‘Tokenmaxxing’ – The wrong AI race to run in Singapore

“Tokenmaxxing” might sound like a side‑effect of AI, but it has a real cost – both for people and for organisations.

Ask any manager what “using more [artificial intelligence] AI” actually looks like in their company and you’ll probably get a long answer, but not a confident one.

Between new tools, dashboards, and intranet updates, the company-wide mandate is clear: everyone should be doing more with AI. What’s less clear is whether all that activity is actually transforming work, or simply encouraging people to burn more tokens so they can say they are “using AI enough.”

In some teams, that pressure is now visible. A developer keeps an AI tab open all day so their usage graph never dips. A project lead runs every document through an AI editor even when the first draft was clear, because “it’s what we’re supposed to do now.” Recent reporting has described tech giants’ employees using internal AI agent platforms to run unnecessary tasks, largely to inflate AI usage scores and stay near the top of internal leaderboards.

This practice is dubbed “tokenmaxxing:” Using AI tools in ways that maximise token consumption, not actual impact. The work doesn’t get better. The numbers just get bigger.

This behaviour isn’t appearing in a vacuum. In markets like Singapore, the gap between AI ambition and day‑to‑day reality is closing fast. Official data from the Ministry of Manpower (MOM) shows that whilst a majority of firms have yet to fully adopt AI, those that have taken the leap are already reporting productivity gains and are starting to move from one‑off pilots to more deliberate workflow redesign. The ambition is real, but so is the pressure that comes with it.

Leaders are being asked hard questions by boards, regulators, and regional headquarters: Where is AI in your business? How are you using it? What’s the return?

HR and business leaders, in turn, feel they need something concrete to point to. A chart. A percentage. A line going up. In that environment, it’s tempting to grab the clearest number on the screen – how many tokens people are burning, how often they log in, which team is leading on usage. Once you tie AI success to raw usage, people do what people always do: They optimise for the metric.

The intent is good but the incentive is off.

The real costs of ‘tokenmaxxing’
Tokenmaxxing might sound like a small, slightly nerdy side‑effect of the AI wave, but it has a very real cost – both for people and for organisations.

The first cost is time and attention. When employees feel they need to prove they are AI‑forward, they start rerouting simple tasks through tools that don’t really need them. A clear email goes through three rounds of AI rewriting. A straightforward report is pushed into a model mainly so it shows up in the logs. Every unnecessary prompt is time not spent with a customer, a colleague, or a hard problem. Over the course of a week, those extra steps add up.

The second cost is trust. In the early days, many people were genuinely excited about what AI could do. It felt like much-needed help that could add a turbo-boost to their day. But when “tokenmaxxing” takes hold, AI starts to feel like another hoop to jump through rather than a source of genuine assistance. Once people associate AI with busywork, it’s much harder to convince them it can help with the big things.

The third cost is to the leaders' visibility. Dashboards light up with activity, but blur the line between thoughtful use and gaming the system. The people who look like your AI champions might simply be your best token spenders. There’s also a cultural effect: When usage becomes a performance signal, even informally, anxiety rises. Employees who are using AI appropriately begin to wonder if a comparatively low AI footprint will be read as resistance, or as a lack of ambition.

That’s not the kind of fluency Singapore is trying to build. The goal is not to raise a generation of workers who are good at feeding models. It’s to raise one that is good at using them wisely.

What good AI use really looks like
But if “tokenmaxxing” is the wrong race to run, what does optimal AI use actually look like?

In the healthiest teams I’ve seen, nobody celebrates the person who used AI the most this week. They celebrate the person who finally fixed a long‑standing bottleneck or freed up Friday afternoons by automating a tedious process. AI shows up in the work itself. You can see its fingerprints in clearer reports, faster response times, and fewer handoffs, not just in a spike in tokens.

Similarly, productive AI use tends to live where people already work.

It’s built into the code editor, the shared document, the HR system or the ticketing tool, rather than sitting in a separate portal people visit to prove a point. A recruiter might use AI to help draft outreach emails and screen large pools of applicants, then spend the time saved in deeper candidate conversations.

A finance team might use it to structure messy data and surface anomalies, then rely on their own judgment to decide what those anomalies mean. The impact is visible to the people who matter most: The teams doing the work and the customers on the other side.

Crucially, productive AI use leaves room for human judgment. People can say, “This suggestion isn’t right for our client,” without worrying that their “AI score” will suffer. They know they’re being trusted to choose when AI helps and when it doesn’t.

At the employee level, this means that AI usage should be uneven by design. Some roles, seasons, and projects will have a heavier AI footprint. Others won’t. That’s healthy. A uniform token trail across the organisation is not a sign of maturity. It’s a sign that people are being nudged to behave the same way, whether it makes sense for their work or not.

Rethinking how we measure AI
None of this means measurement is a mistake. But when ignoring AI metrics altogether isn’t an option, and counting only tokens isn’t a viable solution, leaders are caught between needing metrics and knowing that the simplest numbers they have on hand are often the least meaningful.

The real risk for Singaporean firms is not that employees aren’t using “enough” AI. It’s that they’re using it in ways that look good in a report and do very little for the work itself. In a country that prides itself on being both efficient and forward‑looking, that would be a strange place to land.

That is why any serious approach to measuring AI has to be customised to the business, its work, and its people, rather than lifted wholesale from a vendor dashboard.

The exact mix will look different in every organisation. But the common thread is that the signals are anchored in the work, not just in the logs. Metrics around AI use shouldn’t feel like surveillance, but rather a way to spot where support, training or redesign is needed.

If people feel watched, they will game the system. If they feel supported, they will change their work for the better.

So if your AI programme depends on people chasing token thresholds, it’s time to ask the hard questions and design a measurement system that delivers real impact. Only then will you be measuring the race you actually want to win.