998 and it's true now, and it isn't a new line on the invoice.

And It Does Not Scale Linearly

The second half of the error is assuming review time rises in step with output. It doesn't, because most of what review used to catch was mechanical. The missing disclosure. The wrong figure. The banned phrase. The heading that breaks the house pattern.

Those checks are a file now. They run on every piece in seconds, without anyone remembering to run them, so the human read spends its attention on judgement rather than proofreading. Judgement is a much smaller share of the work than proofreading ever was.

So the shape isn't "more output, proportionally more review". It's more output against a review step that was already budgeted, already staffed, and now doing less per piece than it used to.

The saving is not that the human goes away. It is that the same human, doing the same job they were already doing, now covers five times the output.

What the Skill File Does

The verification most people picture doing by hand is a file, and that file is what does the heavy lifting. Worth being concrete about what's in it, because "we have a process" is what everybody says.

On this site the standard runs to just under eleven thousand words across four plain text files, and a 129-line script checks the output against them before anything gets called finished. It counts banned phrases, sentence-length variance, reading ease, heading hierarchy, link integrity and schema validity, and it fails the piece rather than warning about it.

What It Catches Without a Human

Fifty-eight banned words and phrases. Fourteen structural patterns named, ten of them caught by the script. Em dashes, bracket asides, whole-sentence bolds. Broken internal links, unparsable schema, a title that's drifted out of sync with its heading. Every one of those used to be somebody's afternoon.

What It Deliberately Does Not Catch

Whether the argument is any good. Whether a comparison misleads by leaving something out. Whether a sentence is technically accurate and practically deceptive. Those need a person, they always needed a person, and no part of my pitch pretends otherwise.

The Arithmetic on a Real Team

Take a content team of one editor and three writers producing forty pieces a month, which is a shape I've seen plenty of times.

An illustrative model of a shape I have worked with repeatedly, not a measured case study. Your ratios will differ.

Notice which row didn't change. The editor is still there, still reading, still the last line of defence. What changed is that the two writer salaries went either into headcount reduction or, in every engagement I'd recommend, into five times the content for the same money.

That second option is the one clients take once they see it, because a content team's problem is rarely that it costs too much. It's that it can't produce enough to compete.

What Clients Ask For Instead of a Discount

When a team sees the saving, what they ask for next is almost never "reduce the budget".

The first ask is coverage. Every content team has a list of pages it knows it should have and has never had the capacity to build. The long tail, the comparison pages, the questions real customers ask that nobody has answered. That list tends to be years old and everyone on the team can recite the top of it from memory.

Refresh Is the One They Underrate

Published pages decay, and most operations have no capacity to revisit anything once it ships. A team producing five times the volume can go back over what it published two years ago, and on a mature site that back catalogue is worth more than anything new.

It's the least glamorous line item and the one I'd spend the saving on first. Refreshing a page that already ranks on page two is a shorter route to traffic than publishing a new one that starts at nothing, and no team without spare capacity ever gets to it.

Then Speed to Market

Getting a page live the week a topic breaks rather than the month after. That's the difference between owning a query and turning up to find somebody else already does, and you can't buy it back later at any price.

None of those three is a cost saving on the face of it. All three are what the cost saving buys, and they're why this conversation runs easier than a procurement pitch normally does.

How to Measure It in Your Own Team

Don't take my figures or anybody else's. This is measurable in your own operation, the measurement is straightforward, and it takes one month of ordinary work.

1. Count pieces per editor per month, right now, before anything changes. This single ratio is the whole measurement, and most teams have never written it down.
2. Time the review step honestly for two weeks. Separate the mechanical part, meaning checks a script could do, from the judgement part. The split comes out more lopsided than anyone expects.
3. Write the standard. One page minimum. Whatever your editor checks for that is not judgement goes in it.
4. Automate the mechanical half and run it on everything. Do not change the drafting yet. This step on its own moves the ratio.
5. Then change the drafting, and re-count at the end of the month.

Doing it in that order matters. Teams that change the drafting first and the checking later get exactly the volume-without-verification failure the sceptics warn about, then decide the tooling was the problem rather than the order they did things in.

What It Looks Like Over a Year

Month one is the worst month, and anyone selling you this should say so. The standard gets written, the checks get built, and output barely moves while the team learns to brief rather than draft.

By month three the ratio has moved and the argument settles itself, because pieces per editor is sitting on a dashboard and nobody needs convincing with a slide. That's when the conversation stops being about tooling.

The Second Half of the Year Is the Interesting Part

Volume isn't the compounding effect. The compounding effect is that a team with spare capacity starts doing the work it could never justify. The refresh programme. The internal linking. The pages that serve fifty searches a month and convert at four times the rate of the ones serving five thousand.

On the client behind the figures in the box, the rewritten silos lifted impressions fiftyfold and gained an average of thirty ranking positions. Not because any single page was brilliant, but because the estate finally got attention nobody had ever had the hours for.

The Objection I Take Seriously

One version of the sceptical case survives everything above, and it deserves a straight answer rather than a rebuttal.

It runs like this. Automated checks catch what you thought to check for. The failures that actually damage a business are the ones nobody saw coming, and a team producing five times the volume has five times the exposure to that kind. Scripts don't help there, and the human read is thinner per page than it used to be.

That's correct, and it's why the human read is non-negotiable rather than optional.

Why I Still Think the Trade Is Worth Taking

My answer isn't that the risk goes away. It's that the risk was already there at lower volume, that the mechanical checks free up the human attention that catches the unforeseen kind, and that a written standard makes each new failure cheap to encode so it never happens twice.

Every incident becomes a line in the file. That's a system that gets safer with age, which is the opposite of what happens to a style guide nobody reads, and it's the part of the argument I'd defend hardest.

