Less than people fear for the first useful thing, and more than people expect for the second.

The number that matters is not one price, it is two: what it costs to build, and what it costs to run every month afterwards.

Most quotes cover the first and go quiet on the second, which is how a project that looked affordable in March becomes an awkward conversation in September.

This is what actually moves each number.

The two numbers, and why the second one surprises people

A traditional software build has a big cost to make and a small cost to run. An AI build does not behave that way. Every time it runs, it spends money, because somebody is charging you per unit of work: tokens through a model API, or credits inside a platform.

That changes how you should evaluate it. A build that handles 200 records a month and one that handles 200,000 can cost the same to make and differ by a factor of a thousand to operate. The volume question belongs in the first meeting.

We have written about the run side in detail: what HubSpot's AI agents actually cost to run, and the comparison between platform credits and bringing your own API key at Breeze credits or your own API key.

What actually moves the build number

In rough order of impact.

How many systems it has to touch. One system is straightforward. Two is more than twice the work, because now something has to reconcile them. Three or more and the integration is the project, with the AI as a component.

Whether your data is ready. This is the most underestimated line in any quote. An agent inherits whatever is already wrong in your CRM, so the data work is not preparation for the build, it is part of it. We published the five checks we run first because this is where projects grow without anyone noticing.

How wrong it is allowed to be. A system that drafts something for a person to approve is a good deal cheaper than one that acts alone, because the second needs evaluation, monitoring, reversibility and a much harder test process. Autonomy is the single biggest multiplier on cost and nobody prices it consciously.

Whether a human stays in the loop. Related and not identical. Approval steps are cheap to build and cost staff time forever.

How much of it is judgement. Classifying something into five categories is a small job. Deciding what a customer should be offered is not.

Whether anyone has to explain the decision. If the system makes decisions about people, the disclosure obligation commencing 10 December 2026 applies, which means logging, explainability and a documented position. Design for it at the start and it is modest. Add it after and it is a rebuild.

What we already publish, for comparison

We put real figures on HubSpot builds, because vague pricing wastes everyone's time. From what a HubSpot build costs a 100 person Australian business in 2026: a clean single-hub setup runs A$8,000 to A$15,000, a multi-hub build with migrations and integrations A$15,000 to A$35,000, and revenue architecture with custom objects A$35,000 and up. Separately, HubSpot's own mandatory onboarding runs from A$2,160 for Sales or Service Hub Professional to A$10,080 at Enterprise, and licences from about A$1,380 to A$7,290 a month.

Those numbers are for platform builds, and they are the closest published comparison we have. A custom AI layer sits on top of that work and is scoped separately.

The costs that do not appear in the quote

Four things that are real and routinely left out.

Evaluation. Knowing whether the thing works needs a test set built from your actual cases, and it needs maintaining. Without it you are running on impressions. This is how we test an AI agent before it goes near a customer.

Model changes. Models are replaced and deprecated. Something that worked perfectly can change behaviour after a version change, so budget for periodic re-testing.

The human time. If a person approves 40 items a day, that is a real ongoing cost and it belongs in the business case beside the savings.

Rework from scope you did not know about. Usually the data. Occasionally a system whose API cannot do what everyone assumed.

How to buy it without getting hurt

Ask for build and run costs separately, with the run cost modelled at your real volume. Not per action, per month at the number of records you actually process.

Start narrow. One process, one team, a defined set of cases. The ACSC's agentic AI guidance recommends deploying incrementally starting with clearly defined low-risk tasks, which is a security recommendation that happens to be good commercial advice.

Make the first build prove the number. Pick something where you can measure the before and after, so the second build is funded by evidence.

Price the exit. Ask what you would hold if the relationship ended. If the answer is complicated, the number in front of you is not the real number.

In short

Ask for two numbers, not one. What moves the build cost is how many systems it touches, whether the data is ready, how much autonomy it has, and whether anyone has to explain its decisions. What moves the run cost is volume, and it should be modelled at yours before you sign anything.

Start narrow, measure it, and make the second build a decision rather than a commitment.

Frequently asked questions

How much does a custom AI build cost in Australia? It depends on how many systems it touches, how ready the data is, and how much the system is allowed to do unsupervised. Always get two figures: the build cost and the monthly run cost modelled at your actual volume, since AI systems charge per unit of work.

Why does an AI system cost money every month? Because each run consumes tokens through a model API or credits inside a platform. Unlike traditional software, the operating cost scales with how much work it does, so volume drives it more than headcount does.

What is the most underestimated cost in an AI project? Data readiness. An agent inherits whatever is wrong in the CRM, so cleaning and structuring the data is part of the build rather than preparation for it.

Is it cheaper to use platform credits or your own API key? It depends on volume and on which platform. Credits are simpler and usually cost more per unit as volume grows; your own key is cheaper per unit and adds management overhead.

Sources

Want two real numbers for something specific instead of a range? Talk to us.