The On It StudioOur AI Content Workflow, With a Human Approval Step
Our AI content workflow with human approval has four gates: code checks, a separate editor pass, a fact check on the social copy, and a person who approves before anything posts. It publishes far less than it drafts. In the logged runs from 2 to 4 October 2026, the writer tried 12 topics and published 1. This is how The On It Studio runs its own blog and social channels, with the real counts as of 4 October 2026, and what we would set up the same way for a client.
What does the workflow do, step by step?
A scheduled job drafts a guide, code and a second AI pass try to reject it, and a person approves anything that goes to social media. Each gate can stop the post on its own.
| Step | What happens | Who or what | What stops it |
|---|---|---|---|
| 1. Sources | A weekly scan collects reports and keeps a number only if it appears on the cited page | Scheduled job | Number not found on the page |
| 2. Draft | Three times a day, the writer picks a topic by type and reader and drafts from the sources and our field notes | AI | Nothing yet |
| 3. Code checks | Word count, required links, banned phrases, no external links outside the sources list | Code | Any rule broken |
| 4. Editor pass | A separate AI pass scores six things; it needs at least 4 on each and a 4.3 average | AI | Below the bar after two rewrites, then skipped |
| 5. Social copy | A carousel and captions are made from the published guide | AI | Fact check finds an unsupported line |
| 6. Approval | A reviewer reads the slides and captions on our admin page | Person | Not approved, or sent back with a note |
| 7. Posting | Approved items go to Facebook, Instagram and LinkedIn | Code | Page or image not live yet; max 3 posts per channel per day |
The sources scan is the quiet one. In its first run it saved 17 items and dropped 2 of the stats it had found, because the number was not on the page it claimed to come from.
What did the checks reject in the first week?
Most drafts. The point of the checks is to say no, and the counts show it.
From 2 to 4 October, the writer started 8 times and tried 12 topics. One became a published guide. The other 11 were skipped. The editor pass ran 17 times and passed once. Code checks failed 12 times: 8 drafts were outside the word range, 2 did not link enough related guides, 1 used a banned phrase and 1 had a file format error.
The word-count failures tell us something about our own rules. All 8 were opinion pieces, which have to land between 600 and 900 words, and the drafts kept coming in between 974 and 1,166. Either the drafts pad, or the range is too tight. We have not changed the rule yet. When a check fails the same way eight times, look at the check before you blame the writer.
On 2 October we also re-ran a stricter check over everything already published. The writer had drafted 7 guides since 29 September. Three are live today; the re-check hid the other 4. Of the 46 posts carried over from our old blog, 1 is still public.
The lesson we wrote down after the first week: the editor pass and the rules in code matter more than the drafting. Most of the work is saying no to weak drafts.
What broke: a real number with the wrong owner
The most useful catch was not an invented number. It was a real number credited to the wrong source.
Our guide on a first-time exhibitor's Toy Fair booth gives a rough range for floor space at a regional show, stated as an estimate. The first version of slide 2 in its Instagram carousel said that range came from "our Toy Fair 2026 log". It read well. It sounded like proof. It was wrong, because the guide never says those were our booth costs.
The fact check flagged it, the slide was regenerated without that line, and the reviewer approved a clean version. A person skimming the slide would likely have let it through, because every number on it was in the guide. The error was in who the number belonged to.
If you add one check to an AI copy workflow, make it this one: not only whether a fact is true, but whose fact the copy says it is.
Why a person still approves every social post
Because public posts are hard to take back, and the reviewer is the only gate that knows the business.
As of 4 October, 8 guides have social posts ready. Two were approved by a reviewer on 4 October and went out to Facebook, Instagram and LinkedIn, six posts in all. Five are waiting in review. One is held for manual handling.
The trade-off is speed. Posts wait for a person, and some wait days. A digest email lists what is waiting, and a reminder goes out if something sits for more than 48 hours. The poster also refuses to run until the guide's page and every image return a live response, so nothing links to a page that is not there yet.
The same pattern runs our audit form and job search
The approval step is not only for content. We use the same shape, AI drafts and a person sends, everywhere something leaves the building.
The free ops audit form saves each request to our database. A job on our side drafts the full audit as an email draft, and a person reads it before it is sent. It never sends on its own.
The job scan reads public freelance posts on seven sites and scores each one against our fixed packages. From 29 September to 3 October it scored 139 posts. 102 matched none of our packages, one reached the bar, and we have sent no bids. The full write-up of all four systems is on our studio build page. For the store-side version of this pattern, see our guide to Shopify automation workflows with n8n, and for what this kind of work costs, our breakdown of AI automation agency pricing.
When should you automate content, and when should you write it by hand?
Automate content when you publish often enough that nobody can read every draft twice. Below that, write it yourself.
- Publishing several times a week across a blog and social? A workflow with checks pays off.
- Publishing once a month? Write it by hand. Building and tuning the checks takes longer than the writing.
- Write down the facts the AI may use (your field notes) before it drafts anything.
- Put the cheap checks in code: length, links, banned phrases, numbers that must appear in the source.
- Keep one person on approval for anything public, with a reminder when items sit too long.
- Review your rules monthly. A rule that fails the same way every time may be the problem.
If you would rather hand this kind of admin to a team that runs it and automates it, that is our ops and AI automation service. More stories like this one are in behind the work. Or start with a free ops audit and we will name the three jobs in your week we would take over first.
Frequently asked questions
Can AI write business content without a human reviewing it?
It can produce drafts, but we do not let anything public go out without a person approving it. In our own workflow the checks reject most drafts before a person ever sees them, and the fact check still found a misattributed number that a quick read would have missed.
What should code check before a person reviews AI content?
Check the things a machine can check reliably: word count, required links, banned phrases, no invented external links, and that every number in a social post also appears in the source article. That leaves the reviewer to judge tone and whether the piece is worth publishing.
How long does human approval add to each post?
Approval waits for the reviewer, so a post can sit for hours or days. In our setup a digest email lists everything waiting, and a reminder goes out if an item sits in review for more than 48 hours.
Is it worth automating content for a small business?
Only if you publish often. For one post a month, writing it yourself is faster than building and tuning the checks. The workflow starts to pay when you publish several times a week across a blog and social channels.
