Understanding the cost, the output, and the ownership gap across your agentic workforce

Most organizations running AI agents can tell you what they spent last month. Far fewer can tell you what that money bought. The billing dashboard adds up tokens, seats and API calls, and stops there: it can’t say which agent did the work, whether the work was any good, or whether the same result was available for a tenth of the price.

Over three months, Geordie spoke with more than 50 organizations running agents in production, across financial services, technology, healthcare and private equity. Most came to us through security. The questions they kept asking were about money. Security can usually see what agents are doing, finance holds the budget, and in most of these companies nobody has joined the two. With no owner and no way to separate useful spend from waste, the default response was a freeze.

This research paper sets out what those conversations turned up. Some of it is practical, like the screenshot of a document that costs several times more to process than the same text pasted in, or the free vendor credits that hide the real run rate until they run out. Some of it is cautionary: more than one team that measured AI adoption by tokens consumed found staff using more tokens to look busy, Goodhart’s law in about a week. The paper is also plain about what nobody, Geordie included, has solved yet, starting with the question every leader asked: is my number normal?

What you will take away

• Seven recurring waste patterns, ranked by how often they came up, with who is best placed to fix each one
• What attribution actually requires before a cost can be traced to an agent, an action and a named owner
• How to set a declared budget per agent, and when to revisit it
• Four levels of intervention, from showing a cost to the person who caused it to blocking the action outright
• A maturity path from freeze to devolved budgets, and six questions to put to any vendor selling agent cost management