Corporate AI planning is moving beyond licences and individual use cases. As agents take on more work, boards need an AI token strategy: measure current consumption, forecast future inference demand, remove avoidable token usage, decide which work genuinely needs frontier models, and plan how capacity will be sourced across subscriptions, APIs, private models and owned infrastructure.
Do you know how many AI tokens your business uses today?
Most organisations can tell you their headcount, Microsoft 365 licence count, cloud spend and storage footprint. Far fewer can say how many AI tokens they consumed last month, which users or workflows consumed the most, or what those figures imply for future operating costs.
That is becoming a strategic blind spot. AI expenditure is increasingly spread across employee subscriptions, SaaS products, development tools, API calls and autonomous agents. A useful first step is to build a single view of current AI consumption and identify the top users and top workflows.
The highest-consuming users should not automatically be treated as outliers. They may be leading indicators of what ordinary usage looks like once AI becomes embedded in daily work.
What happens to token demand when agents proliferate?
Human chat naturally limits consumption: a person asks a question, reads the answer and decides what to do next. Agents remove much of that constraint. They can read documents, inspect systems, call tools, search data, draft outputs, check their own work and repeat multi-step processes without a person typing every prompt.
The shift is already visible. Microsoft reported that active agents in the Microsoft 365 ecosystem grew 15x year over year, rising to 18x in large enterprises. OpenAI reported that, by June 2026, Codex generated 64% of combined Codex and ChatGPT output tokens among enterprise customers. Its highest-usage "frontier firms" generated 8.3x as many output tokens per active user as typical firms, up from 2.6x in January.
Sources: [Microsoft 2026 Work Trend Index](https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization) and [OpenAI Enterprise Signals](https://openai.com/signals/enterprise-data/).
For boards, the question is therefore no longer simply "How many employees will use AI?" It is increasingly "How much work will our employees delegate to AI systems, and how much inference will that require?"
Will falling token prices actually reduce AI costs?
Not necessarily. Unit costs can fall while total spend rises because agents consume far more inference than simple chatbot interactions. Gartner calls this the "inference paradox" and forecasts that AI inference costs per agentic workflow will increase more than fivefold through 2028 as workflows become more complex.
Source: [Gartner, 17 August 2026](https://www.gartner.com/en/newsroom/press-releases/2026-08-17-gartner-predicts-ai-inference-costs-per-agentic-workflow-will-increase-more-than-fivefold-through-2028).
This is familiar economics. Computing became cheaper and businesses consumed vastly more computing. Storage became cheaper and organisations stored vastly more data. Bandwidth became cheaper and entire business processes moved online. AI is likely to follow the same pattern.
What should be in a corporate AI token budget?
A token budget should not be an arbitrary cap. It should be a capacity plan. CEOs, CFOs and technology leaders should be able to answer five questions:
- How many tokens are we consuming now, across subscriptions, APIs, SaaS products and agents?
- Who and what are our highest consumers, and are they early indicators of wider adoption?
- How quickly could demand grow under light, medium and heavy agent adoption scenarios?
- Which workloads genuinely require frontier intelligence, and which can run on smaller or private models?
- Where will future inference capacity come from, and what happens if the price or availability of that capacity changes?
How can businesses reduce token demand before buying more capacity?
The cheapest token is the one a workflow never needed to generate. Some tasks should not hit a large language model at all. Deterministic software can often inspect, transform or assemble structured documents far more efficiently than repeatedly sending whole files through an LLM.
This is a principle we have explored through OfficeMaker. In one specific document-processing approach documented in our white paper, combining deterministic document operations with AI reduced token requirements by up to 10,000x versus repeatedly processing entire documents through a language model. The lesson is broader than OfficeMaker: optimise the workflow before shopping for cheaper tokens.
A practical routing model has three tiers: deterministic operations for work that does not require AI; smaller or private models for routine extraction, classification, summarisation and background processing; and frontier models for the difficult reasoning where their extra capability genuinely creates value.
Where should future AI inference capacity come from?
Most businesses currently rent almost all of their AI intelligence through subscriptions and public APIs. That is rational at modest volumes, but the economics can change as consumption scales.
A mature sourcing strategy may combine employee AI subscriptions, metered APIs, committed cloud capacity, smaller private models and on-premise inference. Routine or sensitive workloads can run privately, complex reasoning can be escalated to frontier providers, and unexpected peaks can burst into the cloud.
The objective is not to replace OpenAI, Google, Anthropic or other frontier providers. It is to use expensive intelligence where expensive intelligence creates value, while keeping routine workloads on the most economical appropriate capacity.
What if AI demand starts to outstrip available capacity?
A global shortage is not inevitable, but it is a scenario boards should model. AI capacity ultimately depends on accelerators, networking, data centres, electricity and supply chains. The International Energy Agency projects global data-centre electricity consumption to roughly double to around 945 TWh by 2030 in its base case, with accelerated servers driven largely by AI growing much faster than conventional server demand.
Source: [IEA Energy and AI](https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai).
If demand temporarily outruns available capacity, the likely result is not that 'tokens run out'. It is that guaranteed capacity becomes more valuable, reserved access matters more, less critical work is routed to cheaper models, and organisations with private capacity or highly efficient workflows have more resilience.
Why AI capacity planning starts to look like energy procurement
Large organisations do not simply ask what electricity costs. They forecast demand, improve efficiency, secure supply, diversify sources and sometimes own part of their generation. AI capacity is beginning to deserve the same discipline.
That means thinking in terms of demand, efficiency, capacity, supply, price and resilience. An AI strategy that only lists projects and licences is incomplete if it does not explain how much intelligence the organisation expects to consume and how that capacity will be sourced.
What should CEOs and CFOs ask now?
- How many AI tokens are we consuming today?
- Who are our highest-consuming users and workflows?
- What happens to demand when agents become commonplace?
- What will our inference requirement look like in one, three and five years?
- Where are we wasting tokens on work that could be deterministic?
- Which workloads require frontier models and which do not?
- How much capacity should we rent, reserve or own?
- How exposed are we if inference prices or availability change?
At AI Build, we help organisations map current AI consumption, model future demand, identify unnecessary token usage and design a practical sourcing mix across subscriptions, APIs, private models and owned inference. Talk to AI Build if you want to understand what your organisation's token strategy should look like before agent adoption accelerates.