What an AI Agent Costs to Build and Run: The Four Lines on the Estimate.
Ask what an AI agent costs and you get a number for the model and the platform: so much per month for the seats, so much per thousand actions. It is a real number and it is metered, which is why it is the one that gets quoted. It is also, over a year, usually the smallest of the four lines an honest estimate has.
This is the estimate we write, line by line, and the reason the quoted line is the least important one.
Line one: the model and the platform.
This is the visible cost. The agent platform's licence, the language model's usage, the per-action metering if the platform charges that way. It is easy to estimate because the vendor publishes it and easy to control because it scales with use.
It is also where the comparison between vendors happens, and the comparison is mostly beside the point. Two platforms that differ by thirty percent on this line differ by a few percent on the whole estimate, because of what the other three lines contain. We quote it, we do not negotiate hard on it, and we tell clients not to either.
Line two: the integration the agent stands on.
An agent acts on records. If it is to follow up a lead it reads the lead, the contact, the account and the last email. If it is to chase an invoice it reads the invoice in the accounting system, the deal in the sales system and the open tickets in the help desk. For any of that to work, the systems have to agree who the customer is, which in most mid-market estates they do not. The customer exists under three names in three systems, and the agent that reads them will act on the wrong one with complete confidence.
So the second line is the integration: deciding which system issues the customer id, carrying it across, reconciling the customers that already exist, and building the connections the agent reads through in a known direction. For a company with a connected estate this line is small. For a company with three systems that have never agreed, it is the largest line on the page, and it is the line that decides whether the agent works at all.
Vendors do not quote it because it is not their product and because it is unmetered. It arrives anyway, as change orders in month three, or as an agent that quietly does the wrong thing. We quote it first, because it is the precondition, and a client who learns it here rather than in month three has saved the project.
Line three: the limits that govern it.
An agent needs an identity of its own, a profile and role built from nothing up to what the job needs, fields hidden that the job does not touch, and approval rules on the actions that should wait for a person: amounts, closing stages, bulk changes, customer email. In a suite like Zoho the approval process already exists and governs an agent the same way it governs a rep, so most of this line is configuration and the document that explains it. We walked the controls in the piece on governing agents in a mid-market company.
The line is an afternoon to a few days per agent, plus the goal written in one sentence with a limit in it, plus the exception list extracted from the people who hold it. It is never large. It is also the line most often skipped, because the agent runs without it, right up to the first surprise.
Line four: the watching.
The agent runs. Someone reads what it did. The held proposals wait in an approval queue and someone decides them. When the agent meets a case its goal did not cover, someone adjusts the goal. That is an operating cost, in a named person's hours per week, and it continues for as long as the agent runs.
The watching is where the agent's value is actually realised, because the held proposals are the decisions that were worth a person, and the log is where the company learns what the agent found that nobody knew. A company that budgets for an agent and not for the hour a week to watch it has bought a system that reports success and is corrected by nobody.
This line is small per agent and it adds up across agents. Five agents is five hours a week, and the person doing it is usually someone whose time was already fully allocated. Budget the person, not just the platform.
What the lines look like together.
For a company with a connected estate and one well-defined job, the four lines are modest and the first is the largest. For a company with disconnected systems and a vague goal, line two is several times the other three combined, and line one is a rounding error. Most mid-market companies are closer to the second description than the first, and most quotes they receive describe the first.
The lines also behave differently over time. Line one scales with use and can be turned down. Line two is paid once per connection and then holds, which is why a company that has done it for one agent finds the second agent cheap. Line three is paid once per agent and revisited when the job changes. Line four runs for as long as the agent does and grows with the number of agents, so it is the line that decides how many agents a company of eighty can actually operate, which is fewer than the platform will sell it. A second-year budget that only carries line one has forgotten the person who clears the queue.
That is the whole argument. The agent development cost is not the agent. It is the state of the records underneath it and the discipline around it, and a quote that does not show those lines is a quote for a demo.
How to read a quote.
Three questions, and the answers tell you whether the estimate is real.
Which records does the agent read and change, and are they the same customer, order and item across the systems it touches? If the vendor does not know, line two has not been estimated.
Which identity does the agent run as, what can it see, and which actions wait for a person? If the answer is the administrator's login, line three is missing and the first surprise is already scheduled.
Who reads the log and clears the queue, and how many hours a week? If the answer is nobody yet, line four is an unbudgeted person.
An estimate that answers all three has four lines and can be guaranteed. Ours is. Discovery is no-risk: we read your systems and your candidate job, write the four-line estimate, and you pay only if you go ahead.
Find out what one connected Zoho system would change in the way you run.
In a no-risk discovery we look at your CRM, finance and operations systems and who owns each part, and show what a connected system would do differently. You pay only if you proceed. Or see how we approach it.
More on the same problem:
Comments