This is deliberately narrow: not the total cost of an AI platform, or a self-hosting-versus-managed decision, but the full lifecycle cost of running a single deployed agent, from the day it's built to the day it's retired. Most cost conversations stop at the build invoice. The build is usually the smallest of the four numbers that actually matter.
The four cost buckets that make up an agent's lifetime
Build cost: the one-off engineering time to design, write and test the agent, typically $2,500 to $12,000 for a well-scoped single-purpose agent
Token cost: ongoing per-run API spend, which scales with volume and is the only cost most people track
Monitoring and correction: staff time reviewing outputs, catching drift, and fixing edge cases the agent gets wrong
Maintenance: updates when an upstream tool changes its API, when a prompt needs retuning, or when the underlying model is upgraded
A worked example
A Sydney property management business built an agent to draft arrears-and-maintenance correspondence from tenant emails. Build cost was $6,800. In its first year, token cost ran roughly $900, but the real number was in the other two buckets: about four hours a month of a property manager reviewing and lightly editing drafts before sending, roughly $2,600 across the year at a loaded rate, plus a one-off $1,100 retune when the source email system changed its formatting mid-year. Total first-year cost: close to $11,400, against a build quote that made it look like a $6,800 project on paper.
Why this matters for the build-vs-buy decision
Comparing a $6,800 build quote against a $49-a-month SaaS subscription looks lopsided until you run the full TCO on both sides -- the SaaS tool has its own hidden monitoring and correction cost, it's just spread across a subscription fee instead of itemised on an invoice you can see. The honest comparison is lifetime cost against lifetime cost, not build quote against subscription sticker price, because both numbers hide roughly the same shape of ongoing cost underneath.
The bucket everyone underestimates
The monitoring bucket is the one businesses most consistently underestimate. An agent that's 90% accurate still needs a human catching the other 10%, and that review time doesn't disappear just because the agent is doing most of the work. Budgeting for it upfront, rather than discovering it three months in when a property manager mentions they're spending an hour a day reviewing drafts, is what separates an agent that pays for itself from one that quietly drains staff time nobody accounted for at the start.
If you're costing out a new agent build and want the full lifecycle numbers, not just the build quote, get in touch through /contact and we'll walk through all four buckets against your specific use case.
When the four-bucket model changes the build-or-skip decision
Running the full TCO calculation before building sometimes produces the opposite of the expected answer: a task that looked expensive enough to automate turns out to have low enough monitoring and maintenance costs that automating it is a clear win, while a task that looked cheap to automate on the build quote alone turns out to need so much ongoing correction that the manual process was actually the better economic choice all along. Neither answer is obvious from the build quote by itself, which is exactly why the full four-bucket model earns the extra ten minutes it takes to run.
A useful discipline is revisiting the calculation six months after an agent goes live, using actual monitoring hours logged rather than the original estimate. Estimates made before a single output has been reviewed are guesses; six months of real correction-time data turns the TCO from a projection into a fact a business can actually plan around for the next agent it considers building.
One more factor worth pricing in: retirement cost. Agents built on a specific vendor's proprietary format can cost almost as much to decommission cleanly as they did to build, if the workflow they replaced has since been forgotten and needs reconstructing from scratch. Documenting the manual process an agent replaces, even briefly, before it goes live means a future retirement or handover doesn't start from zero if the agent is ever retired, upgraded, or handed to a new provider to maintain.
A final practical step: build the four-bucket estimate into the original proposal document, not as an afterthought after the agent is already live. A client or a board sponsor who sees build, token, monitoring and maintenance costs laid out together from the start rarely feels blindsided six months later when the ongoing numbers turn out to be real, because they signed off on the honest picture the first time, not a build-quote-only version that was always going to understate the true cost.
