Token spend is becoming a team skill.
A team can get real value from AI and still lose sight of what each task costs. As usage grows, knowing when to use a powerful model and when a simpler one will do becomes part of working well with AI.
A post about Harvey’s AI costs caught my attention this week. Bloomberg reported that Harvey’s gross margin moved from about 50% early this year to negative 50% by June after customer use of its agents surged, citing a person familiar with the matter. Harvey has described its work on an open weight model and its interest in cost efficiency. The reported margin is one company’s experience, but the lesson is much wider: a company can build something people value and still lose sight of what it costs to deliver that value as use grows.
At Ruby, we start by getting to know the people doing the work and the decisions they make throughout the day. That matters for AI cost as much as it does for usefulness, because the team needs to see which tasks call for deep reasoning, which ones are routine, and where a person needs to review the result. Once you understand those differences, model choice becomes a decision your team can make with purpose, instead of a setting everyone passes over.
Token efficiency is not a request for employees to use AI less. It is a way to make sure the intelligence a team pays for is going toward work that benefits from it.
The subscription is only part of the picture.
Many teams first meet AI through a seat based subscription, where usage is included up to a limit. Once a company adds advanced features, credits, agents, or its own API powered workflows, the economics can look different. Business workspaces can draw on optional credits after included usage is exhausted. Enterprise billing depends on the agreement, with some workspaces using credits and eligible contracts charging for token based usage alongside seat fees. API usage has separate billing.
That is why “we have an Enterprise plan” does not, by itself, answer the cost question. Leaders need to know which parts of their work are covered, which are metered, and who can see the usage as it grows. The exact answer will depend on the provider and the contract.
Look at the task before choosing the model.
When every request goes to the most capable model, a team may be paying a premium to reformat notes, sort familiar information, or produce a first draft. Those tasks can sometimes be handled by a smaller model, while complex analysis and consequential decisions deserve more capability and closer human review. The right choice is the model that meets the quality bar for that particular task, not simply the cheapest model on a rate card.
Good model selection also depends on how the work is set up. Long documents pasted repeatedly into a workflow, requests that have to be retried because the goal was unclear, and agents that call models several times for one result all add usage. A clearer task, the right amount of context, and a repeatable process can help the team get to a useful result without starting over each time.
Give employees the judgment to use AI well.
Our goal would be to work alongside the people using these tools, understand the jobs they are trying to get done, and show them what changes when a different model or workflow is used. A team can test a representative task on two models, review the outputs together, and compare the cost per useful result. That conversation makes token spend something employees can understand and act on, rather than a number that only appears on an invoice.
From there, a company can set simple defaults for routine work, make it easy to step up to a stronger model when the task calls for one, and keep a pulse on both quality and spend. Employees should feel comfortable and capable making that choice. If a lower cost model creates more errors or more rework, the company has learned something useful too.
The cost of intelligence will keep changing as models and pricing change. A team that understands its own work can adapt with it, using more capability where it matters and less where it does not. That is how AI becomes something a company can keep using with confidence.
Sources & further reading
Start with the work. Build the right system around it.
We help teams understand where AI is useful, choose the right tools for the job, and build the confidence to keep improving.
Book a Call