Designing an AI-Assisted Content Workflow With Real Approval Gates

Most marketing teams added AI to their content process the same way: the writer now starts from a generated draft, and everything else stays as it was. The calendar is still approved in one meeting, the posts still go to “the manager” for a look, and nobody can say afterwards who checked the product claim in post 14 or whether the Arabic caption was ever read by a native editor. Output goes up. Control goes down.

McKinsey’s The state of AI in 2026, published on 25 August 2026, points at the difference. Among the small group it calls AI high performers (about 6 percent of 1,719 respondents, defined as those who attribute an EBIT impact of 5 percent or more to AI use and call the value significant), nearly three-quarters report fundamentally redesigning workflows because of AI, against about a quarter of other respondents. The report describes high performers as redesigning workflows that are enabled by AI “rather than insert AI into existing ones”. It also notes that revenue gains are most often attributed to AI use in marketing and sales.

The survey does not tell you how to redesign a content calendar. This guide suggests one way to do it: an operating model for a monthly bilingual calendar in which AI drafts and people own the strategy, the facts, the brand fit and the sign-off, with named gates, a responsibility matrix and an approval log you can adapt.

Why the old approval step breaks once AI is drafting

In a manual process, the writer was also the first quality check. They researched the product, asked the engineer, chose the words and knew what they were unsure about. When a model writes the first draft, that knowledge is no longer built into the text. The draft reads fluently whether or not it is right, and fluent text is the hardest kind to review.

Three things usually go wrong:

• Review becomes a single, overloaded step. One person is asked to judge strategy, accuracy, tone, Arabic quality, legal risk and design in the same pass. They check what is easiest to see, which is usually spelling and the visual.
• Nobody owns the facts. Specifications, certifications, delivery claims and prices enter the draft from the model’s general knowledge or an old brochure, and the approver assumes someone else checked them.
• The second language is approved unseen. The English is reviewed; the Arabic is assumed to say the same thing. It often does not.

The fix is not more review. It is splitting review into gates, each with one owner, one question and a clear pass or fail.

The six gates of an AI-assisted monthly calendar

The model below uses six gates; a smaller team may merge some of them, but we would keep the questions separate. Each gate answers a single question. Content should not move to the next gate until the current owner has passed it, and a failed item goes back with a reason, not a rewrite by the reviewer.

Two rules keep the gates honest. First, G0 happens before AI is used at all: a model can suggest topics, but it cannot decide what your company needs to sell this month. Second, the person who approves at G5 should see the item exactly as it will go live, with both captions, not a draft in a document.

Who does what: a responsibility matrix

Gates only work if each one has a single accountable person. The matrix below is a starting point to adjust to your team; it uses the familiar R (responsible, does the work), A (accountable, owns the decision), C (consulted) and I (informed). In a small team one person may hold several roles; the point is that each gate still has one name against the A.

The row that most teams skip is the subject expert. In technical sectors such as fire safety, MEP, industrial products or healthcare, the person who can say whether a rating, a standard or an approval is correctly stated is rarely in marketing. Give them a fixed, short slot each month to review only the claims list, not the whole calendar, and they will actually do it.

When an agency writes the content, the matrix does not change; the agency holds the writer and editor roles, and the company keeps the A at G0, G2 and G5. That keeps the decision where it belongs: the company approves what goes out under its name. RX’s social media and content service works this way: the plan is agreed first, content is written in Arabic and English, and the client approves every item before it goes out.

The approval log: one row per item, kept for the month

The log is what turns “we review everything” into something you can check. It can be a shared spreadsheet or a tab in your project tool. Keep one row per content item with these columns:

1. Item ID and channel (for example, LI-07 for the seventh LinkedIn post).
2. Brief reference: which objective in the G0 brief it serves.
3. AI use: none, draft, translation support, image generation or image editing. This matters later for disclosure and for knowing which outputs to re-check if a tool or prompt changes.
4. Sources used: the datasheet, project record or page each claim came from.
5. Gate status: G2, G3 (Arabic), G3 (English), G4, G5, each with the reviewer’s name and date.
6. Return reason when an item fails a gate: wrong fact, off-brand, weak Arabic, rights issue, off-brief.
7. Final approved version: a link to the exact file and captions approved at G5.
8. Published link and date.

