Maestro AI Plan Orchestration

AI21 introduced Maestro as an AI planning and orchestration system for automating data-intensive enterprise work. Citing AWS estimates, it noted that only 6% of organizations have actually deployed generative AI, pointing to the probabilistic nature of LLMs as the core obstacle to enterprise adoption.

Existing approaches fall into two categories.

• prompt-and-pray: throwing an open-ended task at an LLM/LRM and hoping for the result, which makes errors easy to accumulate in complex multi-step workflows.
• Hardcoded chains: reliable but inflexible, requiring rewrites whenever the environment changes.

Maestro treats LLMs and LRMs as tools rather than final decision-makers, comparing alternative paths to predict success rates and costs before generating and executing a plan. After execution, it verifies whether requirements were met, and shows the entire flow transparently through an execution trace and a verification report.

There are four core capabilities.

• Accuracy: adjusts the amount of reasoning-step computation and selects models/tools, then verifies rigorously.
• Deployment speed: automatically generates a task-specific plan once requirements, tools, and budget are set.
• Environmental adaptation: explores alternative paths through offline simulation to find the optimal strategy.
• Transparency: both the execution process and verification results can be checked.

Benchmark improvements were also presented. On IFEval, GPT-4o rose from about 85% to 91.9%, Claude Sonnet 3.5 from about 88% to 95.2%, and o3-mini from about 92% to 95.7%. On tests requiring multiple requirements to be satisfied simultaneously, improvements of up to 50% were seen, and on FRAMES, it reached 75%, ahead of OpenAI Assistant API (69%) and ReACT with LlamaIndex (59%).

AI21 is currently running an early access waitlist, with general availability in SaaS and VPC forms slated for late 2025.