Accuracy at Scale

AI21 diagnoses that enterprise AI still has not overcome the wall of reliability and accuracy. This problem is even more critical in industries where accuracy is essential, such as healthcare, banking, and manufacturing, where LLM and LRM end-to-end performance remains unstable. Increasing hard coding to compensate for this results in a loss of flexibility.

Companies have only chosen between two approaches so far.

• Prompt & Pray: A method that expects the model to handle things well on its own, where errors accumulate in complex multi-step, multi-tool workflows.
• Hard-Coded Chains: Offers predictability, but is vulnerable to change and costly to rewrite.

Large Reasoning Models have improved in performance through Chain-of-Thought (CoT) and Reinforcement Learning (RL), but results vary from attempt to attempt, and instruction-following and tool use are not stable. It is pointed out that for problems with a large action space and many environment-specific variables, the method of sequentially generating thinking tokens itself is not suited to exploring alternatives.

Existing agentic frameworks also fail to escape the same limitations, as they merely thinly weave together LLM, business logic, and data. In the end, the structure remains one where you must choose between unstable AI and rigid programs.

Instead, AI21 presents Guaranteed AI Performance and Quality Level Agreements (QLAs). When users specify instructions, constraints, and cost limits, the system is designed to satisfy those conditions, and the results have Predictability, Transparency, and Cost Control.

The key point is that AI must not simply stop at thinking, but must plan and execute. For this, it explains that macro-planning, which breaks a task down into multiple steps, and micro-planning, which selects the appropriate model and tools at each step, increases inference-time compute, and runs verification, are both necessary.

This approach is seen as becoming the foundation for automating tasks where accuracy matters, such as contract drafting, research summarization, and customer support. AI21 has announced that more details will be revealed soon.