The case for and against AI token capping

Agencies considering capping token usage must weigh up the risk they could douse their employees’ creativity by limiting their access to the tech.

The case for AI token caps

AI costs — from tokens to licensing and subscriptions — are likely as cheap as they’ll ever be. While manageable today, they’re kept low because of the enormous capital expenditures of companies like Meta, OpenAI and Anthropic. Eventually, those firms will want a return on those investments, meaning today’s prices might not be stable for long.

With that in mind, marketing agencies are building the infrastructure of AI accountability today, rather than wait for a crisis. Performance agency Roast, for example, has capped the amount of AI tokens its staffers can use since March.

“It’s up to the individual to use those tokens responsibly, and I think the benefit of that is then it’s making people think about what use cases they use tokens for,” John Campbell, head of innovation and AI at Roast, told Digiday. 

The company allots a set amount of tokens to each employee within its Claude subscription, Campbell said. Higher individual caps can be provided at particularly busy times. He noted that the limit had stopped some users from relying on the tech for repetitive tasks that might be better delegated to non-AI software tools. 

Video and image production generally represents the most token-intensive activity for creative agencies; media agencies experimenting with agentic tools for planning and buying must also grapple with runaway operating costs.

U.S. media agency PMG places a $50 daily limit on staff AI usage through its Alli4U portal. Staff that hit the cap more than once can be assigned a larger token budget, or receive advice on using different AI models, Dillon Larberg, consulting and strategy director, told Digiday in September. “That provides us a human in the loop moment to meet with teams or meet with individuals,” he said.

While agencies might still want to allow staff the flexibility to experiment and play with AI tools they can use tiered capping systems to provide teams with enough leeway to pilot something new, without burning up the budget.

“If you say you have a limit of $30, $25, $10 a day, you’ll find people become very creative in how they get this stuff done based on that constraint,” said Jonathan Whiteside, global executive vp at Dept, which caps team AI usage rather than individual costs. 

“Constraints are sometimes useful for us to find new ways of working,” Whiteside added. “People start to think much more carefully about, ‘Well, do I actually need to use AI to do this?’”

The case against capping

AI costs, including the price of tokens, won’t stay low forever, so agencies should make hay while the sun’s shining. Intentionally cutting back before compute costs rise would be like abstaining from MoviePass or WeWork in the glorious days of zero interest rates. Agencies can chew through tokens while they develop the best uses for AI tech now, while it’s cheap to do so.

Some agencies use soft AI caps, rather than a hard limit. “When someone maxes out their spend for that tier, they show their work and their output to their department head,” explained Daniel Gilbert, CEO of Brainlabs, in September. He described the company’s approach as one that embraced “checkpoints” rather than caps. “The department head goes through the work with them, and if it’s good work, then they push them into the next tier.”

Such an approach preserves staff autonomy while keeping a lid on AI spending for agencies that want to to let a hundred flowers, or AI agents, bloom.

There are also means of nudging staff into cost-conscious AI decisions, without placing concrete limits on their usage. Gateway systems, or “gating,” refers to a tool which reroutes a user from a token-intensive AI model to a less powerful (and cheaper) model, if it’s better suited to a user’s project. If they’re using ChatGPT to draft their LinkedIn posts, they probably don’t need Astra, its latest and most powerful model.

Whiteside said that model selection is “a big factor” in keeping costs down, but that gating systems or staff training can provide a more appropriate guardrail than an individual token allowance.

As the real costs of AI-enabled marketing no services become clearer, expect more agencies and brands to consider whether hard caps or flexibility are most important.