Only 31% of software projects finish on time, on budget and on scope; 50% are “challenged” and 19% fail outright — and those figures have barely moved in a decade. AI projects are worse: 70–85% never get past proof of concept. The causes are consistent and avoidable: unclear scope, no user buy-in, open-ended budgets, and — for AI — poor data and no plan to scale. Fixed scope, real user involvement, fixed pricing and staged delivery remove most of the risk.
Technology projects have a reputation for going over budget, over time, and under-delivering. The reputation is earned. The Standish Group’s CHAOS data for 2020–2024, summarised in a 2025 research update in the PM World Journal, puts the outcomes at 31% successful, 50% challenged, 19% failed. Compared with 2015–2017 the success rate rose two points and the failure rate didn’t move at all. For generative-AI projects the picture is bleaker still: RAND (2024) and Gartner (2025) put the share that never make it beyond proof of concept at 70–85%.
It’s tempting to blame the technology. In our experience — and in the research — the technology is rarely the real problem. The pattern behind failed projects is consistent, and the fixes are simpler than you’d expect.
Cause 1: Scope that was never really pinned down
Poor requirements gathering and scope creep sit at the top of every failure study. Projects that start with “we’ll figure out the details as we go” tend to sprawl. Requirements creep, timelines slip, and costs balloon. The fix is to scope the work properly before building — define what success looks like, in numbers, and agree it in writing. This is why we scope every project in full during discovery and put it on a fixed price before any build begins: if the scope can’t be written down, it can’t be built.
Cause 2: The people who’ll use it weren’t involved
Software built in isolation from the people who’ll actually use it gets quietly abandoned — the “never used” slice of that 19%. If the front-line team doesn’t find it faster than what they did before, they’ll go back to the spreadsheet. The fix is to design for real users and involve them throughout: weekly demos, testing against real cases rather than a polished sample, and a launch plan that includes training and a period of side-by-side running. In Datacom’s 2026 index, employee concern about AI was the fastest-growing barrier in NZ organisations (18%, up 7 points) — user involvement is how you turn that concern into adoption.
Cause 3: Open-ended budgets and hourly billing
When you’re billed by the hour with no ceiling, there’s no shared incentive to be efficient, and no certainty for you. Half of all projects come in “challenged” — late or over budget — and open-ended commercial terms are a big reason why. Fixed pricing aligns everyone: the scope is agreed, the price is agreed, and surprises are the supplier’s problem, not yours.
Cause 4 (AI projects): bad data and no path from pilot to production
The 2025 research identifies a new set of killers for AI projects: data quality, legal and ethical obstacles, running (inference) costs, and the inability to scale past a proof of concept. Locally, Datacom found 81% of NZ organisations are still in exploratory or implementation stages, and only 15% have scaled AI organisation-wide. The pattern we see: a demo that works on ten hand-picked examples, then collapses on the messy reality of a thousand real invoices or phone calls. The fix is to test on real data from day one, agree the operating cost per month before build, and design the hand-off to humans, the monitoring and the guardrails as part of the first version — not as a phase two that never comes.
The technology almost never fails on its own. Projects fail when scope, people and budget aren’t handled with discipline up front — and AI projects fail when nobody planned for the day after the demo.
How to avoid being a statistic
- Write the success metric down first. “Invoice processing from 3 days to under 4 hours” is a project; “improve efficiency” isn’t.
- Fix the scope, then fix the price. If a supplier won’t commit to either, the risk is yours.
- Involve the people who’ll use it every week. Demo real work, not slides.
- Deliver in stages. A working slice in weeks beats a big-bang in a year — and it lets you stop early if the value isn’t there.
- For AI: test on real data, budget the running cost, and design the human hand-off up front.
Whether you’re planning a new system, a digital transformation, a piece of custom software or an AI agent, the safeguards are the same. If you want an independent view before you commit, that’s exactly what our consulting is for — and every Avtrix project is delivered fixed-scope, fixed-price, in stages, for the reasons above.
Sources
- Arcidiacono, G., Comparative research on IT project failure rates: a 2025 update, PM World Journal, January 2026 — summarising Standish Group CHAOS 2020–2024, RAND (2024) and Gartner (2025)
- Datacom, 2026 State of AI Index — NZ AI maturity and barriers
Last updated 15 September 2026. Written by Ashok Poshamalla, Founder, Avtrix Software Solutions, Taupō.