Tutti gli articoli

The Bill You See Is the Smallest Line

Ask what an AI agent costs and most teams pull up the vendor bill. It is the easiest number to find, and the least useful.

What the vendor bill shows

Take a customer-servicing agent that handles a contact end to end. The vendor bill covers model tokens, tool calls, the platform's share and the infrastructure underneath. In our worked example that comes to $0.45 per contact, and the token line alone is about 2% of the true cost. It is a real cost and it is worth tracking. It is also the part of the picture that looks best, because it is the only part the vendor can see.

What it leaves out

The rest of the cost sits in payroll and operations, on no bill at all. Someone reviews a sample of the agent's work: $0.12 per contact. On 12% of contacts the agent hands off to a person, and that person's time comes to $1.44 per contact once it is spread across all of them. On 6% the customer has to come back, and the repeat contact costs $0.72. Add it up and the true cost of a contact is $2.73, of which the vendor bill is $0.45. The other 79% was invisible to anyone who only read the bill.

This changes which lever matters. Teams negotiate model prices and switch providers to save cents. The biggest number in the example is the hand-off rate. Cut it from 12% to 8% and the agent's cost falls by more than any model discount could deliver. You can only see that when the hand-off is on the books next to the tokens.

Counting what it is worth, carefully

The benefit side needs the same discipline, in the other direction. Time freed is counted at half of what people cost, because salaries do not change the day an agent arrives. Reviews and repeat work count against the agent, not for it. Costs carry a 10% allowance for what we have not thought of, and a 15% safety margin comes off the benefit. The result is stated as money saved and hours given back to your people. It is never stated as the agent being so many times cheaper than a person, because it is not, and because that is the wrong comparison.

This is what the Agent P&L in the Agent Impact Board is built to show: the full cost of each piece of work done right, the return after all costs, and each agent marked grow, hold, fix or shut down on that basis.

If you only see the bill, the agent looks cheaper than it is, and the levers that matter stay invisible.

All figures are an illustrative example built from industry benchmark ranges, not customer data.

Articoli correlati

A low-poly eye with a circuit-like iris, its edge breaking into glowing letters and shards that drift away.
Vision19 agosto 2026·4 min di lettura

Conversation as Visualization

Conversation as Visualization is our vision for the next era of understanding: a continuous exchange between human storytelling and machine visualization, where data questions trigger visual responses and visuals spark new questions.

Nikos MichalakisLeggi l’articolo