I spent the first half of 2023 adding ChatGPT API calls to existing automation workflows. Customer support got a bot. Lead qualification got a summarizer. Data entry got an extraction step.

Everything got a bit faster. Nothing got fundamentally better.

Then I rebuilt one process from scratch with AI as the starting assumption, and the difference was obvious. I'm talking about a 70% time reduction instead of 15%. I'm talking about removing three tools from the stack instead of adding another integration.

The Bolt On Approach

When you bolt AI onto an existing process, you inherit all the constraints that were there before. Here's what I did with a client's lead intake system in early 2023.

Their original flow: form submission goes to GoHighLevel, triggers a webhook to Zapier, Zapier formats data and sends to HubSpot, HubSpot assignment rules route to sales rep, rep gets email notification.

My improvement: added a ChatGPT call in Zapier to score lead quality and write a summary. It worked. The sales team loved the summaries. But we still had five systems in the chain, still had the same 3 to 8 minute delay, still paid for three platforms.

The API costs were negligible, about $0.02 per lead with GPT-3.5. But I added another point of failure and another thing to monitor.

The Ground Up Approach

Two months ago, I rebuilt a content approval workflow for a marketing agency. Instead of asking where to add AI, I asked what the AI made unnecessary.

Old process:

  • Writer drafts in Google Docs
  • Editor reviews, leaves comments
  • Writer revises
  • Editor approves
  • VA copies to WordPress
  • VA formats and adds images
  • Manager spot checks
  • VA schedules

Average time: 4.5 hours per article across three people.

New process: Writer drafts in a custom WordPress plugin I built. On submit, GPT-4 checks brand voice against three example articles (embedded as context), flags inconsistencies, suggests fixes. If the score is above 8/10, it goes straight to a simple approval queue. The same plugin handles formatting and pulls relevant images from their library based on content analysis.

Average time now: 1.2 hours, mostly writing.

I cut out Google Docs entirely. No more version confusion. The editor role changed from line editing to spot checking AI recommendations. The VA role disappeared from this workflow (we moved her to video content where she's better utilized).

What Did Not Work

I tried to use GPT-4 for image selection in the first version. It was terrible. The API can't actually see images well enough to match them to nuanced content needs. I switched to a simpler keyword and tag matching system, and it works better.

I also assumed we could eliminate the human approval step. We can't. About one in twelve articles gets flagged correctly by the AI for tone issues that the writer genuinely missed. But about one in twenty gets flagged incorrectly, usually when the content is deliberately breaking from standard voice for effect. You need a human to distinguish between those cases.

The Real Difference

Bolting on AI means you're still optimizing the old process. Rebuilding means you question whether each step should exist at all.

The approval workflow saves the client about $2,800 monthly in VA costs and editor time. But more importantly, articles publish faster. Their SEO strategy can actually keep up with trends now.

When you're spec'ing your next AI project, start with a blank page. Map what you want to achieve, not what you currently do. Then ask what's actually required to get there if AI handles the parts it's genuinely good at.

You'll end up with fewer steps, not more.