Shared Context Makes HubSpot AI Smarter. It Also Expands the Blast Radius.

Shared understanding is a key concept in HubSpot's new AI framework. The company refers to this as Growth Context, representing a collective comprehension of the business, team, and customer. Context Home facilitates the provision of foundational knowledge across various AI tools. Projects can incorporate instructions, files, HubSpot content, knowledge vaults, and connected applications.

Agents can utilize CRM data and approved tools to fulfill their tasks. The primary advantage is consistency; teams should not need to reiterate the company's details every time they engage with an assistant or create an agent. However, an unintended consequence is the expanded blast radius. For context transparency, HubSpot employs four layers within its AI system: Portal context, Knowledge vaults, Project context, and User and conversation context.

Each layer offers varying degrees of access. The implementation of these layers aims to ensure that teams can reuse knowledge effectively while maintaining control over the influence of shared context. To mitigate potential risks, HubSpot has introduced 'Least-Context Architecture', which recommends granting each AI task the minimum necessary current and authorized knowledge.

This approach prevents overbroad influence from a single bad instruction, outdated methodology, or inappropriate memory. The 'Least-Privilege' principle, which primarily focuses on access control, is insufficient in AI systems. HubSpot's architecture introduces an additional question: even if the system may access the information, should that information influence this decision?

The six gates of this architecture include the Task gate, Audience gate, Source gate, Lifetime gate, Write-authority gate, and Context influence gate. These gates ensure that the context supplied for each AI output is appropriate, current, and authorized, thereby minimizing the risk of unintended consequences.