Custom Software Development14 min read

Custom software costs the hours of skilled work it takes, multiplied by the rates of the people doing it, plus an allowance for what is still unknown when the work starts. That has always been true. What has changed in the last two years is the number of hours. Agentic development, where AI agents draft the design, the code, the tests and the documentation and engineers review and approve every step, has cut the effort a build needs.
So the useful question is no longer what software used to cost, but what it should cost when it is built this way. This guide explains what drives the cost of custom software, where AI agents save effort and where they do not, and how to tell whether a quote reflects the way software is built now. Where it uses figures from outside IT Magnet, it names the source.
What you are paying for
Every estimate for custom software is a number of hours multiplied by the rate for each role, from the developers and testers who build it to the architect and project manager who keep it on track. Rates follow salaries. In Sydney, Robert Half's 2026 salary guide puts the median starting salary for a software developer at $109,000 and for a solutions architect at $190,750, before superannuation, and an employer adds superannuation, payroll tax, leave and equipment on top.
- Software developer$92,650 to $130,800
- Senior developer (.NET)$147,150 to $168,950
- Solutions architect$179,850 to $212,550
Source: Robert Half 2026 Salary Guide, Sydney figures for software developer, senior .NET developer and solutions architect.
Rates have not fallen, and experienced engineers are worth more than ever, because they are the ones who decide what the agents build and check what they produce. What has fallen is the number of hours those people need to spend drafting.
How agentic development changes the cost
IT Magnet builds with its own AI agents for design, coding and review, each working inside a human in the loop process. A design agent turns an agreed specification into wireframes and working prototypes in days rather than weeks. Coding agents draft features, integration code and data scripts against the specification and the existing code. A review agent checks every change against our standards, the tests and known security vulnerabilities, and then a senior engineer reviews and approves it before anything ships.
The speed comes out of the drafting, not out of the checks that follow it. That distinction matters, because the saving is real only where the agents take over work that is well defined, and the review is what keeps the result secure and maintainable.
| Stage | What the agents do | What people still do | Effort saved |
|---|---|---|---|
| Discovery and specification | Summarise workshops, draft user stories and acceptance criteria | Decide what to build, agree priorities and sign off | Small |
| Design and prototypes | Turn the specification into wireframes and working prototypes | Test them with users and choose the direction | Large |
| Development | Draft features, integration code and data scripts | Set the architecture, solve the hard problems, review every change | Large |
| Code review and security | Check every change against standards, tests and known vulnerabilities | A senior engineer approves before anything ships | Moderate |
| Testing | Write and run test cases, including the edge cases people forget | Acceptance testing with your team | Moderate |
| Documentation | Keep architecture notes and API references current with every change | Check they are right | Large |
| Integrations and data migration | Read the other system's documentation and draft mappings | Work with that system's owners and resolve data problems | Small |
| Project management | Draft status reports and release notes | Make decisions, set priorities, talk to stakeholders | Small |
The pattern is clear in our own projects. Work that is well defined and checkable gets much faster. Work that depends on people, decisions and other organisations gets only a little faster, and on many projects that is where most of the time goes.
Why AI tools alone do not cut the cost
Giving developers an AI assistant is not the same as building with agents, and the research shows why. In a controlled experiment by GitHub and Microsoft researchers, developers using Copilot built a small web server 55.8 percent faster than those without it. But in a 2025 trial by the research group METR, experienced developers working on large codebases they already knew took 19 percent longer with AI tools, even though they expected to be 24 percent faster and afterwards believed they had been 20 percent faster.
- New, well defined task (Copilot study)44% of the time
- Experienced developers' forecast (METR)76% of the time
- Their estimate afterwards (METR)80% of the time
- Measured on mature code (METR)119% of the time
Sources: Peng and others, 2023; Becker and others, METR, 2025. METR studied tools from early 2025.
Our reading of these results matches what we see. AI saves the most on new, clearly specified work and the least when a person is fitting changes into a large system the tool does not understand. That is why our process puts the effort into the specification first, gives each agent a narrow job, and keeps a senior engineer reviewing every change. The tools have also improved a great deal since early 2025, but the lesson about process still holds.
What moves the cost
Two projects with the same number of screens can still differ greatly in cost. These are the things that make the difference, roughly in order of how much they move an estimate, and most of them are the kind of work agents speed up least.
