# Media agencies build audit tools after AI buying agents hide true procurement costs

*Opaque agentic costs force buyers to track token spend before scaling automated purchasing.*

By **Jenny Huang Goodman MPA MSc MHSA, Principal** — The Stash Edge, Hako Shikin LLC.
Published 2026-09-06.

Canonical: https://www.pops4.com/stash/articles/media-agencies-and-agentic-ai-buyers-2026-09-06t06-7
Subject: Media agencies and agentic AI buyers
Tags: agentic ai, procurement cost, token economics, supply chain automation, sourcing tools, ai audit

---

Media agencies testing AI agents for programmatic ad buying have discovered a cost accounting problem that physical product buyers should watch closely: the systems work, but no one knows what they actually cost to run, according to Digiday. The opacity has prompted agencies to build internal audit tools before they scale agent deployment.

The issue isn't whether the AI completes the buy — it does. The problem is token consumption. Agentic AI systems make multiple reasoning calls per decision, each call burns tokens, and token costs compound fast when an agent evaluates thousands of inventory options or negotiates placement terms. According to Digiday's reporting, agencies discovered true costs only after reviewing bills that ran multiples higher than initial estimates. Without visibility into per-decision token spend, procurement teams couldn't budget or compare agent performance against human buyers.

This matters for physical product marketers because the same AI agent architecture is coming to merchandise sourcing. Agentic tools that search supplier databases, compare spec sheets, negotiate volume pricing, and route approval workflows will face identical cost opacity. A founding team testing an agent to source **10,000** units of custom tote bags might see the agent complete the RFQ process in minutes — then discover the token bill exceeded the margin on the entire order. The speed gain evaporates if the infrastructure cost isn't tracked at the task level.

The mechanism causing cost drift is agent looping. An AI buyer doesn't make one decision — it makes a reasoning chain. It queries inventory, evaluates fit, checks compliance, proposes terms, revises on feedback, and logs results. Each step consumes tokens. If the agent revisits a decision or runs parallel evaluations, token count multiplies. In programmatic buying, an agent might evaluate **50** placement options per campaign. In physical sourcing, an agent might compare **200** SKU variants across **15** suppliers. Without task-level logging, the buyer has no idea which decisions cost pennies and which cost dollars.

The steal is straightforward: if you test agentic sourcing tools, build cost tracking before you buy at scale. Start by running the agent on a single, documented sourcing task — say, finding a supplier for **500** units of a known product. Log every API call the agent makes: supplier search, spec comparison, pricing request, compliance check. Multiply call count by your token rate (OpenAI GPT-4 runs roughly **$0.03** per 1,000 input tokens, **$0.06** per 1,000 output tokens as of early 2025). Compare total token cost to the labor cost of a human buyer completing the same task in one hour. If the agent cost exceeds **20%** of the human hourly rate, the tool isn't ready for volume. If it's under **10%**, test a second task with more complexity — multiple suppliers, custom spec negotiation — and log again. Build a simple spreadsheet: task type, token count, cost per task, decision quality. Run **10** tasks before you route live sourcing through the agent.

For teams already using agents, retrofit the audit. Pull your API logs for the past **30** days. Total token consumption. Divide by number of sourcing decisions completed. Calculate cost per decision. Compare that to your average margin per order. If agent cost per decision exceeds **5%** of unit margin, either the agent is over-indexing on low-value decisions or the reasoning chain needs pruning. Most agent platforms let you limit reasoning steps or cap token spend per task — use those controls. The media agencies building audit tools aren't rejecting AI buyers; they're controlling cost before the tool controls budget.

The broader lesson is that agentic AI introduces variable cost where you used to have fixed cost. A human buyer costs the same per hour whether they source **10** items or **100**. An agent's cost scales with task complexity and decision volume. That's fine — if you measure it. The agencies that document per-task cost now will scale agents profitably. The ones that assume speed equals savings will find their procurement budget absorbed by infrastructure they can't see.

## The takeaway

Log token cost per sourcing decision before scaling AI buying agents — speed gains vanish if infrastructure cost isn't task-level visible.

---

## Publisher

**Hako Shikin LLC** — Virginia Beach, Virginia. Founded 1997. ASI 217876 · DUNS 18-204-6339.
Principal and author: **Jenny Huang Goodman MPA MSc MHSA**.

- Author: https://www.huanggoodman.com/about
- LLM context: https://www.pops4.com/stash/llms.txt
- MCP endpoint, for AI agents: https://mcp.pops4.com/mcp
- Client dashboard: https://dashboard.pops4.com/
- Catalogue: 70,000+ products, 200+ brands