01 the business
Group of three restaurants with a catering arm
90 staff, two marketers
Two marketers run a newsletter, three social accounts, a blog and seasonal campaigns for three venues. Generic AI tools give them copy that sounds like everyone else and gets menu details wrong. The brand guide is a PDF nobody opens.
02 before
The problem
A one- or two-person marketing team has to fill a newsletter, several social channels and a blog. Generic AI tools produce copy that sounds like everyone else and gets product details wrong. The brand guide sits in a PDF that nobody opens.
- Every campaign starts from a blank page.
- Prices and dishes in posts sometimes contradict the live menu.
- Each venue's voice drifts depending on who wrote the post.
03 the build
What the AI does
- Step 1: Learns your voice from the brand guide, approved past campaigns and product or service sheets stored in Google Drive or Notion.
- Step 2: Takes a short brief, such as a new menu, a course launch or a seasonal offer, and asks for anything missing.
- Step 3: Drafts the pieces you need for each channel, from a Mailchimp email to Instagram captions and a blog outline.
- Step 4: Checks prices, dates and product names against your source documents and marks any line it could not confirm.
- Step 5: Saves drafts to your shared folder or straight into WordPress or Webflow as unpublished posts.
paused: waiting for a person to approve
How it was put together
- Loaded the brand guide, ten approved past campaigns and the current menus from the shared drive.
- Built a short brief form that asks for anything missing before drafting.
- Checked every price, date and dish name in a draft against the source menu and marked anything unconfirmed.
- Saved drafts as unpublished posts in the website and newsletter tools.
04 human in the loop
Where a person stays in charge
Every draft stays unpublished until someone on the marketing team edits and approves it. Anything that mentions a price, offer or health claim is highlighted for a closer read.
- Nothing publishes without a marketer's approval.
- Lines that mention prices, offers or allergens are highlighted for a closer read.
- The assistant uses only documents the team has loaded, not the open web.
05 rollout
From kickoff to live in 2–3 weeks
- Week 1Collect brand material and approved examples; agree the brief form.
- Week 2Draft one real campaign in parallel with the team and compare.
- Week 3Use draft-first for all routine content; keep big launches human-led.
06 after
What changes
The team spends its time on ideas and editing rather than first drafts. Copy sounds like your business across every channel, and product details match what you actually sell.
Assume 40 pieces of content a month at 45 minutes each to write from scratch, and 15 minutes to edit a draft.
40 × 45 minutes = 30 hours today. 40 × 15 minutes = 10 hours with drafts. About 20 hours a month go back to ideas and photography.
What to measure
- Time from brief to approved draft.
- Edits needed per draft.
- Factual errors caught before publishing.
- Posting consistency across channels.
07 evidence
What the research says
In a study of 453 professionals doing writing tasks from their own jobs, access to ChatGPT cut the time taken by 40% and raised quality, as rated by evaluators, by 18%.
Google rewards original, high-quality content however it is produced, but using automation mainly to manipulate rankings breaks its spam policies.
Google Search Central, Google Search's guidance about AI-generated content (opens in a new tab) · 2023
About 75% of generative AI's potential value sits in four areas: customer operations, marketing and sales, software engineering, and R&D.
08 lessons
Mistakes to avoid
- Training on everything ever published. Choose the examples you are proud of.
- Skipping the fact check. The brand voice is easy; the details are where trust is lost.
- Publishing straight from the tool. Keep the approval step.
09 questions
Common questions
- Will Google penalise AI-written posts?
- Google's guidance is that it rewards helpful, original content however it is produced. Content made mainly to game rankings breaks its spam policies.
- Can it write in more than one language?
- Yes, and the same checks against your source documents still apply.