How much does an AI agent cost to build?

The question is reasonable. The reason it goes unanswered is not evasion — it is that two builds described with identical words differ by an order of magnitude in effort.

A booking agent wired to one calendar is a fortnight. A booking agent that authenticates the caller, checks availability against an ERP, respects three tiers of cancellation policy, and hands to a human on anything it cannot close is a different job wearing the same sentence.

How many systems it touches. One integration is arithmetic. Five is combinatorics — every pair of systems that has to agree on a record is a place where they can disagree.

Whether it writes or only reads. A read-only agent that drafts a reply is straightforward. An agent with write access to your system of record needs authorisation boundaries, an audit trail, and a rollback path, because the failure mode is no longer a bad sentence, it is a bad record.

How defined the escalation path is. If you already know what the agent should do when it cannot close something, that is a short conversation. If nobody has decided, that decision is part of the build.

Volume and latency. An agent handling forty calls a day and one handling four thousand are the same logic on different infrastructure.

Insurance. A number given before anyone has looked at your systems has to cover the worst plausible version of the job, because the person quoting has no way to rule it out.

You pay for that uncertainty whether or not it materialises. The alternative is not a longer sales process — it is one call, then a written scope, then a real number.

The one manual process eating the most hours this month. What you actually run — CRM, ticketing, spreadsheets, whatever it really is rather than what the org chart says. And the date something has to be working by, even if it is a guess.

Three answers, and the scope conversation takes twenty minutes instead of three weeks.