The license is the cheap part

Every AI automation proposal opens with a number designed to look cheap: a license, a per-user price, a per-agent fee, or close to zero if the model is public. That number is real, and it is the smallest one you will pay. Your budget is decided by what it doesn't mention: the team to run it, the maintenance that never ends and the token meter that climbs with volume.

Pricing AI by its license is like pricing a car by its key. Here is the rest of the iceberg, layer by layer.

First hidden cost: the team you didn't know you were hiring

Calling the model costs almost nothing. The people who turn it into something that holds up in production do cost. If you build it in-house on public agents, you need AI and data engineers, someone to run the models, someone to govern them, and the security, testing and monitoring work that never shows up in the demo. Before the first token, you already have a new department.

For a mid-sized company, salary is not usually the deciding factor. The question is whether it makes sense for your technology team to build and run AI when AI is not what you sell. Every engineer maintaining agents is one fewer improving the ERP or strengthening cybersecurity.

Agent platforms hide the same thing more elegantly. The annual license is the entry fee; after that you still hire architects, developers and support staff to run it. The license is one piece of the total cost, not the total cost.

Second hidden cost: maintenance that never ends

Automation is not bought once: you rent a tuning problem that lasts as long as you use it. It is RPA's open secret. According to HfS Research, between 70 and 75 % of the total cost of an RPA program is maintenance, not development. The vendor changes a screen and something breaks. Every exception asks for a new rule.

What looked like a one-off project ends up as a fixed line on the payroll, and it grows with every new automation.

Third hidden cost: the token meter

In the pilot it looks like a tip; in production it blows up. Every reasoning step, every retry and every document an agent reads consumes tokens. Gartner found in 2026 that agentic AI uses between 5 and 30 times more tokens per task than a standard chat, because it plans, calls tools, checks and tries again. With ten test cases, the meter shows a few dollars. In production, with long flows and tasks running overnight, it is among the largest and least predictable costs.

The trap for finance is assuming that, since the price per token is dropping fast, the bill will shrink on its own. The opposite happens: consumption grows faster than the price falls. According to the FinOps Foundation's State of FinOps 2026, 73 % of companies went over their AI cost projection, and some organizations spent three times their token budget.

Without an orchestrator that sends simple steps to small models and saves the large ones for the hard part, the meter has no ceiling. That is why it doesn't pay to tie your operation to a single expensive model: token strategy is cost strategy.

The four options, with the full bill

• RPA: a low license plus maintenance that takes three quarters of the cost.

• Public agents built in-house: almost free to call, but with your own team of specialists and tokens with no ceiling. A hiring plan disguised as a tool.

• Agent platform: a high annual license plus a team to run it, and you absorb the tokens.

• A service that answers for the result: you pay for the work delivered and the provider carries the team, the maintenance and the tokens.

What total cost means

What it costs to run, at real volume, over its whole useful life, with the people included. If you add up license, team, maintenance and tokens, the ranking almost always flips: the "cheap" or "free" option turns out to be the most expensive, because the money didn't disappear; it moved from the invoice you see to the payroll and the meter you didn't model.

And what looked expensive, paying for a result while someone else carries those layers, is often the option with the lowest real cost. The question is no longer only who gets there faster, but who eats the part of the cost that never appears on the first slide.

That is the logic we work by. A Yunt digital collaborator is a role, not software: you don't buy the tool, you buy the result, and the software comes inside. The team that keeps it working comes with the role, as at an airline, where the fleet flies thanks to the mechanics on shift.

The next time an AI proposal reaches you, do the math the vendor left out, at real volume and over several years. The license is the entry fee. Ask what the full season costs: that is the number your CFO is going to live with.

Sources

• HfS Research, maintenance cost in RPA programs.

• Gartner (2026), token consumption of agentic AI.

• FinOps Foundation, State of FinOps 2026.