The AI honeymoon is over. Now businesses must make it work

After three years delivering AI training and workshops across New Zealand and Australia, I have seen a consistent pattern emerge. AI adoption is rarely linear.

Usage often rises rapidly when people first discover ChatGPT, Microsoft Copilot or Claude. The early gains can be immediate: emails are drafted faster, meetings summarised, long documents reviewed, research accelerated and presentations developed in a fraction of the time.

Then adoption often dips.

Users encounter hallucinations, inconsistent answers and features that do not always perform as impressively as the demonstration they watched online. Expectations have been fuelled by vendor marketing, social media and carefully constructed examples showing AI operating at its best.

The reality is less dramatic, but more useful. Copilot, ChatGPT and Claude can be exceptional at some tasks and average at others. Mature users know how to provide context, select the right tool and recognise when an answer needs to be challenged.

When organisations invest in practical AI training and workshops, adoption often rises again. The second wave is based on capability rather than novelty.

Many organisations are moving too quickly past the basics. Agents and automation dominate the conversation, while substantial gains remain available from helping individuals and teams use existing productivity tools properly.

For organisations operating primarily in Microsoft 365, Microsoft Copilot is a natural place to start. Used well across Outlook, Teams, Word, PowerPoint and Excel, it can reduce time spent on everyday knowledge work almost immediately.

Meeting preparation, summarisation, drafting, document comparison, research and reporting are not glamorous AI use cases. Multiply small gains across hundreds of activities and the impact becomes material.

New Zealand's AI Forum reported in August 2025 that 91% of surveyed businesses were seeing efficiency gains from AI, while half reported a positive financial impact. The opportunity is real. The challenge is making those gains repeatable rather than leaving them with a handful of enthusiastic users.

The progression I encourage organisations to think about is straightforward: individual productivity, team productivity, then better workflows, then broader organisational productivity.

Governance should guide that journey, not chase it.

In many organisations, employees started using generative AI before formal governance existed. Policies are now being developed around behaviours that are already embedded. I regularly see staff using free or unapproved AI tools because they solve an immediate problem faster, while business leaders have limited visibility of what information is being entered into them.

Simply banning those tools rarely fixes the problem. It can just make the usage less visible.

Good governance should make safe adoption easier. Employees need clear answers: Which tools are approved? What information can be used? When must an answer be verified? Who is accountable? Where is experimentation encouraged?

That clarity gives people confidence to use AI productively without creating unnecessary security or compliance risk.

It also helps address another growing problem: too many tools.

Businesses can quickly accumulate Copilot, ChatGPT, Claude and specialist AI applications with overlapping functionality. Every additional platform brings another security consideration, governance requirement, licence cost and learning curve.

Most organisations would gain more from mastering one or two core platforms. For a Microsoft-based organisation, Copilot may naturally sit at the centre of everyday productivity, while Claude or ChatGPT may be stronger for particular work.

The useful question for IT leaders is not, "Which AI is best?" It is, "Which tools should our people use for which work, with which data, under what conditions?"

Experience matters. AI gives less experienced employees access to capabilities that once took much longer to acquire, but access to an answer is not the same as having the experience to assess it.

An experienced engineer, accountant, consultant or manager can read a polished AI response and recognise that an assumption is wrong, a constraint has been overlooked or the recommendation would not work in practice. They challenge it, refine it and use AI as a thinking partner.

Someone without that domain experience may see the same polished response and assume it is correct. That is one route to AI slop: professional-looking material that lacks accuracy, insight or judgement.

As AI makes content easier to produce, human judgement becomes more important, not less.

AI can also make individuals more productive while making organisations less collaborative. Previously, an employee might ask an experienced colleague; today they can ask an AI assistant.

Managers therefore need to protect informal learning and knowledge sharing. Teams should share prompts that work, demonstrate useful approaches and discuss failed outputs as well as successful ones. AI should increase access to expertise, not reduce interaction with the experts already inside the organisation.

Once individuals and teams become competent users, the next opportunity is to stop looking only at isolated tasks and examine how work actually flows.

"Use Copilot to write this report" is a task. A workflow asks where the information came from, where the data sits, what happened before the report was created, who reviews it, what happens afterwards and where human judgement is required.

This matters because the smartest model is not necessarily the most useful model if it cannot securely access the information required for the workflow. McKinsey's 2025 State of AI research found that workflow redesign was the organisational factor most associated with bottom-line impact from generative AI, yet only 21% of respondents using generative AI said their organisations had fundamentally redesigned at least some workflows.

The next phase does not need to start with another platform or complex agent programme. The near-term opportunity is simpler: establish clear governance, select trusted tools, help individuals use them properly, share what works across teams, then improve the workflows where AI can remove friction.

And measure it.

AI now has to demonstrate return on investment. Licence costs, token usage, implementation, training, security and governance all add up. "Our people are using AI" is no longer a meaningful measure of success. Leaders need to know whether work is being completed faster, quality is improving, repetitive activity is reducing and the capacity created is being redirected towards higher-value work.

The AI honeymoon may be over. That is not bad news.

It means we are moving past the hype and into the phase where the technology has to earn its place in the business.