Context
When you are asked to approve a budget, you usually get one figure and a justification. With an AI agent the figure often comes split into several pieces, and each piece behaves differently. One is paid once, another every month, and another according to how much is used. If you mix them up, you compare proposals badly and learn the real cost around the third invoice.
This lesson has no prices. They change by company, scope and model provider, and a loose number would not help you. What helps is knowing what the cost is made of and what to ask.
The core idea
The cost of an agent has three layers, and it pays to look at them separately.
The first is Discovery. It is preparatory work, and it is paid, to understand how you really work: who does what, with which tools, where the process gets stuck and what data it touches. It helps you decide whether an agent makes sense and where. It is worth asking that it end with a written scope.
The second is implementation. This is building and installing the agent: connecting it to your systems, giving it instructions, testing it with your own cases and leaving it working alongside your team. What it includes and how it is charged is defined in the proposal.
The third is the monthly fee, which corresponds to the Maintenance phase. It covers keeping the agent running and handling adjustments within the helpdesk hours. It includes a pool of tokens and a pool of helpdesk hours, meaning support for when something fails or you need an adjustment. If you use more than what is included, the excess is paid by consumption.
This is where the token comes in. A token is a piece of text, similar to a syllable or a short word, that the AI model reads or writes, and it is the unit model providers use to measure usage. More work from the agent means more tokens. That is why the monthly fee comes with an included pool and the excess is metered, like extra data on a mobile plan.
How it is done
Reading a proposal means separating those layers and asking concrete questions. These are the ones that help most:
- What does Discovery cover, and what do I receive when it ends. A scope document you can review with your team is a good sign.
- What does implementation include, and what stays out. Every system that must be connected is work, so ask for the list.
- What does the token pool include, what do the helpdesk hours include, and what happens when they run out.
- How is the excess charged, and how will you see the consumption. Ask for a report that someone in finance can read, not only someone technical.
- Where does the agent run. With dedicated infrastructure, the virtual machine, the Anthropic account and the agents are yours, with no shared infrastructure.
- How long is the contract, and when is it signed. At Omni Agent the 12-month contract is signed after Discovery, when you already know what you are buying.
To think about return, do not start from a savings percentage someone promised you. Start from your own operation. Pick one process, for example the month-end reconciliation, and measure today how many hours your people spend on it, how many errors they correct and how many days it takes. That is your baseline. After a few weeks with the agent, measure the same things. The comparison is yours, with your data.
Add the cost of review as well. Agents make mistakes, and someone has to check their work as they would a colleague's. That review takes time and belongs in the math. And ask yourself what your people will do with the hours they get back, because that is where the value is: a broader role, with more time for what only a person can decide.
An illustrative case
Imagine a distributor of auto parts. It is an invented scenario, and the numbers are invented for the example too.
The finance director receives two proposals for an accounts payable agent. The first has a single figure, with no breakdown. The second separates Discovery, implementation and monthly fee, with its token pool and its helpdesk hours. She sends both the same questions. What does the pool include. How much does the excess cost. How will she see consumption each month. Who owns the model account.
The second proposal answers everything in writing. The first answers that it depends. That is a difference no figure shows.
Before starting, she measures her baseline. Her accounts payable team spends 60 hours a month entering and reconciling invoices. After two months with the agent, the team spends 15 hours reviewing what the agent leaves ready. The analyst who used to enter invoices now negotiates payment terms with suppliers, something there was never time for. With that data, the director compares against the full monthly fee, including the token excess of the heavy months.
None of this is guaranteed. It is a way of looking at cost with your own hours, measured before and after.
Takeaways
- The cost of an agent has three layers: Discovery, implementation and monthly fee. Each is paid differently.
- The monthly fee includes a pool of tokens and helpdesk hours. The excess is paid by consumption.
- A token is a piece of text the model reads or writes. More work from the agent means more tokens.
- Ask for the scope in writing, the list of what stays out and a consumption report you can read.
- Calculate return with your own baseline, and count review time as part of the cost.
- Be wary of anyone who promises exact savings or a firm timeline before knowing your operation.