What Will This Actually Cost?
External quotes cover the build. They do not cover your own team’s time specifying, testing and correcting it, which for most AI projects is a comparable amount. This estimator includes both, plus the contingency nobody budgets and always uses.
The three-year total is the number worth planning against. Projects budgeted only on the build cost run out of money in year two.
AI Project Cost Estimator
Including the internal time that never appears on an invoice.
Cost rises faster than linearly with each one.
Specifying, subject-matter input, testing, review.
20-30% for a first AI project.
Licences, infrastructure, inference.
Build plus internal time plus contingency. Excludes running costs.
An indicative planning estimate, not a quote. Day rates vary considerably across the Australian market. Data quality, permission complexity and regulatory requirements are the factors most likely to move the real figure above this estimate.
What Gets Missed
Three categories of cost are absent from almost every AI project budget, and together they typically account for a substantial share of the true total.
Your own team’s time
Specifying the process, providing subject-matter input, testing outputs, and working through the first months of edge cases. This is real cost against real salaries, and it is routinely left out entirely because no invoice arrives for it.
Maintenance is higher than for normal software
AI systems need more ongoing attention than conventional applications. Models get deprecated, behaviour drifts as inputs change, and quality needs periodic re-measurement. Budget fifteen to twenty-five per cent of build cost annually.
Contingency you will actually use
AI projects encounter more genuine unknowns than conventional builds, largely because data turns out to be messier than assumed. A project with no contingency is one that will need an awkward conversation partway through.
How the Estimate Is Built
Five components. The first two are what most quotes cover; the rest is what makes the number realistic.
External build
Estimated build days multiplied by the day rate you expect to pay, adjusted upward for integration complexity.
Internal time
Your own staff days spent specifying, providing input, testing and reviewing, at their loaded daily cost.
Contingency
A percentage applied to the combined build and internal cost, covering the unknowns that AI projects reliably surface.
Annual running costs
Licences, infrastructure and inference costs for the ongoing operation of the system.
Maintenance
An annual percentage of the build cost covering model migrations, quality re-evaluation and configuration updates.
What Drives the Cost Up
Four factors that separate a modest project from an expensive one, ordered by how much they typically add.
Number of systems that must connect
Cost rises faster than linearly with each system in scope. Each integration brings its own authentication, data model, error handling and testing surface, and each one is a dependency that can change without warning.
- One system is straightforward; four is more than four times the work
- Legacy systems without proper APIs are dramatically more expensive
- Check licence tiers: API access is often not included in yours
- Each integration is an ongoing maintenance liability, not a one-off
Permissions and access control
The most consistently underestimated component of enterprise AI work. Ensuring users only see what they are entitled to see is frequently harder than everything else combined, and it cannot be skipped without creating a genuine disclosure risk.
- Budget significant effort where multiple user groups have differing access
- Existing permissions are often inaccurate and need remediation first
- This is not optional: getting it wrong is a serious incident
- Costs rise sharply where permissions have been managed informally
The state of your data
Clean, structured, accessible data makes a project straightforward. Scanned documents, inconsistent formats, duplicates and information locked in systems without export capability all add substantial cost, and are usually discovered after the quote.
- Assess data quality before requesting quotes, not after
- Scanned and image-based documents need additional processing
- Duplicate and inconsistent records require remediation first
- Data preparation frequently exceeds the application build itself
Regulatory and assurance requirements
Projects in regulated sectors carry documentation, evaluation and audit requirements that add real cost. This is legitimate and unavoidable, but it needs to be in the budget from the start rather than discovered at a security review.
- Formal evaluation and documentation add meaningful effort
- Security review and penetration testing should be budgeted separately
- Data residency requirements may constrain and increase infrastructure cost
- Allow calendar time for approvals, not just effort
Next Steps
AI Business Case Calculator
Take this cost figure into a risk-adjusted business case.
Build the case →AI Consultant Cost in Australia
The longer written guide to what AI consulting actually costs locally.
Read the guide →Frequently Asked Questions
A reasonable planning ratio is around forty to sixty per cent of the external build days. For a forty-day external build, budget sixteen to twenty-four internal days across everyone involved, process owners explaining how things work, subject-matter experts defining correct answers, testers reviewing output, and a project owner coordinating. Organisations that budget zero internal time either discover it later as an unplanned cost or, more damagingly, fail to provide the input and end up with a system built on the vendor’s assumptions about their business.
Twenty to thirty per cent for a first AI project, fifteen to twenty for an organisation with relevant experience and known-good data. That is higher than you would apply to a conventional software build, and deliberately so, AI projects surface more genuine unknowns, most commonly that the data is messier than anyone realised or that the process has undocumented exceptions. If your data quality is unknown, use the upper end. Contingency you do not spend is a good outcome; contingency you did not budget is an awkward conversation.
Three reasons specific to AI. Model providers deprecate versions on a faster cycle than most software vendors, forcing periodic migration and re-testing. System behaviour drifts as the inputs it processes change, so quality needs re-measuring rather than being assumed stable. And prompts, retrieval configuration and evaluation sets all need updating as your business changes. Fifteen to twenty-five per cent of build cost annually is realistic. Budgeting nothing produces a system that quietly degrades until someone declares the project a failure.
The default is indicative rather than a market quote, and rates vary considerably by firm type, seniority and location. Independent specialists, boutique consultancies and large firms occupy quite different price points for nominally similar work. Rather than anchoring on any single figure, get two or three quotes against the same clearly documented scope and compare the totals including ongoing costs. Wide variation between quotes usually indicates they have interpreted the scope differently, which is itself worth investigating.
Yes, and running one is frequently the best money in the whole project. A small, time-boxed pilot answers the questions that most affect the full estimate, whether the data is usable, whether the approach works for your specific case, whether users will adopt it. It costs a fraction of a full build and it converts the largest unknowns into knowns before you commit. Organisations that pilot first tend to produce far more accurate full-project estimates, because they are estimating something they have already partially built.
Investigate rather than celebrating. The most common explanations are that the quote covers the build only and excludes your internal time, that it assumes clean and accessible data, that it excludes permissions and access control work, or that it excludes ongoing maintenance. All four are legitimate scope decisions if made explicitly, and all four become variations later if they were not. Ask the vendor directly which of these are in scope, and get the answer in writing before signing.
Want a Real Estimate?
Describe the use case, the systems involved and the state of your data. We will give you a realistic range and tell you what would move it in either direction.