The most important budget in your business now runs on tokens.
We help you see it clearly.
AI is the fastest-adopted technology in business history.
We love this. We are pro-AI, pro-experimentation, pro-employees choosing the tools that make them excellent.
The job ahead is not slowing any of it down. It’s making it legible: to the engineer choosing a model, the CTO sponsoring a rollout, the CFO signing the bill, and the board approving the strategy.
This is the gap Flowstate exists to close.
Eight frontier models. Five hundred tasks. Four runs each.
In April 2026, researchers at Stanford, Michigan, MIT, Google DeepMind and All Hands published the first systematic study of agent token consumption. Their findings are a useful map of how this new workload behaves.
Bai et al., “How Do AI Agents Spend Your Money?” arXiv 2604.22750Agentic tasks are a different shape from chat.
Token usage is naturally variable.
Your AI bill is not a single number. It is a distribution.
Accuracy peaks at moderate spend.
Beyond the second-cheapest cost quartile, additional tokens deliver diminishing returns.
Different models suit different tasks.
On 230 problems every model solved correctly, the heaviest models used 1.5 million more tokens than the leanest.
Matching the model to the work is where the gains are.
Models can’t predict their own spend reliably.
This isn’t a flaw. It’s a property of the workload.
It’s also a planning problem with a measurement solution.
Every one of these findings is a place where good information helps.
The paper measured it on a benchmark. We see the same patterns on customer devices.
“We thought our staff were creating new PowerPoints. They were iterating on the same deck for hours. Each ‘change this word on slide three’ billed as a full document regeneration.”
Once they saw it, the fix was easy and the relationship with the tool got better.
“Everyone was reaching for the biggest model by default. Once they had visibility, they let people keep using whatever they liked, and added a gentle nudge for the lighter model where it was clearly enough.”
Bills dropped. Outcomes didn’t.
“What are these five people doing? They’re always running out of tokens.”
They turned out to be the most productive people in the company. The answer was to raise the budget, not the eyebrow.
In every case, the answer wasn’t to use less AI. It was to use it with eyes open.
Two things are true at once.
The vendor layer is the engine.
AI providers are doing exactly what good product companies do: making it easy to use more of what they offer. That’s not a critique, it’s how great products grow. We benefit from it. So do customers.
The customer layer is the dashboard.
The CFO needs to reconcile spend back to projects and cost classes. The CTO needs to understand which workloads are running, on which models, with which outcomes. Employees need to be trusted with the tools they choose. Boards need to see the trajectory.
Cars work better when both exist.
We call this Workforce Engineering.
AI agents are a new kind of workforce. They accrue cost like contractors, deliver value like employees, and are capitalised like software.
Variable rate, by output. Never fixed.
Persistent productivity. Compounding capability.
ASC 350-40 eligible. Audit-grade evidence required.
The discipline is too large to leave on spreadsheets.
On every surface AI is used.
Mac, Windows, Linux, browser, terminal, IDE, production workload. Every prompt, every token, every model choice attributed to a person, a project and a cost class, in real time.
Attribution gets better with every customer added. New customers ship in their second week with the benefit of every prior month of training.
We deploy to production several times a day. The customer-facing surface keeps pace with the model market.
Provider-agnostic by design. Every model, every tool, one view. We complement what providers offer; we don’t replace it.
Where this is going.
AI spend will be capitalised within five years.
The accounting profession is already moving. Continuous CapEx classification will become routine, the same way continuous deployment did for code.
The buyer-side layer will become standard.
Every cost category this large has eventually grown its own customer-side visibility tools. AI will be no different. Vendors and visibility tools coexist comfortably in every other category. They will here too.
The workforce is becoming hybrid permanently.
Contractor-to-agent swaps are already happening. The companies that build a single ledger across humans and agents first will run smaller, cheaper and faster.
Engineering finance is the next category.
Sales has Salesforce. Marketing has HubSpot. Finance has NetSuite. Engineering, the largest cost base in most technology companies, has spreadsheets. That changes now.
Returns are shifting from labour to capital.
For two centuries, productivity gains accrued primarily to labour. AI flips that. The companies that measure deployment accurately will be the ones that compound the gains.
One ledger. Four views.
AI spend by project, team and cost class in real time. Audit-grade reconciliation back to the vendor invoice.
Which workloads are running, on which models, with what outcome. Live.
Your portfolio’s AI trajectory, and where it is creating durable value.
The freedom to use the tools that make you excellent, knowing the company has your back.
We built it because the future of business runs on AI, and the people deploying it deserve to see it clearly.