Skip to content
Briefing 019 minute readCost and business case

Why most AI pilots never reach production

The research is consistent: the model is rarely what fails. What fails is everything around it. Here is what the numbers say, the five gates a pilot has to clear, and the questions a board should ask before funding the next one.

Tom BuckleySpurbuck Advisory
5 in 100 pilots deliver rapid revenue growthMIT NANDA, 2025

The short version

  • MIT's 2025 study of enterprise generative AI found about 5% of pilots deliver rapid revenue growth; most stall with little measurable effect on profit and loss 1.
  • In S&P Global's 2025 survey, 42% of companies abandoned most of their AI initiatives, up from 17% a year earlier, and the average organisation scrapped 46% of proofs of concept before production 2.
  • The causes named are the same across studies: poor data, weak risk controls, rising costs, unclear value and tools that never connect to how work actually gets done 1, 3.
  • Australia's Guidance for AI Adoption sets out six essential practices that map neatly onto what separates a pilot from a production system 5.
5%of generative AI pilots achieve rapid revenue acceleration 1MIT NANDA, 2025
42%of companies abandoned most of their AI initiatives in 2025, up from 17% 2S&P Global Market Intelligence
46%of proofs of concept scrapped before production, on average 2S&P Global Market Intelligence
30%of generative AI projects predicted to be abandoned after proof of concept by the end of 2025 3Gartner, July 2024

Read the numbers carefully, then take them seriously

Headline figures about AI failure travel fast and lose their footnotes on the way. The MIT figure comes from 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public deployments 1. It measures rapid revenue acceleration, which is a high bar. The S&P Global figures come from a survey of more than 1,000 organisations in North America and Europe 2. Gartner's figure was a prediction made in mid-2024 3.

So the exact percentages deserve care. The direction does not. Three different methods, from three different organisations, land in the same place: most AI work started in the last two years has not made it into the way the business runs, and the rate of abandonment rose sharply in 2025.

For a mid-sized organisation in Australia or New Zealand, that is useful information. It means the hard part is known, it is not the technology, and it can be planned for before money is spent.

Abandonment rose sharply in a yearShare of companies that abandoned most of their AI initiatives, by survey year 2.

Where pilots actually break

Gartner named four reasons projects are dropped after proof of concept: poor data quality, inadequate risk controls, escalating costs and unclear business value 3. MIT's researchers put it differently but arrive at the same place. The problem is not the quality of the models; it is a learning gap in how tools and organisations fit together, and integration that never happens 1.

In practice the failure looks ordinary. A pilot runs on a clean extract of data that nobody maintains. It sits in a separate tool that staff have to remember to open. No one has decided who signs off what the agent does, so legal or risk quite reasonably stops it. Usage costs climb with no one measuring the return. When the sponsor moves on, so does the pilot.

None of that is solved by a better model. It is solved by treating the pilot as the first release of a production system, with an owner, controls and a measure from day one.

A pilot proves the model can do the task. Production proves the organisation can live with it.

The pattern behind the numbers
The five gates between a pilot and productionEach gate matches a reason the research gives for abandonment. Clear them in the pilot, not after it.
PilotProduction12345
  1. 1An owner and a baseline. One named person accountable for the outcome, and a measure of how the work performs today, so value can be shown rather than asserted 3, 5.
  2. 2Data the agent can trust. Live access to the systems of record, not a one-off extract, with the gaps and duplicates known before the agent relies on them 3.
  3. 3Controls risk can sign. Least-privilege access, human approval for anything consequential and an audit trail of every action. This is what turns a no from risk into a yes 3, 5.
  4. 4Inside the tools people already use. The work arrives in Microsoft Teams, the CRM or the finance system, not in a new tab staff have to remember. This is the integration gap MIT describes 1.
  5. 5A running cost and a monitor. A known cost per transaction, a dashboard of what the agent did and how often a person corrected it, and someone who reads it 3, 5.

Buy, partner or build

One of the more practical findings in the MIT research concerns who does the work. Buying specialised tools and building through partnerships succeeded about 67% of the time; internal builds succeeded roughly a third as often 1.

The reason is not that internal teams are less capable. It is that a partner who has taken agents into production before brings the integration patterns, controls and operating habits that a first internal build has to discover. The best results come from combining the two: a partner who builds with the internal team and hands over something they can run.

Success rate by approachVendor and partnership approaches succeeded about 67% of the time; internal builds about a third as often, shown here as 22% 1.

