AI orchestration fixes cut enterprise task costs 41%
Enterprises can slash AI deployment costs by more than 40% by fixing the software orchestration layer surrounding foundation models, according to new research from Writer.
New research from AI company Writer demonstrates that enterprises can cut the cost of automated tasks by 41% simply by optimizing the orchestration software wrapped around foundation models. The study found that this optimization slashed token consumption by 38% and reduced task latency by 44%, all without altering the underlying AI model or degrading output quality.
The findings address a growing financial drain in corporate AI deployment known as "tokenmaxxing." Rather than engineering efficient systems, developers often stuff massive context windows into AI models and rely on brute-force computation to complete tasks. "Your invoice is tokens-per-task times price-per-token, and most teams only watch the second number," said Waseem AlShikh, CTO and co-founder of Writer. "In agentic workloads, tokens-per-task compounds — every loop iteration re-transmits the growing context — and it compounds faster than prices fall."
Current industry efficiency techniques, such as prompt compression or speculative decoding, fail because they try to fix the model itself rather than the system architecture. Writer’s research focuses on the "harness"—the orchestration layer that routes prompts and formats data. By treating this layer as a core piece of software rather than disposable glue code, companies can dictate their actual unit economics.
The researchers tested a tuned harness against a standard agent loop across six models, including Claude Sonnet 4.6 and Gemini 3.1, using 22 locked enterprise tasks. Blended cost per task dropped from 21 cents to 12 cents, while success rates actually improved marginally, moving from 78% to 81%. "Enterprises spend months on model evaluations and then rent their orchestration off the shelf," AlShikh said. "Whoever owns the harness owns your unit economics, and an open framework tuned for demos is not tuned for your invoice."
However, the study revealed limits to this approach. Delegating tasks to specialized sub-agents—a key cost-saving mechanism—only proved reliable on the most capable models. Claude Sonnet 4.6 and Writer’s Palmyra X6 scored 0.85 and 0.86 in reliability, respectively. Lighter models like Gemini Flash 3.5 and Qwen 3.6 scored below 0.45, making them unsuitable for complex multi-agent routing.
For executives managing AI budgets, the research dictates a shift in engineering focus. Developers must separate static instructions from dynamic queries to trigger API prompt caching, and move conversation history out of the active context window to avoid re-paying for data the model has already processed. Most critically, spending guardrails must be hardcoded into the software. "The core principle is that you never ask the model to police its own spending," AlShikh said. "The fence has to live below the model, in code, on your side of the API."
