TL;DR: In 2025, AI adoption has hit a “Failure Paradox.” While 71% of organizations are experimenting, nearly 95% of pilots fail to reach production. This article explores the staggering cost of “shoehorning” AI into businesses without a strategy—from A$439,000 hallucinations to legal precedents in customer service—and provides a roadmap for adoption led by subject matter experts, not just marketing hype.

AI is the only technology in history that leaders are rushing to ‘shoehorn’ into their businesses before they’ve figured out what it’s actually for. Like buying a solution to a problem you’re not sure you really have.

It’s easy to simply start implementing AI into your operational model. With enough dedication it can be shoehorned into nearly every process, system, and workflow that makes up your enterprise but would it be meaningful?

AI can be costly not just in terms of the outlying costs but there’s also the indirect costs to be considered, after all employees need to be upskilled in a tool (that could potentially never deliver an ROI).

Then there’s the potential for costly mistakes.

Examples of these mistakes

Deloitte - Australian Government 2025

For example, in 2025, Deloitte was contracted by Australia’s Department of Employment and Workplace Relations for A$439,000 to produce an “independent assurance review” of their welfare compliance system.

The report was riddled with AI hallucinations including:

• Multiple citations to nonexistent academic reports
• Fake references supposed from University of Sydney and Lund University in Sweden
• A completely fabricated quote from a federal court judgment in a robodebt case

A researcher at Sydney University flagged the errors to media and government officials.

After being caught, Deloitte admitted using Azure OpenAI GPT-4o to help produce the report. This led to them refunding a reported amount of A$97,000.

But the damage doesn’t end there. Some reports have suggested that this prompted other clients to review their reports from Deloitte, which has resulted in refunds “in the millions of dollars” for similar issues. Not to even begin to fathom the reputational damage it’s done to their brand.

The irony? This happened amid Deloitte’s $3 billion AI investment plan through 2030.

Air Canada Chatbot (2022-2024)

In November 2022, Jake Moffatt needed to fly to his grandmother’s funeral. Upon chatting to Air Canada’s chatbot, it told him that he could claim a bereavement discount within 90 days after travelling. He booked $1,400 CAD in flights.

Only, on applying for the discount with the death certificate, Air Canada denied it. Their actual policy requires requesting bereavement fares before travel. They further argued that the chatbot was “a separate legal entity responsible for its own actions”. A tribunal would go on to call this a “remarkable submission”.

Ultimately, the British Columbia Civil Resolution Tribunal ordered Air Canada to pay $812.02 in damages, further ruling that “It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.”

A legal precedent was also set that companies are liable for AI chatbot information and ultimately, in April 2024, the chatbot was removed from their website.

So, although the pay for damages seems minor when compared to the Deloitte case, you have to factor in the reputational damage again, as well as the costs involved in first rolling out this chatbot only to retract and remove it again 2 years later.

Neither of these examples is meant to scare anyone away from adopting AI, but rather to highlight the gap between AI deployment enthusiasm and proper verification systems.

The AI Reality Check: 2025 by the Numbers

While the marketing hype suggests a seamless transition to an automated world, the data tells a story of “Pilot Purgatory”—a state where many projects start, but very few cross the finish line.

1. The Adoption Paradox

There is a massive disconnect between individual experimentation and enterprise-grade integration.

• The Surge: 71% of organizations now use Generative AI regularly, and the global user base has hit 378 million.
• The Friction: Despite that surge, 9 out of 10 US businesses still report no formal AI integration at the firm level. We are seeing a “Shadow AI” movement where individuals use tools that the enterprise hasn’t officially mastered.

2. Why “Pilot Purgatory” is Getting Crowded

The “Failure Rate” of AI is nearly double that of traditional software projects.

Key Insight: Only 25% of AI initiatives are delivering the ROI expected by the C-Suite. This is largely because 62% of leaders still struggle with the “Data Basement”—poor data access and integration.

If anything is to be gleaned from the above, it’s that there is an ROI to be had with AI adoption if you approach it with critical thinking and a better understanding of the technology itself.

3. What the “Successful 5%” Are Doing Differently

Success isn’t accidental; it follows a specific pattern of behavior.

• Buy over Build: Organizations partnering with external vendors see a 66% success rate, doubling the odds of those trying to build in-house.
• Back-Office First: The fastest ROI isn’t found in flashy customer-facing bots (like Air Canada), but in “boring” functions: operations, finance, and procurement.
• Transformation over Deployment: Winners treat AI as a business model change, not just a software update.

Accountability

Accountability is probably the biggest detractor for AI enthusiasm. Obviously, there is a lot of concern around people losing their jobs to AI but the reality boils down to accountability, and let me tell you, AI has no skin in your game.

When you ask an employee to produce output, they are responsible for that output. They understand that, and they realise that they are receiving a salary based on that accountability.

AI, as it stands, has zero accountability for its output. Anthropic, OpenAI, Google, etc., will absolutely laugh at you if you tell them you have lost revenue because of a mistake their specific model has made. They skirt responsibility for the output from the very start, after all.

What I’m trying to say is that wherever you adopt AI, there will be a human being who is held accountable for the output, even if that person is you.