What I Recommend When I Leave

An engagement that ends with nobody able to run what I built is a failure, however good the numbers looked in month three. So the handover is part of the work.

What I recommend, every time: a real human keeps reading the output and checking anything that sounds off. Not every word of every piece, but a genuine editorial read by somebody who knows the subject and is allowed to stop things. That was always the arrangement, and it doesn't change because the drafting changed.

Legal and Regulated Copy Gets More

Where a page carries legal consequence, the final human read isn't a recommendation, it's a requirement, and I put that in writing before an engagement starts. Contracts, regulatory disclosures, anything a lawyer would want to see. Those get read by a person every time, no exceptions, and the saving on those pages is smaller because it should be.

That isn't a caveat that undermines the pitch. It's the reason the pitch is credible on the other ninety per cent of the estate.

And Some Clients Keep Me On

A few days a month, for the judgement calls and to keep the standard current. Some prefer that to building the capability internally, and for a small team it comes in under the cost of a hire. It's an option rather than a condition, and plenty of clients run it themselves after handover.

What I won't do is make the retainer load-bearing. A system that only works while I'm attached to it is a dependency dressed up as a service, and the standard sitting in four plain text files the client owns is the deliberate answer to that.

Where the Saving Does Not Apply

Three situations where I'd tell you the numbers above don't hold, because a pitch that applies everywhere applies nowhere.

• Very low volume. Under about ten pieces a month the setup cost of writing the standard exceeds the saving. Keep the freelancer.
• Nobody will write the standard. The saving is downstream of a written rulebook. No rulebook, no automated checking, no saving, and you are back to a human reading everything at higher volume, which is the failure mode the sceptics correctly warn about.
• Original reporting. If the value of the piece is an interview nobody else got, the drafting was never the expensive part and there is little to collapse.

What It Costs to Set Up

A saving quoted without its cost is marketing rather than analysis, so here are the three things you have to buy.

A subscription, which is the trivial one. The time to write the standard, which is the real one and runs to days rather than weeks for most teams. And the discipline to keep it current, which is the one that quietly fails at most organisations about six months in.

Against a saving measured in writer salaries, all three are small. But they aren't nothing, and anyone quoting you an 89% reduction without mentioning them is selling rather than advising.

Questions I Get Asked

Are AI Content Cost Savings Actually Measurable?

Yes, and it's simpler than most people expect. Count pieces published per editor per month before and after. That one ratio captures the whole effect and it's hard to argue with, because it needs no assumptions about token prices or time-per-task.

Does the Quality Drop?

Not if the standard is written down and enforced, and measurably yes if it isn't. On the client behind the figures in the box, the rewritten silos lifted search impressions fiftyfold and gained an average of thirty ranking positions, which isn't what a quality drop looks like.

Will I Need Fewer Writers?

Fewer people producing first drafts, the same number exercising judgement. Most teams I work with redeploy rather than reduce, because the constraint was never too many writers. It was not being able to publish fast enough to compete.

What About Google Penalising AI Content?

Google's own published position is that it rewards helpfulness rather than authorship, and its helpful-content guidance says the same. What gets punished is thin, unchecked, undifferentiated work, which describes a bad process rather than a tool. Governance is what keeps you the right side of that line, and it's why the saving and your search performance aren't in tension.

How Long Before It Pays Back?

For a team publishing forty pieces a month, weeks. The standard takes days to write, the checking is automated from the first piece, and the volume increase starts immediately. The slow part is organisational, not technical.

Which Tool Should We Use?

Less important than the standard, which is the point of the comparison I wrote separately. The saving comes from governed volume, and you can get that on any of the major tools if somebody writes the rules down.

Do We Have to Disclose That AI Was Involved?

No general rule requires it and most publishers don't. I disclose here because the disclosure is the argument, which isn't most people's situation. Where it does matter is regulated copy and anything carrying a byline that implies personal testimony, and that's a question for your legal team rather than your content team.

What Happens When the Standard Is Wrong?

It will be, repeatedly, and that's the system working. A rule that turns out wrong gets corrected in one file and every piece after it inherits the correction, which is the opposite of how corrections travel through a team of freelancers. The standard here has been revised more than a dozen times, twice because a rule imported from another project was damaging the voice.

Is This Just Cost-Cutting With Extra Steps?

It's cost-cutting only if you take the saving as cash. Most teams take it as capacity instead, which is a different decision with a different outcome. Same money, several times the output, and a back catalogue that finally gets maintained. The ones who take the cash tend to be the ones who were over-resourced to begin with.

How Do I Know These Figures Are Real?

The three industry numbers are labelled as market shape and came from independent sources that agree with each other, which is the most anyone can honestly claim for a statistic they didn't generate themselves. The three client numbers are mine and measured. The team model says on its face that it's illustrative. Anything I couldn't verify was left out rather than estimated.

Can a Small Team Do This Without a Consultant?

Yes, and a fair number do. The standard is the hard part, and it's hard because it demands decisions rather than expertise. What your copy must always do, what it must never do, and who gets to say. A team that can answer those three can build this without me. What I sell is mostly the time and the pattern, not a secret.

The Short Version

The drafting cost collapses. The review cost you already had stays roughly where it was, because it was never the thing AI made harder, and the mechanical half of it is a file that runs in seconds now. Everything between those two facts is the saving.

A human still reads the output, and on legal work a human reads all of it. That was always true, before any of this existed, and nobody sensible is proposing to change it. What changed is that the same team, doing the same job, now ships five times as much, and the rules that used to live in a document nobody read fail a build instead.

That's the whole of the offer. Not a machine that replaces the editor, and not a discount on the salary line. A drafting cost that falls close to zero, a checking step that stopped being manual, and an editor who finally has the hours for the part of the job that needed them all along.