OpenAI just dropped GPT-4o mini at 15 cents per million tokens. That's 60 times cheaper than GPT-4 was two years ago. Anthropic's pricing on Claude keeps falling. Even the smaller players are racing to the bottom.

I've been integrating AI into client projects for the past 18 months, mostly CRM automation and content workflows. The cost of API calls used to be a real line item. Now it barely shows up on the invoice.

But here's what I'm seeing. Clients come to me thinking cheaper AI means their problems got cheaper to solve. They did not.

The Real Cost Moved Upstream

Last month I worked with a client who wanted to automate their lead qualification process in GoHighLevel. They'd done the math on API costs and figured it would be pennies per lead. They were right about that part.

What they hadn't budgeted for was figuring out what a qualified lead actually meant for their business. Their sales team had three different definitions. Their marketing team had another one. We spent four hours in meetings just mapping out the logic before I wrote a single line of code.

The API calls cost maybe $12 that month. My time cost them a lot more.

This keeps happening. The pattern is consistent across WordPress implementations, HubSpot integrations, custom Keap workflows. The AI part got cheap. The thinking part did not.

What Actually Takes Time

Here's where the hours go on a typical AI automation project:

  • Mapping existing processes that only lived in people's heads (6 to 10 hours)
  • Defining what success looks like with actual metrics (3 to 5 hours)
  • Building test datasets that represent real edge cases (4 to 8 hours)
  • Iteration when the first version misses something obvious in hindsight (always happens, add 30 percent)

The actual API integration? Usually two hours, maybe four if the CRM has weird webhook limits.

I tried to shortcut this once. Client wanted AI to categorize support tickets in their Ontraport system. I built it fast, used GPT-3.5 turbo, got it running in an afternoon. Felt efficient.

It worked for about two weeks. Then they realized it was miscategorizing a whole class of technical questions because I hadn't understood their product architecture well enough. We had to rebuild the prompt logic from scratch. That quick win cost us three extra revision cycles.

Strategy Is Still Manual Labor

The questions that matter have not gotten easier:

What data do you actually have? Not what data you wish you had. I've seen clients assume their CRM had complete conversation history when it only captured form fills.

What happens when the AI is wrong? Every model has a failure rate. You need a human review process or a confidence threshold. That requires workflow design, role definition, training.

How does this fit with the tools you already use? The fanciest AI pipeline is useless if it requires your team to check three different dashboards. Integration architecture still takes planning.

Does this actually save time or just move work around? Sometimes automation just shifts manual work from one team to another. You have to trace the whole process.

What This Means Practically

If you're budgeting an AI project, spend less time worrying about token costs. Spend more time on discovery.

Block out real hours to document your current process. Include the weird exceptions. Talk to the people who actually do the work daily, not just management.

Build a small test first. Run it for a week with manual review of every output. See what breaks.

The AI will be cheap. Your time to fix a strategy mistake will not be.