Individual users

It could stand to reason that any user who understands a technology, including its shortcomings, would benefit more from it than one who doesn’t. We see this already with products like Photoshop, Office, Kubernetes, etc.

We call this the power-user feedback loop: Users with stronger foundational capabilities (prompt engineering, data literacy) acquire higher-quality information, which makes them more inclined to continue using the tools, further sharpening their cognitive skills. It’s a similar passion we’ve seen when comparing passion-driven engineers to career-driven engineers (though the two are not mutually exclusive).

The reason I raise this is to highlight a single (and very effective) way to meaningfully adopt AI into your organisation. Rely on subject matter experts.

These experts will happily identify where and when to introduce AI into your delivery pipelines.

Case Study: The SuperGroup Approach

At SuperGroup, our AI drive started with engineers stepping forward to showcase what they had achieved with GitHub Copilot. These same engineers keep stepping up, showcasing demos on things like MCP, AI guiderails, etc. Ultimately, they are increasing our adoption and showing a strong potential for ROI on the technology.

On the flip side, where we’ve had companies market their own versions to us, we’ve looked at it skeptically. Even when trying one out, we’ve seen complete failure, often down to lacklustre use of the product.

Understanding the human impact of AI adoption

A study published in PubMed Central (2025) explored AI literacy through the lens of Self-Determination Theory. It found that users with higher AI literacy didn’t just work faster; they felt better.

That being said, there is a flip-side when AI is seen to do creative tasks like generating artwork, or used as a communication medium. No one wants a GPT-written document from HR, afterall. AI can simulate intelligence, but it cannot simulate empathy.

If I look at engineering in isolation, where AI shines is in allowing engineers to focus on what they enjoy most, while avoiding the parts of their job that bring less job satisfaction (writing boilerplate code and unit tests, for instance).

If we can focus on this aspect of the adoption, we could actively increase job satisfaction for our employees.

What it boils down to

It’s easy to get lost in the marketing. Microsoft has been pushing Copilot on all of us, hard. But one person’s decision isn’t going to encourage everyone in the organisation to adopt it. You need to find the productivity and ROI link in the noise.

One way to do so is to follow the passionate adopters, who are willing to point out when a product is simply not worth the time.

Another approach is to give employees strategic training in the matter. A 2025 report by Thomson Reuters highlights a widening gap between general optimism and strategic mastery. This report highlights that professionals who know how to effectively use AI have projected time savings of up to 5 hours per week.

This is critical when attempting to redesign a job to incorporate AI.

Final Verdict

AI, as it stands, is not a magic solution to all your problems.

It won’t write accurate financial reports for you, but it could help you research topics for the reports and potentially give your report a once-over to ensure you’re using optimal language based on your audience.

It won’t produce production-level code for you, but it could ease the Software Development Life Cycle (SDLC) your engineers have to follow and help in producing better code, highlighting potential issues, training them in new methodologies, etc.

AI art is almost instantly identified and has shown to damage brands and reputations. It won’t replace your graphics designers, but it could help them come up with concept sketches which they can take to fruition.

I don’t buy into all the hype around AI; there is a clear agenda behind it, but I don’t discredit it either. There is immense value to be had, provided you stop looking for a magic wand and start looking for a better tool for your experts.

The Appendix

Just for full transparency on the state of the market, here are the aggregated benchmarks referenced in this piece:

Adoption Rates:

• Over 80% of companies are using or exploring AI in 2025
• 65-71% of organisations regularly use generative AI in at least one business function (doubled from previous year)
• Individual AI user base reached 378 million people worldwide in 2025
• However, at the firm level, 9 out of 10 US businesses report NOT using AI (showing the individual vs. enterprise adoption gap)

The Failure Paradox:

• 42% of companies abandoned most AI initiatives in 2025 (up sharply from 17% in 2024)
• 88-95% of AI pilots fail to reach production (varies by study)
  • MIT NANDA: 95% failure rate
  • IDC: 88% don’t scale (4 out of 33 pilots reach production)
  • Average enterprise scrapped 46% of AI proof-of-concepts before production
• Over 80% of AI projects fail (double the failure rate of non-AI tech projects)
• Only ~25% of AI initiatives deliver expected ROI

Key Challenges:

• Data Issues - 62% of leaders cite data access and integration as top obstacle
• Organisational Readiness - Lack of skills, processes, and infrastructure
• The “Pilot Purgatory” Problem - Proof-of-concepts abound but scaled success stories are rare
• Change Management - Only 1/3 of companies prioritize training and adoption support
• Cost & Unclear ROI - Expectations for productivity gains are underdelivering
• Integration Complexity - Moving from sandbox to production requires different skills

What’s Working (The 5%):

• 31% of use cases reached full production in 2025 (double from 2024)
• Organizations partnering with external vendors: 66% success rate vs. 33% for internal builds
• Companies that treat AI as business transformation (not just tech deployment) succeed
• Focus on back-office functions (operations, finance, procurement) shows faster ROI than customer-facing use cases
• Agile delivery organizations strongly correlated with achieving value

Emerging Trends:

• Shift from “generative AI” to “agentic AI” - systems that can act autonomously
• 23% already scaling agentic AI, 39% experimenting
• Focus moving to composable architectures to avoid vendor lock-in
• Trust, governance, and compliance now board-level concerns