TL;DR — Key Takeaways
- LangGrant has launched an open source initiative to develop a standard way to capture the reasoning, evidence, policies and approvals behind AI-driven decisions.
- The proposed model is designed to let humans inspect, modify, reuse and govern reasoning across multiple AI agents, models and enterprise data sources.
- A common reasoning standard could improve auditability, compliance and portability while reducing dependence on any single frontier-model provider.
LangGrant today launched an open source Enterprise Reasoning Initiative to drive the development of a standard for a human-readable shared information model for conversations with and between artificial intelligence (AI) agents and applications.
That reasoning would manifest in a durable artifact containing the sources, steps, evidence, human and AI contributions, semantics, policies, versions, and approvals behind an analysis or decision. Humans would then be able to review, modify, re-execute, and reuse that reasoning collaboratively across workflows much like any other artifact, such as source code, that is included in an application.
That common representation for reasoning would make it possible for people, software tools, and AI models to work collaboratively over time. Human judgments, such as changing an assumption, applying domain expertise, rejecting an inference, introducing a business rule, or refining a definition, can become part of the reasoning process.
The proposed standard is specifically focused on six capabilities: Structured reasoning; human judgment throughout the reasoning process; reasoning lifecycle management; reasoning across multiple enterprise information sources; progressively evolving semantic intelligence; and attribution of reasoning and decisions to business outcomes.
Those capabilities, in effect, would then make it simpler to both apply analytics to AI agent conversations and enforce safety and compliance mandates before a decision is made or a task performed, says LangGrant CEO Ramesh Parameswaran. We want to provide insights into the reasoning versus just being able to look at the output, he adds.
An enterprise IT team, for example, could define a well-documented set of standardized reasoning processes that could be consistently applied across those workflows, notes Parameswaran. The overall goal is to enable humans to better understand and become part of the reasoning process used by an AI agent to complete a task, he adds.
At launch, other companies that have joined the initiative include Almaden AI, Causal Dynamic Labs, Conflux, DeepGraph, GirardAI, LEIT Data, Ngentix, Proof Analytics, Skyhook and The Knowledge Graph Guys.
No provider of a frontier model has thus far signed up for this initiative, but it’s inevitable that some type of reasoning standard will emerge, says Parameswaran. That standard would go a long way to demonstrating that providers of AI agents and applications will be able to apply policies and controls to AI agents, he adds. The absence of such a standard only invites more regulation, notes Parameswaran.
It’s not clear how long it might take to achieve such a goal. If IT teams are able to define a set of reasoning workflows for AI agents in a way that can be applied across multiple frontier models, the providers of AI models will be able to exercise less control over those workflows. In effect, the reasoning workflows become portable across multiple backend AI services.
One way or another, humans will find a way to safely collaborate with AI agents. The challenge now is putting in place the standards that enable humans to supervise those AI agents to prevent them from ever going rogue in the first place.