Pilots that stall

  • Chosen because the technology is interesting
  • Run on a one-off copy of the data
  • Live in a separate tool or chat window
  • No one has defined what the agent may do alone
  • Success is described, not measured

Pilots that reach production

  • Chosen because a process is slow, costly or error-prone today
  • Connected to the live systems of record
  • Deliver work where staff already are
  • Permissions, approvals and audit agreed with risk up front
  • A baseline and a target set before the build starts

What this means in Australia and New Zealand

Adoption here is still uneven. The National AI Centre's SME AI Pulse found 44% of small and medium businesses using AI in February 2026. Among those not using it, around 65% pointed to distrust of AI decision-making or a preference for keeping people in control, and 19% said they simply did not know how to start 4.

That concern about control is reasonable, and it is also answerable. In October 2025 the Australian Government replaced the Voluntary AI Safety Standard with the Guidance for AI Adoption, which condenses ten guardrails into six essential practices: decide who is accountable, understand impacts and plan accordingly, measure and manage risks, share essential information, test and monitor, and maintain human control 5, 6. Read side by side with the five gates above, they are close to the same list.

New Zealand took a deliberately light-touch path. Its AI Strategy and MBIE's Responsible AI Guidance for Businesses, released together in July 2025, take a proportionate, risk-based approach with no AI-specific law, pointing organisations instead to existing obligations such as the Privacy Act 2020 and the Fair Trading Act 1986 7.

For a board on either side of the Tasman, the practical conclusion is the same. The controls that regulators describe are the same controls that get a pilot into production. Doing them early is not a compliance cost; it is the work that makes the investment pay back.

Where to start: the back office

MIT found more than half of generative AI budgets going to sales and marketing tools, while the clearest returns came from back-office automation 1. That matches what we see in mid-sized organisations: the payback is in reconciliations, reporting, compliance tracking, onboarding and the hand-offs between systems that staff currently do by copy and paste.

Those processes have three things in common. They are measured already, so a baseline exists. They run on structured systems, so the data is reachable. And the cost of an error is known, so the right level of human approval is easy to agree.

Questions to ask before you fund the next pilot

Six questions that separate a pilot built to be shown from one built to run. If any answer is unclear, that is the work to do first.

  1. Who is accountable for the outcome, by name, after the pilot ends?
  2. What does this process cost or how long does it take today, and how will we measure the change?
  3. Which live systems will the agent read from and write to, and who controls that access?
  4. What can the agent do on its own, and what needs a person to approve it?
  5. Where will staff see and use the result, and what do they stop doing?
  6. What will it cost to run each month at full volume, and who watches that number?

Want this done for your organisation?

Our AI Opportunity Assessment works through these questions against your own processes and systems, and ends with a costed, sequenced plan for the two or three automations most likely to pay back.

See the AI Opportunity Assessment

Sources

  1. Fortune, MIT report: 95% of generative AI pilots at companies are failing (18 August 2025), reporting MIT NANDA, The GenAI Divide: State of AI in Business 2025
  2. CIO Dive, AI project failure rates are on the rise, reporting S&P Global Market Intelligence (March 2025)
  3. Gartner, Gartner predicts 30% of generative AI projects will be abandoned after proof of concept by end of 2025 (29 July 2024)
  4. National AI Centre, AI adoption insights: December 2025 to February 2026, SME AI Pulse by Fifth Quadrant
  5. Department of Industry, Science and Resources, Guidance for AI Adoption (21 October 2025)
  6. Allens, Governance doesn't stand still: 9 FAQs on the Government's new Guidance for AI Adoption (November 2025)
  7. DLA Piper, Quick on the uptake: New Zealand's new strategic approach to artificial intelligence (July 2025), on the AI Strategy and MBIE's Responsible AI Guidance for Businesses released 8 July 2025

Questions people ask us

Why do most AI pilots fail to reach production?

Studies by MIT NANDA, S&P Global and Gartner point to the same causes: poor data quality, weak risk controls, rising costs, unclear business value and tools that never integrate with how work is done. The model itself is rarely the problem.

What percentage of AI projects are abandoned?

S&P Global Market Intelligence found 42% of companies abandoned most of their AI initiatives in 2025, up from 17% in 2024, and the average organisation scrapped 46% of proofs of concept before production. Gartner predicted at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025.

What are the six essential practices in Australia's Guidance for AI Adoption?

Decide who is accountable; understand impacts and plan accordingly; measure and manage risks; share essential information; test and monitor; and maintain human control. The guidance was published in October 2025 and replaced the Voluntary AI Safety Standard.

Start with a free discovery session.

Thirty minutes with Tom on how this applies to your organisation, your systems and the decisions in front of you. You leave with a clear view of what is worth automating, what it would take, and what it would return. No obligation.

We use the details you give us only to arrange and hold your session. See our privacy policy.