Why government IT projects fail in the same three ways

Photo: Scott Webb via rawpixel (CC0 1.0)
The Royal Commission into the Robodebt Scheme handed down its report in July 2023, and among its findings about an unlawful debt-recovery programme was a quieter observation that anyone who has covered government technology for two decades would recognise: the system did what it was designed to do, the design was wrong, and the people who could have said so either weren’t asked or weren’t heard. I’ve read the audit reports on a great many federal and state technology programmes since the early 2000s, and the failures cluster. Not into a hundred causes. Into three.
One: the policy was decided before anyone asked whether it could be built #
Robodebt is the extreme case. The policy, averaging annual income across fortnights to raise debts automatically, was adopted, budgeted and announced before the question of whether the calculation was lawful or accurate had been properly answered, and the technology was then built to execute it. The royal commission’s report, available in full from the commission’s site, describes a public service in which the people who understood the data and the people who made the policy were in different rooms and the door between them was shut.
The milder, more common version is the budget measure. A minister announces a programme with a start date, the date is in the budget papers, and the agency is now building a system to a deadline it didn’t set, to a scope that was written by a press secretary. The systems that work in government, and there are more of them than the headlines suggest, tend to be the ones where the technology people were in the policy room early enough to say “that will take three years and here’s why”. The ones that fail tend to be the ones where they were told the date.

Two: nobody owned the whole thing #
The myGov user audit, chaired by David Thodey and released in early 2023, looked at a platform that tens of millions of Australians use and found that it had been built up over a decade by a series of agencies, each responsible for its own piece, with no single owner accountable for whether the thing worked as a whole. It recommended, among other things, that myGov be treated as critical national infrastructure with dedicated funding and a clear owner. That’s a governance finding, and it’s one you can find in some form in every major review of a federal digital programme going back to the Gershon review in 2008.
The structural reason is that government programmes are funded by measure, delivered by agency and reviewed by portfolio, and a system that crosses all three, which is every useful system, falls between them. The Digital Transformation Agency exists in part to solve this, and its investment oversight framework is the current attempt to make someone accountable for a programme’s whole life rather than its first budget cycle. Whether it has the teeth is a question I ask every year and the answer is improving slowly.
Three: the contract was the plan #
State payroll and health record projects, of which Queensland Health’s payroll remains the most expensive, share a pattern that my colleague Anjali describes from the inside in her piece on why ERP projects blow out: a large integrator contract signed early, a scope that changed, and a government that had, by the time it understood the problem, no practical way to walk away. The contract becomes the plan. The plan can’t be changed without a variation. The variations are where the money goes.
The fix that the audit offices keep recommending, and that the Commonwealth’s digital sourcing rules now encourage, is smaller contracts, shorter terms, and an internal capability that can judge whether the vendor’s work is any good. That last part is the hard one. Two decades of outsourcing left many agencies without the engineers to evaluate what they were buying, and rebuilding that capability is slow, unglamorous and competes with every other budget line. The DTA’s Digital Marketplace and the panels that succeeded it were meant to make buying easier. They did. Buying well is a different skill.
The state version of the same three #
The states have their own versions and the audit offices in Sydney, Melbourne and Brisbane have documented them just as thoroughly. Victoria’s HealthSMART programme in the 2000s, the New South Wales LMBR schools finance system, Queensland’s payroll: each of them appears in a state audit report with a version of the same findings. Ownership that moved between agencies as ministers changed. A go-live date fixed by a budget or an election. A contract with an integrator that outgrew the department’s ability to manage it. The states have also had the successes, and for the same reasons; Service NSW’s digital platform was built by a team that stayed together for years under an agency that was allowed to own it, and it shows.
The one distinctly state-level failure mode is the shared-services programme, where several agencies are consolidated onto one platform to save money. These are announced with savings figures, run for years, and end with a report explaining why the savings didn’t arrive. The reason, almost always, is the first of my three: the policy decision to consolidate was made before anyone asked whether the agencies’ processes could be reconciled, and they couldn’t.
What the successes look like #
They exist, and they’re worth naming, because the failure stories give a distorted picture. The tax office’s online services, for all the complaints, process the country’s tax returns at a scale and reliability that most private companies would envy, and did so through a decade of platform rebuilds. The electronic conveyancing platform that moved property settlement online in most states has run for years without becoming a headline. The Bureau of Meteorology’s data services, the passport system, the electoral roll. Each of them had a long-lived owner, an internal team that understood the system, and a scope that was allowed to be boring.
The common thread is patience. The successful programmes were built over years by people who stayed, on platforms that were maintained continuously rather than replaced in a single heroic project every decade. That’s the opposite of the budget-measure model, and it’s why I’d argue, mildly, that the biggest technology reform available to any government is a funding rule that pays for systems the way it pays for roads: as assets with a maintenance line, not as projects with an end date.
There’s a fourth ingredient in the successes that’s easy to miss because it’s an absence: nobody announced them. A system that’s built to be boring, on a platform that’s maintained rather than replaced, doesn’t get a launch event, and a programme without a launch event doesn’t get a launch date, and a programme without a launch date can take the time it needs. The single most reliable predictor of a government technology failure in my files is a media release with a month in it.
The accountability gap #
The royal commission recommended, and the government accepted, a legislative framework for automated decision-making: a set of rules about when a computer can decide something that affects a person’s entitlements, what oversight applies and how a person can challenge it. That’s the single most important structural change to come out of the whole affair, because it puts the question “should this be automated” ahead of “can it”, which is the reverse of the order in which Robodebt was built. The framework has been slow to arrive. The pressure to automate, particularly with the tools my colleague Sophie covers in her piece on what’s actually deployed, is not slow at all, and the gap between the two is the thing I’d watch.
What I’ve learned from reading the reports #
I’ve sat in Senate Estimates and watched a departmental secretary explain a $100 million overrun as a “rephasing”, and I’ve read enough Australian National Audit Office reports to know that the auditors find the same things every time: unclear ownership, optimistic assumptions, contracts that outran the capability to manage them. The reports are public, they’re well written and almost nobody outside the public service reads them. If you want to understand why the next big government system will run late, the answer is already on the ANAO’s website, in a report about the last one.
The Robodebt report changed one thing that the audit reports never could, which is that it made named individuals answerable for a system’s design. Whether that changes behaviour across the service is a question for the next decade. My guess, for what it’s worth, is that it changes the behaviour of the people who read it, and the job of the rest of us is to make sure the next cohort of secretaries do.
I’ll admit I read the full 990-odd pages of the commission’s report over a long weekend in 2023 instead of doing anything useful with a radio chassis I’d promised to finish. The radio’s still in bits. The report was worth it.
— Marcus Tully, Editor, Technology Spectator