The honest answer: it depends, but the ranges are knowable
Most AU agency pricing pages say "it depends" and stop there. The honest answer is that it does depend, but the ranges are not mysterious. After three years of running engagements across financial services, government, mining, healthcare, and software businesses, the bands are predictable. Below is what real builds for real operators have cost. Numbers are AUD, ex-GST, current as of mid-2026.
The bands at a glance
Floors below are the entry point. Upper bounds depend on scope, system count, compliance posture, and timeline. We do not cap engagements; the figure your build lands at is set at scoping.
- Discovery sprint. From $15,000. Two-week scoped engagement. Architecture, integration plan, fixed-price quote for the build.
- AI agent build. From $40,000. Single-workflow near the floor. Multi-step, multi-system, action-taking scales up.
- Bedstone OS deployment. From $80,000 first deployment, plus monthly operate retainer from $8,000.
- Custom software. From $60,000. CRUD apps near the floor. Multi-tenant SaaS with complex domain logic scales higher.
- Financial systems. From $50,000. Quote-to-cash, invoicing, reconciliation, board-pack automation.
- Security audit. From $15,000. Web/API audits near the floor. Full app + infra + cloud + IRAP-aware scales higher.
- Penetration testing. From $20,000. Web-only near the floor. Full external + internal red team scales higher.
- Cloud infrastructure (one-off). From $40,000. Greenfield AWS or Azure landing zone, migrations, IaC adoption.
- Monthly retainer. From $12,000 per month, scaling with capacity reserved.
- Embedded engineer. From $25,000 per engineer per month, three-month minimum.
The full structured comparison, with engagement types and trade-offs side by side, lives on the pricing page. The rest of this post explains what shapes the band you'll land in.
What actually drives the cost
If you understand these eight factors, you can pre-estimate any AU agency quote within 30% without seeing the proposal. In rough order of impact:
1. Number of source systems
Every additional system adds connector work, schema discovery, identity mapping, and rate-limit handling. The first three systems set the baseline. Each system after that adds roughly $8,000 to $20,000 depending on API quality. A "10-system Bedstone OS" is not just 3.3x the cost of a "3-system Bedstone OS"; it's closer to 2.5x, because the platform overhead amortises. But it is not 1x. Operators commonly underbudget here.
2. Access control complexity
If everyone in your business can see all the data, integration is fast. If access scoping has to mirror role-based, field-level, or document-level permissions in source systems, integration takes considerably longer. Financial services and healthcare engagements almost always land in the second category; software startups commonly land in the first. Adding identity-aware retrieval mid-build is the single most expensive change-order pattern we see.
3. Data quality of the source systems
If your CRM is reasonably tidy, building against it is fast. If it has six years of inconsistent field usage, duplicate accounts, free-text where structured data should be, and abandoned custom objects, the project includes a data quality pass before the AI work starts. That pass commonly adds $20,000 to $50,000 and is the single biggest source of cost surprises on AI builds. Audit your data layer before you commission an AI build, not during.
4. Compliance posture
Australian Privacy Act and OAIC obligations apply by default. APRA CPS 234, IRAP, sovereign data residency, SOC 2, and sector-specific frameworks (My Health Records Act, SOCI Act) add evidence trails and architectural constraints. A PROTECTED-classified deployment is roughly 2x to 3x the cost of an unconstrained commercial deployment of the same scope, mostly because of evidence work and infrastructure choices, not the AI itself. Plan accordingly.
5. Identity and SSO maturity
Clean Entra ID, Okta, Google Workspace, or Auth0 setup makes integration fast. Fragmented identity across separate user directories adds a consolidation pass. If your business runs three different identity providers (one for the legacy ERP, one for the SaaS stack, one for the development tooling), expect to deal with that before the AI work matters.
6. Existing cloud account state
Greenfield deploys to a new account are fastest. Deployments into existing accounts with legacy IAM, networking, and observability take longer because we work around what's already there. If you've never run an account hygiene pass, factor it in. AU mid-market businesses commonly carry technical debt in their cloud accounts that gets exposed by AI work.
7. Model and provider choice
Commercial APIs (OpenAI, Anthropic, Google) are fastest to integrate. Open-weight models in your own VPC require GPU provisioning, inference serving, and ongoing operations work that adds to both the build and the retainer. Sovereign requirements that force in-VPC inference can add 20-50% to the deployment cost and meaningful monthly cost in GPU capacity. We covered this in detail on the sovereign AI Australia page.
