Master tokenomics or broke-enomics
Anthony Vella & Darren Chua | Lumyra AI Growth Catalyst | Aug 2026
Every now and again, a new term enters the business zeitgeist. One term we’re hearing a lot more is tokenomics. The term and the underlying mechanics of tokenomics are something project and product teams, CFOs, and CIOs alike need to get familiar with.
Tokens are the chunks of text an AI model reads and writes. Models don’t process letters or whole words; they break text into pieces, roughly 3-4 characters each. “unbelievable” might split into un + believ + able. Common words are usually one token; rare words, names, and code split into several. A rough rule of thumb: 1 token ≈ 0.75 English words, so 1,000 tokens ≈ 750 words.
This matters as LLM pricing models are based on token generation. Tokens are the fuel in the new world of AI and some organisations are suffering sticker shock.
Earlier this year, Fortune reported that Uber had burned through its entire 2026 AI budget in four months, leaving its COO questioning whether the spend was worth it. Microsoft faced the same token cost crunch, limiting Claude Code licences and cracking down on “tokenmaxxing” by its engineers. All this happened while the price of AI was falling.
Understanding what drives AI cost and designing an organisation’s use of AI so the spend stays deliberate is quickly becoming a key priority for business leaders.
AI consumption costs don’t behave like software licenses
Traditional software was a fixed cost per person, which made financial management and forecasting predictable. AI began the same way, with model providers offering seat-based agreements. Through 2026, however, the commercial model has shifted to a combination of seat fees and usage. Each licence has a fixed fee component that provides access to the models, with usage then billed on a consumption basis.
The major model providers have both moved to this structure. Anthropic’s published enterprise pricing for Claude is per seat, with usage then billed on top at its standard API rates, so the total cost scales with the models used and the volume of work put through them. OpenAI structures ChatGPT Enterprise the same way. Each seat includes a base level of access, and organisations then purchase a shared pool of credits that is drawn down staff make heavier use of advanced capabilities such as deep research and coding.
The consequence is that AI cost is now partly decoupled from headcount. Seat counts still set a floor and give a rough sense of scale, but consumption can rise disproportionately, and a small number of heavy users or a single agent-driven process can account for a large share of the bill.
Business casing needs to change
That changes how a five-year business case is built. Alongside the seat count, a credible case needs a forecast of how much work will flow through the models and how adoption will spread, unit-cost assumptions, and sensitivity ranges in place of a single line. The most reliable inputs are measured rather than assumed. Metering what tasks consume in the first months of use will provide a far better basis than a vendor estimate.
Discipline has to survive past the approval gate. Analysis by the RoAI Institute found that only 4 per cent of organisations running unmeasured pilots achieve ROI success, while 85 per cent of those with formal, regular reporting of AI economics do. The biggest driver of AI returns is not the technology but whether anyone keeps measuring after go-live. That means usage reporting from day one, monthly review against forecast, and re-baselining the business case at each stage rather than assuming the original approval holds.
For dev teams, striking a balance between ideation and scarcity is important. Tech companies that created leaderboards for developers using the most tokens as a sign of success are now retracing their steps and putting token budgets in place.
Not all AI, all the time: the importance of workflow design
Maslow’s hammer states that if all you have is a hammer, everything looks like a nail. There can be an uncomfortable push by boards and executives, “we need to use more AI”. The response to this refrain should always be “for what?” Organisations are fortunate that they have more than just an AI hammer in their tool bag, and there are other tools that can get the job done.
One decision framework that can help organisations decide whether probabilistic reasoning and tools are more appropriate vs. a deterministic approach is known as 3C decision-switching, a research framework we’re developing for the governance of AI systems. It weighs three factors for each step: Confidence, how far the model’s outputs can be trusted for the task; Criticality, the consequence of getting the decision wrong; and Calibrated risk appetite, how much risk the organisation has decided to accept in pursuit of its goals. High confidence widens the range of steps a model can be given; high criticality narrows it; risk appetite sets the threshold between the two.
A high criticality process or decision may be best handled by deterministic, rules-based modelling. In addition to increasing control for high-stakes decisions, deterministic systems don’t consume a costly number of tokens needlessly! For example, an organisation that receives a large number of refund requests via email may use an AI parser to triage the requests to the right queue, and potentially also populate text to help process the refund; however, a deterministic check would be undertaken to reconcile the refund amount and authorise, with additional human-in-the-loop escalations.
This end-to-end process balances the benefit of AI and automation with guardrails and humans in the process. By taking a range of differently formatted emails in free text, AI helps to sort and triage, taking away from human manual work, but in the critical step in the process (where a financial decision needs to be made), the process switches to deterministic, and/or human in the loop. Tracking error rates over time for AI parsing vs. the time and cost saved of eliminating human effort in undertaking the triage makes for pretty easy maths to understand and assess the viability of the business case.
Treat AI spend like cash and put controls around it
Emerging economic theory speculates that AI tokens may one day be fiat, interchangeable between organisations and potentially even fungible with other types of fiat. This might be a bit of a clue as to how to treat tokens, a little like cash.
Organisations are familiar with cash management and have built robust governance around it. The same discipline needs to be applied to tokens, with spending caps and delegations put in place to avoid runaway or unapproved spend. This extends to accountability structures: alerts when usage departs from the expected pattern, and each team charged for what it consumes rather than drawing on an unmetered central pool, because unmetered units get used carelessly.
None of the challenges raised are reasons to slow investment in AI capability. Falling unit prices mean each dollar of AI spend buys more work every year, and the organisations that benefit will be the ones that direct that spend deliberately. Competition from Chinese models such as Kimi K3 and Qwen could see a price war for LLMs break out in the near term, as the frontier models battle for market share. IT and AI architects will take a multi-model approach in the future, triaging between closed models like Claude/ChatGPT for highest complexity workflows and running private, open-weights models that they can train on their own data and run at lower cost (but with capex investment).
Managing AI cost well means doing the simple things right: designing work so AI is used where it delivers ROI, building the business case with the right value drivers, giving the board a number and someone accountable for it, and putting controls around the spend. None of these disciplines are new. Organisations already apply them to cash and to every other business case; the same rigour now needs to extend to AI.
In our next post, we’ll delve deeper into the 3C decision-switching framework.
