The Business of AI Adoption: Why the Hard Work Hasn't Changed

The more things change

In 2017, we wrote about the gap between AI excitement and AI readiness. At the time, vendors like Salesforce, Google, Amazon, and Microsoft were embedding machine learning into their platforms and making bold claims about what was now possible. The technology was genuinely impressive. But the practical guidance for organisations wanting to move beyond experimentation was remarkably thin.

The argument then was straightforward: successful AI adoption is not primarily a technology challenge. It is a business discipline that demands clarity of purpose, data readiness, thoughtful integration, and a willingness to invest in the unglamorous work of preparation before chasing results.¹

Nearly a decade later, the technology landscape has transformed almost beyond recognition. Large language models (LLMs) can generate text, code, and analysis that would have seemed like science fiction in 2017. Generative AI tools are available to anyone with a browser. Yet the fundamental challenge remains stubbornly unchanged: most organisations still struggle not with accessing AI capabilities, but with turning those capabilities into sustained business value.

The accessibility trap

The rise of LLMs and generative AI has dramatically lowered the barrier to experimentation. Where building a machine learning model once required data scientists, specialised infrastructure, and months of development, a business analyst can now prompt a foundation model and receive sophisticated outputs in seconds. This is a genuine revolution in accessibility.

But it has created a dangerous illusion: that because AI is easy to try, it must be easy to adopt. The distinction matters. A proof of concept that impresses a boardroom is not the same as a capability embedded into business operations, governed appropriately, and delivering measurable returns. Deloitte's research into enterprise AI adoption found that while organisations are investing heavily — with 53 percent spending more than US$20 million annually on AI — fewer than half reported having a high level of skill in selecting AI technologies and suppliers. More telling still, 56 percent agreed that emerging risks were actively slowing their adoption efforts.²

Stanford's 2026 AI Index Report reinforced this picture, identifying "a widening gap between what AI can do and how prepared we are to manage it."³ The technology is advancing faster than the organisational capability to absorb it responsibly.

Four fundamentals that still hold

The planning framework we outlined in 2017 was built around four critical aspects of AI adoption. What is remarkable is how durable these foundations have proved, even as the technology itself has been reinvented.

Clarify your purpose

Before selecting a tool, an organisation must define the question it is trying to answer or the outcome it is trying to achieve. In 2017, this meant establishing which business problem warranted a machine learning approach. Today, with LLMs capable of generating plausible answers to almost any question, the risk of skipping this step is even greater. The ease of getting an output masks the difficulty of knowing whether that output is the right one, or whether AI was the right approach at all.

Purpose-driven adoption means starting with the business problem — not the technology. What decision are you trying to improve? What process are you trying to transform? What capability gap are you trying to close? Without this clarity, organisations end up with a portfolio of impressive demonstrations and no measurable impact.

Understand your data

In 2017, data readiness meant locating relevant datasets, restructuring them appropriately, and ensuring quality. The principles of tidy, well-governed data remain essential, but the landscape has expanded. LLMs introduce new data considerations: what proprietary knowledge should be made available through retrieval-augmented generation (RAG)? How do you prevent sensitive information from leaking through model interactions? How do you maintain data lineage when AI is synthesising across multiple sources?

Organisations with mature data governance have a compounding advantage. Those without it find that generative AI amplifies their data problems rather than solving them — surfacing inconsistencies, hallucinating from poor-quality sources, and creating outputs that cannot be traced back to authoritative information.

Choose the right approach

The original framework focused on selecting appropriate algorithms — distinguishing between deep learning, classical machine learning, and their respective applications. Today, the decision landscape is broader but the principle is identical: different problems demand different approaches.

Not every business problem needs a large language model. Some are better served by traditional analytics, rules-based automation, or classical machine learning. An LLM that generates customer service responses is a different proposition from a predictive model that forecasts equipment failure. Choosing the right approach requires understanding the trade-offs: cost, accuracy, explainability, latency, and risk tolerance. The organisations that treat every problem as a prompt engineering exercise will waste resources and introduce unnecessary complexity.

Plan for integration

A model that works in isolation is a prototype. Value is created when AI capabilities are integrated into existing business processes, systems, and decision-making workflows. In 2017, this meant navigating API integrations and agile development cycles. Today, it also means managing prompt chains, embedding models into enterprise architectures, establishing human-in-the-loop oversight, and designing for graceful degradation when AI outputs are uncertain or incorrect.

Integration also now encompasses governance: how outputs are reviewed, how decisions are audited, how bias is monitored, and how the organisation maintains accountability for AI-assisted outcomes. These are not afterthoughts — they are architectural requirements that must be designed in from the start.

What has genuinely changed

While the fundamentals endure, the context around them has shifted in ways that demand attention.

The governance imperative is no longer optional. In 2017, AI governance was a forward-looking concern for regulators and academics. Today, it is a board-level requirement. The EU AI Act, NIST's AI Risk Management Framework, and Australia's own voluntary AI Ethics Principles all signal a regulatory environment that expects organisations to demonstrate responsible AI practices — not just good intentions. Organisations adopting AI without a governance framework are building on foundations that will need to be retrofitted, at significant cost.

The talent equation has inverted. The scarcity in 2017 was data scientists and machine learning engineers. Today, foundation models have commoditised much of the technical capability, but the scarce resource is people who can bridge business strategy and AI implementation — those who understand both what the technology can do and what the organisation actually needs. Deloitte's research confirms this: organisations consistently report that the ability to integrate AI into existing business processes and evaluate AI technologies are among their most significant capability gaps.²

The cost of getting it wrong has escalated. When AI was confined to back-office analytics, a failed experiment was a sunk cost. When AI is generating customer-facing content, making lending decisions, or summarising medical records, the consequences of hallucination, bias, or data leakage are material — reputational, financial, and regulatory. The speed at which generative AI can produce outputs at scale means that errors, too, can scale.

From experiment to enterprise

The path from AI experiment to enterprise capability follows a pattern that has been consistent across every wave of technology adoption: the organisations that succeed are those that invest disproportionately in the design and planning phases. As uncertainty decreases through structured iteration, execution accelerates and outcomes compound.¹

This means treating AI adoption with the same rigour applied to any significant business transformation:

• Start with a business case, not a use case — define the measurable outcome before exploring the technology
• Assess organisational readiness honestly — data maturity, governance capability, integration capacity, and change management readiness all determine whether AI will deliver value or create new problems
• Design for production from day one — prototypes that cannot be governed, scaled, or maintained are not stepping stones; they are dead ends
• Build the governance framework alongside the capability — not as a compliance exercise after deployment, but as an integral part of the design
• Invest in the people who bridge the gap — between what the technology can do and what the business needs it to do

The enduring lesson

The technology will continue to evolve — rapidly and unpredictably. What will not change is the fundamental truth that successful adoption is a business discipline, not a technology project. The organisations that recognised this in 2017, when machine learning was the frontier, have a structural advantage today as generative AI reshapes the landscape. They have the data governance, the integration thinking, and the organisational muscle to absorb new capabilities without starting from scratch.

For those still at the beginning of the journey, the good news is that the playbook is well established. The hard work has not changed. But it remains, stubbornly, the work that matters most.