Social Rocket
Agencies are beginning not only to control their AI costs, but also to bill them to clients in new ways. Digiday reports on 17 September 2026 that growing tool subscriptions and usage-based token costs are changing the existing logic of technology budgets and agency fees.
This builds on a trend already visible with autonomous media and marketing agents: those who scale AI operationally need not only better models, but budgets, logging, model routing and clear rules on who pays for usage.
Update: AI Costs Move Into Client Models
At Markacy, according to Digiday, the main costs arise from enterprise access to Claude as well as additional ChatGPT access. According to Crispin's Chief Transformation Officer, the agency uses around 100 enterprise AI tools for tasks ranging from media planning to social listening and influencer discovery. Many contracts are structured as fixed annual fees to limit monthly fluctuations.
Dept, in turn, uses Claude, ChatGPT and Google offerings, among others. This is where the largest token costs arise. At the same time, the company says total AI costs still make up less than a double-digit share of the budget; personnel costs remain significantly higher.
What is new above all is the question of re-invoicing. Omnicom CFO Phil Angelastro said at the Goldman Sachs conference that the group is building a model to measure and manage token costs and to recover additional costs. In the long term, he expects a stronger shift toward outcome-based compensation models.
Another agency representative, who was not named, told Digiday that his company is testing two variants: token costs as a separate "material" item alongside human work, or as an embedded component of a combined fee. For clients, this makes the contract question more concrete, which we have already described in our piece on AI clauses in agency contracts.
Audit Logs Are Meant to Make Autonomous Decisions Traceable
As early as the beginning of September, several agencies had described additional control mechanisms for AI agents in media buying. According to Digiday, Rise, an agency from the Quad group, uses an audit log feature from PubMatic. It makes it possible to reconstruct which change an agent made and why it deviated from an original brief or parameter.
PubMatic presented the underlying governance architecture in August. Guardrails and audit trails are meant to let buyers control which actions autonomous systems actually execute.
This builds on a trend we already see with AI in media buying and platform automation: the more decisions are automated, the more important independent control over decision logic becomes.
Token Budgets Become the New Cost Lever
Other agencies address consumption directly. According to earlier Digiday research, Brainlabs allocates AI tokens in tiers and reviews additional quotas. PMG works with daily limits. Dept also uses a central "AI Gateway" that can route requests to suitable models based on cost, performance or legal criteria.
A current look at agency budgets shows why such systems are becoming more important. Alongside classic software subscriptions, consumption-based costs are increasingly arising, while several models must remain available for different clients and tasks. Without a central view of usage and benefit, a stack of many small, hard-to-allocate cost items threatens to emerge.
For teams building agentic workflows, good prompting is therefore not enough. What is needed are budget limits, logging, model routing, usage analytics and approval thresholds. Similar governance questions arise with social media MCPs and AI assistants.
What Advertisers Should Check in Agency Offers
For clients, it matters how AI usage is calculated. A separate token item can be transparent, but it is only meaningful if it is clear which models and tasks lie behind it. A flat fee creates more predictability, but may obscure actual usage. Outcome-based models, in turn, shift the discussion away from working hours and tokens toward results – but they require reliable measurement metrics.
Regardless of the model, agencies should be able to explain which tools they use, how costs are limited and which work continues to be reviewed by people. The bigger trend is therefore not a simple swap of personnel for AI. Instead, agency services gain an additional variable infrastructure cost layer that must be managed economically and contractually.