8. Workflow action scope
Read-only agents are fastest. Agents that take actions (create tasks, send emails, update records, fire payments) require identity-bound API access, reversibility patterns, and approval gates. Adding action capability to an existing read-only agent is roughly $15,000-$40,000 of additional work depending on how many downstream systems are involved.
Where AU operators commonly underbudget
Four predictable surprises:
- The operate retainer. The build is not the cost. The first six months of operate work — adding integrations, tuning retrieval, responding to feedback, model upgrades — commonly runs 40-80% of the initial build cost. Budget for it from the start.
- Internal change management. No agency line item, but the cost of getting your team to actually adopt the system is real. Allow 5-15% of the build cost for internal enablement (training, documentation, change champions). Most operators allocate zero and then wonder why adoption stalls.
- Compliance evidence work. If you're in a regulated industry, the evidence work is often invisible until it's invoiced. Ask any agency to break out the compliance line item. If they can't, they're either hiding it or doing it badly.
- Model API costs. The LLM calls themselves are a real operational cost, not a one-time build cost. For a mid-sized Bedstone OS deployment, the model API bill typically lands between $2,000 and $15,000 per month. Negotiate enterprise pricing with the provider early; the retail rates are not the rates that mid-market AU operators should be paying.
What the cheap end looks like (and why to be careful)
You will find AU agencies quoting AI agents at $15,000-$25,000. Sometimes this is a small, well-scoped piece of work; more often it is a demo wrapped as a build. The pattern:
- A working demo over hardcoded data takes a senior engineer two to three weeks.
- Hardening that demo to handle real customer data, real access controls, real edge cases, and real operational load is the rest of the cost.
- The cheap quote almost always covers the demo and stops there. The hardening becomes a change order, or a series of them, by week six.
If the quote is materially cheaper than the bands above, ask exactly what's in scope, what acceptance criteria look like, and how change orders are handled. If the answers are vague, walk.
What the expensive end looks like (and when it's justified)
You will also find AU consultancies quoting AI agents at $500,000+. Sometimes this reflects genuine complexity (PROTECTED classification, ten-plus systems, multi-tenant compliance requirements). Sometimes it reflects a slide-deck-heavy delivery model where the engineering is subcontracted at lower rates and the consultancy captures margin. The pattern:
- Eight-week discovery phase that produces an architecture document and a "phase 1 roadmap".
- Engineering work that begins three months after the contract signs.
- Significant time spent on governance ceremony, steering committees, and stakeholder workshops.
This shape can be appropriate for large agency procurement processes where steering ceremony is genuinely required. It's almost never the right fit for mid-market AU operators who need a working system in three months, not eight.
R&D Tax Incentive: get the cost back
Custom AI agent builds and novel software work commonly qualify for the Australian R&D Tax Incentive. At the small-company tier (turnover under $20M), that's a 43.5% refundable offset on eligible expenditure. On a $200,000 build, that's up to $87,000 back. The conditions are specific (the work must address technical uncertainty, you need contemporaneous documentation, and the activities must fit the scheme), but most production AI engagements clear the bar with the right delivery partner. We covered the eligibility detail in our R&D Tax Incentive piece.
How to compare AU agency quotes apples-to-apples
Ask every agency the same five questions:
- What's the scope by milestone? Real milestones with acceptance criteria, not phases described in adjective-heavy paragraphs.
- What's the change-order process? If the answer involves "we'll work it out as we go", you'll find out by week four how it gets worked out.
- Who owns the IP at the end? The right answer is you do, with no retained rights on the bespoke work. Any agency that retains usage rights on your specific code is taking value that should belong to you.
- What does the operate retainer cost? If they can't answer until you've signed the build, you're underwriting a cost they're not ready to commit to.
- What's the discovery cost? Free discovery is rarely free. The cost is built into the build quote, which means you can't unbundle if you decide the build isn't for you.
These five questions, applied to every quote, will tell you more than 90% of what you need to know about whether the quote is real.
Bedstone's commercial structure
Reading this and curious how we structure it: we quote fixed-price for defined scope, monthly retainer for continuous work, and embedded engineering for organisations growing in-house capability. We charge 100% on signature by default, falling back to 50/50 only where procurement requires staged payment. We don't do hourly billing, open-ended time-and-materials, or loss-leader first projects. The full layout, with engagement types compared, lives on the pricing page. If you want to know what your specific build would cost, the fastest path is a 30-45 minute scoping call. Bring the five things on the call-prep page and we'll have a clear sense of the band by the end of the call.