CRM AI Context

Seven rules you can write today without research, as a copyable YAML file. The whole starting library for grounding CRM AI, and why it's supposed to look this small.

A canonical schema for CRM AI context entries: the governance envelope, field entries, rule entries, the thirteen rule categories, and a JSON Schema file you can validate against.

Three things that get substituted for each other and answer different questions. Why embedding your field descriptions doesn't work, and where a semantic layer genuinely overlaps.

Every term used across this site, defined once: context layer, rule categories, phased retrieval, hard and soft controls, resolution states, and the failure modes each one prevents.

MCP standardizes how an AI application connects to your CRM. A context layer defines what the data means and what the agent may touch. You need both, and one will not do the other's job.

Authority is a ladder that doesn't bend; specificity breaks ties among equals. Where a live user instruction sits, and what to do when two active rules genuinely contradict.

Store AI guidance rules in Salesforce as admin-maintained records, enforce object and field access in Apex, and load only the rules a question needs before the agent writes a line of SOQL.

The HubSpot implementation of a CRM AI context layer: teach agents which custom properties to trust, what their values mean, how associations behave in practice, and how to manage the rules with files, prompts, or Enterprise custom objects.

A working, CRM-independent context layer. The reusable knowledge and prompts live in git, an app runtime loads only the relevant rules, and a rep gets a grounded first-draft email.

Object, field, and row scope; running as the asking user; why a permission-suppressed null looks exactly like a blank; who is allowed to edit the rules; and a five-level model for letting an agent act.

A regression suite for your context layer: targeted and broad questions, an evaluation rubric that goes beyond pass/fail, and a 30-question starter set as executable YAML.

What has to be true before a CRM AI agent answers for real users: enforcement in code, always-load rules, phased retrieval, disclosure, and attributable logging. Every item is done or not done.

What actually happens between a question and an answer: phase-1 payload, rule index, selection, query plan, tool-boundary validation, and the audit log. Plus the same question failing on purpose.

Your CRM AI gave a confident wrong answer. A symptom-to-cause index for the twenty failures that actually happen, what each one means, and the rule that prevents it recurring.

Eleven ways this gets built wrong, starting with dumping your CRM schema into the prompt. What each one looks like, why it fails, and what to do instead.

A generated summary field is more dangerous than a stale one, because it's fresh, fluent, and it asserts things. The two failure modes, compounding fabrication, vocabulary decay, and the controls that contain them.

The part after it works: a maturity model, the rule lifecycle, expiry and review SLAs, what to do when the CRM schema changes underneath you, and the metrics worth watching.