Four questions to answer before you switch on an AI agent

AI agents are being switched on inside CRMs faster than anyone is deciding what those agents are allowed to touch.

Why it matters: A chatbot that answers badly is an embarrassment. An agent with write access that acts on bad data is a data-quality incident in your system of record — and it scales at machine speed.

Ask these before go-live

• Whose permissions does it run as? If the answer is "an admin," stop. An agent should hold the narrowest profile that lets it do its job, and nothing more.
• What can it write? Reading is low risk. Creating, updating and emailing are not. Decide field by field, not system-wide.
• What happens when it is wrong? Not if. You need the audit trail and the rollback path decided before the first run, not after the first bad record.
• Who reviews the output, and how often? An unreviewed agent is an unmonitored employee.

The pattern that works

Let the agent draft and recommend. Let a person commit. You keep the speed and lose almost none of the safety — and the review step gets faster as trust is earned.

The big picture

Gartner projects that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from under 5% in 2025 (reported here). The question stops being whether you deploy one and becomes what it is permitted to do.

The bottom line: Scope the permissions before you scope the ambition.

Working through this in your own org? Talk to an Expert — OneAlgorithm builds and governs Salesforce and automation for regulated and growing businesses.