- Integrations: every system the software talks to, such as an ERP, payments, a POS or a government API, has to be understood, built against and tested
- Data migration: moving years of records out of spreadsheets or an old system is often a project of its own
- Security and compliance: government work to the Information Security Manual, privacy obligations, penetration testing and audit logging all add effort
- Platforms: a web app, an iPhone app and an Android app are three things to build, test and maintain
- User types and workflows: each role with its own screens, permissions and approvals multiplies the work
- Unknowns: the less is settled before the build starts, the wider the estimate has to be
| Type of project | What it usually involves | What drives the cost most |
|---|---|---|
| Integration between two systems | Moving data between, say, a CRM and an accounting system, with error handling and monitoring | The quality of the other systems' interfaces and data |
| Internal workflow tool | Replacing a spreadsheet or paper process for one team, with approvals and reporting | How many exceptions the process really has |
| Mobile app with a backend | An iPhone and Android app with its own API and an admin portal | Offline use, payments and app store requirements |
| Customer or partner portal | Secure sign in, self service, documents and integration with internal systems | How many internal systems it reads from and writes to |
| Business system | A CRM, job or operations system replacing several tools | Data migration and the number of roles and workflows |
| Platform with many integrations | Several user types, devices or IoT, heavy compliance and national scale | Integrations, compliance and staged delivery |
The unknowns are why an honest early estimate is a range. Until someone has looked at your data, your integrations and the way your team actually works, nobody can price the job precisely. A short discovery turns the range into a fixed price.
How to tell if a quote reflects how software is built now
Quotes differ because they are rarely quoting the same thing. One includes discovery, testing, project management and a warranty period, another assumes you will test it yourself. One is fixed price with the risk priced in, another is time and materials with the risk left with you. And now one may be priced on agentic development while another is priced on the hours the same work took a few years ago.
Compare quotes by their hours and assumptions, not their bottom line. Ask:
- How many hours does the estimate assume for each stage and role?
- How do you use AI agents in the build, and how is their work reviewed before it ships?
- What is included: discovery, design, testing, project management, deployment, warranty?
- What assumptions is the price based on, and what happens if one proves wrong?
- Is the price fixed, and if so, for which stage of the work?
- Who owns the source code, and is it held in your organisation's own repository?
Be as wary of a quote that is suspiciously low as of one that has not moved in years. Agentic development cuts drafting, not review. If a quote has no time for senior review, testing and security scanning, the saving has come out of the checks.
Budget for after launch
Software keeps costing once it is live. Hosting and licences, monitoring, security patches, the yearly updates Apple and Google require for store apps, support when something breaks, and the improvements your team will ask for once they use it. Plan for a support arrangement and hosting as a monthly cost, and treat larger improvements as their own projects. Documentation that agents keep current with every change also makes later work, and any future handover, cheaper.
You will often see maintenance quoted as a fixed share of the build cost each year. We could not find a primary source for that rule of thumb, and in our experience the real cost depends far more on how much the system keeps changing than on what it cost to build.
Can the R&D Tax Incentive reduce the cost?
Sometimes, but less often than many businesses hope. The R&D Tax Incentive supports experimental work to resolve genuine technical uncertainty, where the outcome could not be known in advance. Building a portal or business system with established tools usually does not qualify, while developing a new algorithm or a technique nobody has proven may. The government announced changes in the 2026-27 Budget that start on 1 July 2028, and the current rules apply until then. Talk to your accountant or an R&D tax adviser before you count on it.
How to keep the cost under control
Overruns are common. A study of 1,471 IT projects by Bent Flyvbjerg and Alexander Budzier of the University of Oxford found an average cost overrun of 27 percent, and one project in six overran by an average of 200 percent. Faster drafting does not change the habits that keep a project on budget.
- Pay for discovery first, so the build is priced from a specification rather than a guess
- Fix the price one stage at a time, so a surprise in stage two does not reprice the whole project
- Build the smallest version that delivers value, then add to it
- See working software every two weeks, so problems surface while they are cheap to fix
- Put every change through one written process with its cost agreed before work starts
What we would do
Start with a short discovery that produces a specification, a working prototype of the parts people will use and an estimate by stage. With agentic development the prototype comes early enough to test the idea before you commit to the build. Then fix the price for the first stage, keep a contingency for what discovery could not settle, and set aside a monthly budget for running the software once it is live. It is how we run every custom software project.


