What an AI agent can safely do in your business
The difference between an assistant that answers and an agent that acts, and the guardrails that make agents safe to run.
An AI assistant answers questions. An AI agent takes actions: it reads your CRM, drafts emails, updates records, books appointments and moves work between systems. That makes agents far more useful, and it means the rules around them matter more.
Assistants answer, agents act
An agent is given a goal, a set of tools and permission to use them. Asked to “follow up with this week’s new leads,” it can look them up in the CRM, check the calendar, draft personalised messages and log what it did. The Model Context Protocol (MCP) gives a standard way to connect a model to tools, so the agent works through approved connections rather than open access.
Good first jobs for an agent
-
Research and summaries
Pulling together a customer’s history, a week’s leads or a project’s status from several systems.
-
Drafting
Follow-up emails, quotes from templates and replies to common requests, ready for a person to review.
-
Data hygiene
Tagging, de-duplicating and updating records according to clear rules.
-
Routine handoffs
Moving a job from sales to operations with every detail attached.
The guardrails
-
Least access
The agent gets only the tools and data the job needs, read-only wherever possible.
-
Approval for what matters
Anything that spends money, contacts customers in a new way or changes important records waits for a person.
-
A full log
Every step the agent takes is recorded, so you can see what it did and why.
-
Limits
Spending caps and rate limits, so a mistake can’t repeat a thousand times.
Start with the agent drafting and a person approving. Loosen the rules only after you’ve watched it work.
Where agents shouldn’t act alone
Pricing, refunds, legal or medical questions, hiring decisions and anything that can’t be undone. An agent can prepare the information for these decisions, but a person should make them.
How to start
Choose one repetitive task with clear rules that takes real time each week. Run the agent alongside the person who does it today, compare the results, and expand only when it’s consistently right. We build agents with Claude and OpenAI models, chosen per task. In Boardroom, a multi-agent system we built, every consequential action waits in an approvals queue for the owner.
Find your first agent.
A 30-minute call to pick one task worth handing to an agent and the guardrails it needs.