Why ERP projects blow out, from someone who was paged when they did

Photo: rawpixel via rawpixel (CC0 1.0)
Queensland Health went live with a new payroll system in March 2010 and spent the next several years paying people wrong. Tens of thousands of staff were underpaid, overpaid or not paid, the fix involved hundreds of extra staff processing pay by hand, and a commission of inquiry led by Richard Chesterman QC reported in 2013 that the total cost to the state would run to somewhere around $1.2 billion over eight years. It is the canonical Australian enterprise software failure, it gets taught in project management courses, and every large ERP programme I’ve worked on since has had at least one person in the room who was quietly certain it couldn’t happen to them.
It can. The failures repeat, and they repeat for reasons that have very little to do with the software.
The process was never written down #
Every ERP implementation begins with a promise that the organisation will “adopt standard processes”. That means the way the vendor’s software expects the work to be done, rather than the way the organisation actually does it. It’s the right promise. It’s also the one that gets broken first, because when the analysts sit down with the payroll team, or the warehouse, or the accounts payable clerks, they discover that the real process exists in the heads of eleven people, half of it contradicts the policy manual, and three of the eleven are retiring next year.
The Chesterman inquiry’s findings on Queensland Health were, at their core, about this. The award structure for health workers was extraordinarily complex, the existing processes for handling it were poorly documented, and the new system was expected to encode rules that nobody could fully state. The software did what it was told. The problem was what it was told.
The check: before configuration starts, can someone walk you through the ten highest-volume processes end to end, on paper, with the exceptions? If the answer is a workshop that runs over by a day, you’re not ready, and the go-live date is a fiction.

Customisation is debt with a grace period #
The second promise that gets broken is “minimal customisation”. Someone senior can’t live with the standard screen, or a legacy report needs to look exactly as it did, and a developer writes a modification. Then another. Each one is small. Each one has to be retested, re-migrated and often rewritten at every upgrade, and the sum of them is why organisations end up running a decade-old version of their ERP with a support contract that costs more than the licence.
I ran systems for a freight company where the ERP had 340 custom objects by the time I arrived, and about 60 of them did things nobody could explain. We spent eight months just finding out which ones were still used before we could touch an upgrade. The vendor’s standard process would have handled most of them. Someone in 2008 had wanted a button.
The check: keep a register of every customisation with the name of the person who asked for it and the business reason. If the reason is “that’s how we’ve always done it”, it’s a training issue, not a development ticket.
Data migration is the project, not a phase of it #
Every plan has a box called “data migration” somewhere in the middle. In practice it is most of the work and most of the risk. Thirty years of customer records with six different address formats. Product codes that were reused. Employees who exist twice. The old system tolerated all of it because it grew up with it; the new one won’t, and the cleansing has to happen before the load, which means someone has to decide, record by record, what the truth is.
The organisations that do this well start the data work on day one, run it in parallel with everything else, and do at least three full trial migrations before the real one, comparing the totals every time. The ones that don’t discover on the first payroll run that 4,000 employees have no bank account on file. That’s not a hypothetical. That’s a Tuesday I remember.
Testing that tests the demo, not the business #
User acceptance testing tends to be scheduled for the last six weeks and then compressed into three because everything before it ran late. The test scripts cover the happy path, because those were written from the design documents, and the design documents describe how the process is supposed to work. Nobody tests the fortnight where the award rates change halfway through a pay period and a nurse worked a public holiday across a shift boundary while on higher duties. Then the system goes live in exactly that fortnight.
The check: give the test scripts to the most cynical person in each business team and ask them to break it. Pay them for the time. Their scenarios are the ones the analysts didn’t think of, and they are the ones that will page you at two in the morning after go-live.
The go-live that couldn’t be stopped #
The single most common thread in the post-mortems I’ve read, from Queensland to the well-publicised overseas disasters at councils and retailers, is a go-live date that had become politically impossible to move. The old system’s support was expiring. The executive had announced it. The budget for the next year assumed it. Everyone in the room knew the readiness criteria hadn’t been met, and everyone went anyway.
A go/no-go decision only means something if “no” is a real option with a real plan behind it: an extension of the old system’s support, a budget contingency, an executive who’d rather explain a delay than a failure. Build that option before you need it. The Chesterman report is worth reading on exactly this point, and it’s available in full from the Queensland government.
The integrator’s incentives are not yours #
Most large ERP programmes are delivered by a systems integrator on a time-and-materials basis, or on a fixed price with a change-request process that quickly becomes time-and-materials in disguise. Neither structure rewards the integrator for telling you that your processes aren’t ready, that a customisation is unnecessary, or that the go-live date should slip. All three of those conversations reduce billable hours. I’ve worked with excellent integrators who had them anyway, and I’ve worked with ones who took the change request and said nothing, and from the outside the two looked identical until about a month after go-live.
The structural fix is to have someone on your side of the table whose only job is to disagree with the integrator: an independent programme assurance role, reporting to the steering committee rather than the programme director, with the standing to say “not ready” and be listened to. Governments have got better at this after their own failures; the Queensland inquiry recommended exactly this kind of oversight. Private companies mostly still don’t do it, because it looks like a cost, and it’s a cost that’s a rounding error against the number in the first paragraph of this article.
The second fix is contractual. Tie a meaningful part of the integrator’s fee to outcomes measured after go-live: pay run accuracy for three months, order-to-cash cycle time, the number of manual workarounds in use at day 90. Integrators resist this, reasonably, because a lot of the outcome depends on the customer. Agree the measures anyway. What gets measured at day 90 gets attention at day minus 90.
The training that gets cut #
End-user training is the last line in the budget and the first one cut when the programme runs over, and the consequence arrives on day one of go-live as a support queue that never clears. A payroll clerk who’s been doing the job for fifteen years on the old system and gets a two-hour webinar on the new one will find a way to make the new one behave like the old one, usually through a workaround that quietly breaks the data the system depends on. I watched a warehouse team keep a parallel spreadsheet for eleven months after go-live because nobody had shown them where the stock adjustment screen was. The ERP reported one number, the spreadsheet another, and the spreadsheet was right.
What good looks like #
It’s not glamorous. It’s a programme that spends the first quarter documenting what actually happens, says no to nine in every ten customisation requests, treats data migration as a workstream with its own manager, tests with the cynics, and has a written plan for not going live. The federal government’s Digital Transformation Agency publishes guidance on managing large digital investments that says most of this in more formal language, and it’s the better for being boring.
I’d add one opinion. The industry’s habit of blaming the vendor, or the integrator, after a failure is mostly a way of not reading the findings. The software is rarely the fault. The organisation that couldn’t describe its own payroll was. That’s uncomfortable, and it’s the reason it keeps happening.
Anyway. Go and find out how your pay is actually calculated. Ask three people. If you get three answers, you’ve found your next project’s biggest risk, and its not in the software.
— Anjali Mehta, Enterprise & Cloud Editor, Technology Spectator