Before: every client starts from zero
An agency with fifty retainer clients runs the same few jobs over and over, by hand.
The on-page team writes meta titles and descriptions one page at a time, across sites with 100 to 500 pages each, on WordPress, Shopify and Squarespace. Every title has to fit the length, use the right keyword, name the right city, and not break a rule the client set months ago.
Account managers read every blog post before it goes to the client: four to eight posts for each client, every month. They check tone, banned words, service area and keywords against what they remember about that client, because the rules live in old emails and in their heads.
Writers get re-briefed on each client, again, every time the work changes hands. And the paid team writes responsive search ad copy by hand: up to 15 headlines of 30 characters and 4 descriptions of 90 characters per ad, counting characters as they go.
None of this is hard. All of it is slow, and all of it depends on someone remembering the right client's rules at the right moment.
What does the workflow look like?
With the agency's files organized the way I described in the last post, a request like "draft new meta titles and descriptions for Example Dental's service pages" sets off a sequence:
- The system reads Example Dental's instruction file: voice, banned words, service area, approved keywords.
- It reads the agency's meta-title procedure: length limits, format, what never changes.
- It reads each page's own content and drafts a title and description from that page, not from a template.
- It checks every draft against the client's rules and the character limits.
- It builds a review page for the person who owns the work.
How do you check AI drafts before they go live?
The system prepares drafts. A person has to check them, and that check has to be fast or it doesn't happen.
On the review page, each new title and description sits next to the page it came from and the one it would replace. Character counts are shown. Anything that breaks a client rule, such as a banned word, a city outside the service area or a keyword the client hasn't approved, is flagged. The reviewer approves, edits or rejects each line. Only approved changes get implemented on the site.
Ad copy works the same way: every headline and description is drafted within its character limit and checked against what that client is actually offering, so one client's promotion never ends up in another client's ad.
How much time this saves is the number I measure at each agency. [TBD: measured result from first agency build]
What about blog QA?
The QA pass doesn't rewrite the writer's post. It reads the post against that client's own voice file and rules, then flags what doesn't match: a banned word, an off-topic section, a service the client doesn't offer, a missing approved keyword. The account manager reads the flags instead of re-reading every post line by line from memory, and decides what to send back.
The rules being checked are the client's written rules, so a new account manager checks against the same standard as one who has had the account for years.
Where does the client data go?
Everything runs on business-tier AI accounts in the agency's own name, with training on your data excluded. Client page copy, briefs and reports stay in the agency's storage. Client website and ad-account passwords never go into a chat tool. Approved changes are made the way your team already makes them.
What else can the same setup do?
Once clients are organized this way, new requests take a sentence: a summary of everything done for a client this month before the report goes out, a list of posts still waiting on client approval, a first draft of next month's topics built from that client's approved keywords. Each one is built on the same organized files, so each one starts out pointed at the right client.
Why is training half the job?
The biggest obstacle I see isn't technical. It's that staff don't know what to ask for. People keep doing the slow version of a task by hand because no one showed them it could be one sentence, with the client's rules already applied.
Showing people those opportunities is half of every setup I do. The other half is making sure that once they start asking, the answers stay right for every client.
Common questions
Can AI write meta titles and descriptions for hundreds of pages?
Yes, if each draft comes from that page's own content and that client's rules. Drafts land on a review page with character counts and rule checks, and a person approves them before anything changes on the site.
Does AI blog QA rewrite the writer's post?
No. It checks the post against the client's voice file and rules and flags what doesn't match. A person decides what to change.
Do we have to switch from WordPress, Shopify or Squarespace?
No. Drafts and checks happen in the agency's own files. Approved changes are implemented on whatever platform the client's site runs on.