Hypothetical example: a clinic group planning 20 items in a month. It might see in the log, at month end, that most returns at G2 came from treatment and service details pulled from last year’s brochure. That is a source problem, not a writer problem, and the fix is to give the writer the current service sheets before the next G1. The return-reason column is the part of the log that improves the workflow; without it the log is only a record.

Data, likeness and labelling rules to settle before the first prompt

G4 is the gate most teams improvise. Write these rules once, in the same document as your voice and terminology sheet, so the writer applies them at G1 rather than the reviewer discovering problems at G4.

What never goes into a prompt

Customer names, phone numbers, inquiry details, staff records and anything else that identifies a person should stay out of general-purpose AI tools unless your company has approved the tool for that data. The UAE, Saudi Arabia and Egypt all have personal data protection laws in force; our guides to Saudi PDPL consent for direct marketing and Egypt’s data protection law and direct marketing set out the dates and obligations. Check the current rules with your adviser; this guide is not legal advice.

Faces, voices and real people

Generated or altered images of real people, including your own staff, need consent you can show. Arab News reported in May 2026 that deepfakes guidelines from the Saudi Data and AI Authority (SDAIA) say content creators must secure explicit consent before using any individual’s likeness and keep auditable consent records, and must apply visible, tamper-resistant watermarks to synthetic content. SDAIA’s own copy of the document could not be opened for this guide, so read the current version on SDAIA’s website before relying on it. Even where your market has no equivalent rule, a consent record and a clear label are the safest default.

Platform labels

Platforms set their own disclosure rules, and they change. Two examples: TikTok’s community guidelines say synthetic or manipulated media showing realistic scenes “must be clearly disclosed”, for example with a sticker or caption, and Meta applies an “AI info” label when it detects industry-standard AI indicators or when people disclose AI-generated content. Paid ads carry their own, stricter rules, which we cover in a separate guide to AI labels on Google and Meta ads. For the workflow, the point is simple: the AI-use column in the log tells the G4 owner which items need a platform check, and the check is made against each platform’s current terms, not last year’s.

Running the month: cadence and what to measure

A monthly calendar fits the gates more easily if the work is batched rather than run post by post. A typical rhythm looks like this:

• Before the month: G0 brief agreed; source pack updated (current datasheets, approved claims, offers, dates, glossary); AI instructions refreshed if the voice sheet changed.
• Drafting batch: G1 for the whole month, with the claims list produced alongside.
• Expert slot: G2 in one sitting, on the claims list only.
• Language pass: G3 in Arabic and English, by separate editors, on the designed items rather than plain text, so direction and on-image text are checked too.
• Clearance and sign-off: G4, then G5 on the full month in calendar view, followed by scheduling.
• During the month: any new or reactive item goes through the same gates in short form, with the same log row. Speed is not a reason to skip G2 for a product claim.

Measure the workflow, not only the posts. At month end, look at three things in the log: how many items passed each gate first time, the most common return reasons, and how many items were published without a complete row. The first two tell you where to fix the inputs (the brief, the source pack, the voice rules). The third tells you whether the gates exist in practice.

If the voice and terminology rules are the weak point, our guide to turning brand guidelines into rules AI tools can follow covers how to write them so the drafts at G1 need less editing.

Signs your gates are for show

A workflow can look complete on paper and still let errors through. Watch for these:

• The same person is accountable at G1, G2 and G5. That is one review, not three.
• G3 is done on the English only, or the Arabic is “checked” by comparing it with the English rather than read as Arabic.
• The approver signs off a document of captions and never sees the designed post.
• Return reasons are always “minor edits”. Either the drafts are excellent, or reviewers are fixing problems quietly instead of sending them back, and the inputs never improve.
• Nobody can say which items used generated images, so nobody can say which ones needed a label.

To start, pick next month rather than redesigning everything at once: write the G0 brief, set up the log with the eight columns above, and book the subject expert’s claims slot before any drafting begins. At the end of that month, read the return-reason column first; it tells you which input to fix before the second month.