Use more AI. But mind the bill.

For the past year, the message to employees has been fairly consistent: embrace AI. Experiment. Find ways to work faster. Build your capability so you don’t get left behind.

Now another message is arriving: be careful how much you use.

A video about Accenture’s AI spending recently caught my attention. According to reporting on an internal announcement, the company has introduced a centrally funded allowance of US$29 per person per month for Claude, with additional usage charged to the relevant project or team. It is a change in how usage is funded, rather than an absolute limit on what employees can spend.

Even with that distinction, the tension is interesting. Organisations want people to move faster with AI. They also need to manage the bill.

I’m seeing versions of this in New Zealand too. Tokens are becoming hot property, with access increasingly tied to demonstrating business value.

I agree with that accountability. But where does experimentation fit when people need to demonstrate the value before they have had a chance to discover it?

The cost was always there

Tokens are the units of information AI models process. With usage-based services, the amount processed contributes to what an organisation pays. Different tools and commercial arrangements package those costs differently.

The cost of an individual interaction can seem insignificant. Across teams, repeated tasks and automated workflows, consumption adds up. Accenture’s own work on AI economics acknowledges this: falling token prices can sit alongside rising overall spending as usage expands.

That makes “use AI more” an incomplete instruction.

People also need to understand which tools suit which tasks, what consumption costs, and when the likely benefit justifies it. If leaders encourage adoption without providing that guidance, employees should not be surprised later by rules they were never equipped to navigate.

Cost awareness belongs in AI literacy from the beginning.

Faster needs to mean something

Speed alone is a poor measure of value.

Producing a document in ten minutes looks impressive. If someone then spends two hours correcting it, checking its sources and making it usable, the initial speed tells us very little.

Equally, an AI interaction that looks expensive in isolation might be worthwhile if it removes hours of repetitive work or helps a team avoid a costly mistake.

We need to assess the whole task: the AI cost, the human effort, the quality of the result and what happens next.

Consider a hypothetical workflow that costs NZ$20 in AI usage and saves two hours. That could be excellent value. Spend the same amount generating material nobody needs, and it is waste.

Even time saved needs careful interpretation. It may create capacity to serve customers, reduce a backlog or give people breathing room. It does not automatically become a cash saving.

Those benefits still matter. They just need to be described honestly.

Experimentation needs its own budget

This is where I think organisations need to make a deliberate distinction.

An established workflow should have evidence behind it. We should understand its purpose, expected costs and whether it delivers.

An experiment begins with uncertainty. Its purpose is to help us decide whether something is worth pursuing.

Requiring both to demonstrate the same certainty creates a problem. The work with an existing business case gets funded. The unfamiliar idea struggles to get permission to develop one.

I would give teams a modest, protected experimentation allowance, separate from the budget for proven uses.

Access would come with straightforward expectations: identify the problem, set a spending and time limit, use approved tools and data, and share what happened.

The question could be simple: can AI help us make these customer instructions easier to understand without losing essential information?

At the end, the team might recommend expanding the approach, changing it or stopping. Any of those can be a useful outcome if the experiment produces evidence.

Learning that something does not work can prevent a much larger investment in the wrong direction.

Keep the permission proportionate

There is little value in creating an approval process that costs more than the experiment.

Small, bounded tests should be easy to start. Larger commitments should require stronger evidence. Spending alerts and clear limits can provide oversight without requiring a business case for every prompt.

Access also deserves attention.

If experimentation depends entirely on a manager’s enthusiasm or a well-funded project, opportunities may concentrate among people who already have influence and resources.

Frontline staff, administrators and customer service teams often understand repetitive work and customer frustrations exceptionally well. They need a practical route to test ideas too.

Otherwise, we risk funding the people best placed to ask, rather than the ideas most worth exploring.

Leaders need to own the balance

There are sensible ways to improve AI economics: choose an appropriate model, avoid unnecessary processing, reuse effective approaches and stop automated work that is no longer useful.

Employees need training to make those decisions. They also need visibility. It is difficult to manage a cost you cannot see.

Leadership needs to provide clear budgets, usable guidance and consistent expectations. Asking people to be inventive while making every unsuccessful attempt feel like wasted money sends a confusing message.

For organisations with tighter budgets, including many here in Aotearoa, this balance matters enormously. Experimentation needs boundaries. It also needs enough room to produce a discovery.

I want businesses to ask what value their AI spending creates. I also want them to recognise that some of that value is knowledge: understanding what works, what fails and where human judgement remains essential.

Before asking people to move faster with AI, leaders need to decide what they are willing to fund, what they want to learn and how they will judge the result.

Curiosity needs a budget too.