TIL: don't forget to include the cost of the data pipeline feeding the agent.

Hey folks,

I've been deep in the weeds building an agentic workflow for customer support ticket classification, and I just had a major "aha!" moment I need to share. When we all calculate the TCO for these fancy new AI agents, I think we're almost universally missing a massive line item: **the data supply chain.**

We get so focused on the per-token cost of the LLM call or the monthly fee for the agent platform that we forget the agent is only as good as the data we feed it. Building that pipeline—the one that reliably gathers, cleanses, structures, and delivers the context—is a whole separate project with its own costs. I didn't fully account for this, and it blew my initial ROI projections out of the water.

Let me illustrate with my project. The agent itself (using an off-the-shelf platform) was maybe $500/month. But to give it the right context from our internal systems, I needed:
* **Real-time ticket data** from our help desk (Zendesk).
* **Customer purchase history** from our ERP (Netsuite).
* **Recent support interactions** from our chat logs (Intercom).

Suddenly, I'm building a multi-step integration just to *feed* the agent. The costs stacked up fast:

* **Development Hours:** Building the middleware (I used Make) to normalize and merge data from three different APIs.
* **Ongoing Maintenance:** API schemas change. Error handling for when one source is down. Logging and monitoring.
* **Data Enrichment:** Sometimes the raw data isn't enough. I had to add a step to fetch product details, which meant another API call (and cost) *before* the LLM even saw the prompt.
* **Infrastructure:** A small VM to run the middleware, plus blob storage for chat history attachments.

```json
// A simplified look at the pipeline cost centers:
{
"agent_platform_monthly_fee": 500,
"middleware_hours_development": 120,
"middleware_monthly_maintenance_hours": 10,
"additional_api_calls_for_enrichment": "~$200/month",
"infrastructure_vm_and_storage": "$50/month"
}
```

The "agent" cost became the *tip of the iceberg*. The real work—and ongoing expense—was the data pipeline silently powering it. My advice? When you model your TCO, map the entire data journey first. Factor in:

* The integration work to get data **out** of your source systems.
* The transformation logic to get it into a useful **context window**.
* The reliability measures (retries, fallbacks, alerts) to keep it **flowing**.

If you're only counting the cost of the LLM call, you're seeing maybe 30% of the picture. Would love to hear if others have hit this same realization and how you've quantified it. Any good spreadsheet templates out there that properly account for the integration layer?

-- Ian