Blog Content Approval Bottleneck — AI Training Solution

Overview

Citrus America has a persistent pattern of rejecting first-draft blog content, even when the team directly addresses the client's stated feedback. The cycle — write, submit, get rejected, revise, resubmit — is consuming significant time and producing frustration on both sides. Discussed in the ^[1] call.

The Problem

• The client (primary contact: Miriam) consistently rejects first drafts, regardless of whether prior feedback was incorporated.
• In the most recent round, Ben wrote the copy and directly addressed all noted comments. The client's response was still "you're way off" — without clear explanation of what changed.
• The team suspects the core issue is a domain knowledge gap: Citrus America operates in a technical industry, and the team doesn't fully understand the nuances well enough to anticipate unstated preferences.
• This is compounded by the client's apparent desire to position himself as an industry influencer, which creates subjective, hard-to-pin-down voice expectations.

"He's never happy with the first round, right? Whatever. So then we do another round. This last time, Ben actually ended up doing copy on it. We sent it over and he's like, 'you're way off.' And I'm like, but we addressed the comments that you provided."
— Melissa Cusumano

Proposed Solution: AI Model Training on Client Feedback

Mark proposed building a dedicated AI training workspace that accumulates all past client feedback and uses it to constrain future content generation.

How It Works

1. Collect all past drafts submitted to Citrus America alongside the client's written feedback on each.
2. Feed both into a single AI workspace — the draft as input, the feedback as the correction signal.
3. Repeat iteratively so the model builds a growing picture of what this client accepts and rejects.
4. When writing new content, the AI draws on this history and avoids known failure patterns.

An optional enhancement: conduct a structured interview with the client (recorded via Fathom), then feed the transcript into the workspace as an explicit rules document.

"We write some content, we get his comments, and we keep putting it in one place. And then when we let the AI write, it will take everything he's ever said into consideration. It may not be perfect, but we won't make the same mistakes over and over."
— Mark Hope

Practical Framing

The team also noted that the blog's primary purpose is SEO — getting target keywords on the page without factual errors. The client's desire to have a distinctive voice matters less for website content than he believes. This framing may help set expectations in future conversations.

"The most important thing with these blogs is to get the keywords on the page and not to say anything wrong. It's less about his voice and him becoming this influencer."
— Mark Hope

Action Items

• [ ] Mark Hope — Set up the AI training workspace for Citrus America blog content, seeding it with all past drafts and client feedback
• [ ] Melissa Cusumano — Brainstorm with Carly (on her return) on additional process improvements for the blog approval cycle
• [ ] Team — Consider a structured client interview (Fathom-recorded) to capture explicit content preferences as a rules document for the AI workspace

Status

• Carly was out at time of meeting; full implementation deferred until her return.
• Mark committed to setting up the workspace infrastructure.
• No timeline confirmed for the client interview component.

Related

• ^[2]
• ^[1]
• ^[3]