---
title: AI marketing agents case study: 99% customer approval | Codivated
canonical: https://www.codivated.com/case-studies/ai-marketing-agents/
---

[Home](https://www.codivated.com/)/[Case studies](https://www.codivated.com/case-studies/)/Marketing agents

# AI marketing agents customers approve 99% of the time.

Five AI agents. Nothing posts without the customer's yes.

Search ⌘KAG

Queue 3ScheduledPublishedAgents

Blog · SEO agentDraft · waiting for you

5 spring maintenance tips for homeowners
Spring is the best time to check the parts of your home that took a beating over winter. A little attention now saves a bigger repair bill later.

- Clear gutters and downspouts
- Check the roof for loose shingles
- Service your air conditioning before the heat

▣Hero image · generated

ApproveAsk for changesDeny

Checks before you see it
Brand voice

No prices

No competitor names

Reviewer agent

Agent notesTone matched to the brand guide. Two claims removed.

Illustration with example content

Customer decisions on agent drafts99%approved

2.5M+ actions~18,000 customers5 agents

What our team built

## Five agents, one approval queue.

- 01Lead nurtureFollows up with the customer's own leads.

- 02SEOPlans and drafts search content.

- 03BlogPublishes to the customer's website.

- 04SocialPosts to connected social accounts.

- 05SupportHandles support questions.

Every draft from every agent lands in one queue in the customer's account: approve, deny or ask for changes. The same pattern works for anything AI sends in your company's name: emails, quotes, posts, replies.

What broke

## At 50 rules, the prompt stopped listening.

The first version put every brand and safety rule into one long prompt. Past about 50 rules, the model stopped following them reliably. So we moved the rules out of the prompt.

Before: one prompt (example rules)
- Never mention prices
- Match the brand voice
- No competitor names
- Use the business name exactly
- …46 more rules

After: three layers
- Checks in codeFixed rules, checked every time.
- Reviewer agentsSmall agents, one job each.
- The customerApproves every piece in the queue.

What we learned running it

01

### Approval is the feature.

Customers let agents post for them because nothing goes out without their yes.

02

### Rules belong in code.

A long prompt is a wish list. Checks in code and small reviewers hold every time.

03

### Run it by hand first.

The blog engine ran manually before it was automated, and it was rebuilt several times. Each pass showed what to automate next.

## Questions buyers ask us.

### Can AI agents post for our customers without hurting their brand?

Yes, if nothing goes out without the customer's approval. Codivated puts every draft in an approval queue and checks it against fixed rules first.

### Can we sell AI agents as a paid feature?

That is what this product did. The agents shipped as a paid feature and are used by about 18,000 customers.

### What happens when a customer says no?

They can deny a draft or ask for changes, and the agent revises it. About 99% of drafts were approved.

Codivated is an AI consulting and agent development firm that builds production AI agents for mid-market companies.

## Want AI that never speaks for you without your yes?

A 30-minute call. We look at your product and where agents would earn their keep. We'll tell you what it takes from your team.

[Talk through AI features](https://www.codivated.com/contact/)

Next case study[An enterprise AI platform with 100,000+ sessions a month](https://www.codivated.com/case-studies/enterprise-ai-platform/)
