AI agents for business: what executives should approve, and what to hold back
An agent does not just answer; it takes steps in your systems. That is where the value is, and it is why the decision about how much it may do alone belongs to the executive team.
An AI agent is software that uses an AI model to plan and carry out a multi-step task using tools, such as reading a CRM record, drafting a reply and updating a ticket. For executives the decision is not which model to use but how much autonomy to grant: which tasks an agent may complete alone, which it may prepare for a person to approve, and which it may not touch.
Anthropic
OpenAI
Microsoft 365Salesforce
HubSpot
ServiceNow
Workday
NetSuite
Slack
Where the numbers come from
These counts come straight from the lists on this page. Nothing here is a survey result or a client outcome.
What is the difference between an assistant, a workflow and an agent?
An assistant answers when asked. A workflow follows steps a person designed. An agent decides its own steps toward a goal, using the tools it has been given.
- Use a workflow when the steps are known and do not change; it is cheaper and more predictable.
- Use an agent when the path varies from case to case.
- Many useful systems are workflows with one or two agent steps inside them.
Anthropic's guidance on building agents makes the same point: start with the simplest design that works and add autonomy only where it clearly helps.
How much autonomy should an agent have?
Decide by what a mistake would cost and whether it can be undone. Start one level lower than you think you need and raise it when the record supports it.
- Acts alone: low-cost, reversible, checked by sample.
- Prepares for approval: the agent does the work and a person releases it.
- Not for agents: moving money, changing access, and decisions about people's jobs, credit or rights.
What controls should be in place before an agent goes live?
The same controls you would require of a new employee with system access, written down and tested.
- A named business owner for the agent and its results.
- Its own account with the minimum access the task needs, never a shared admin login.
- A log of every action it takes, kept where the owner can read it.
- Limits: what it may spend, send or change without approval.
- A way to pause it immediately, known to more than one person.
- A test set of real cases it must pass before launch and after every change.
How do agents connect to your systems?
Through the systems' own APIs and through connectors that expose them to the model. The Model Context Protocol is an open standard for those connectors, supported by several AI vendors.
- Access should follow the permissions the agent's account holds in each system.
- Prefer connectors your team or a trusted partner controls.
- Review what each connector can read and write before turning it on.
Questions executives ask about this
Are AI agents ready for a mid-size business?
For bounded tasks with a review step, yes. For open-ended work with no checkpoint, treat claims with caution and test on your own cases first.
Will agents replace staff?
They change what people spend time on. The tasks in the table are parts of jobs, not whole jobs, and the review steps are done by the people who know the work.
Who is accountable when an agent makes a mistake?
The company, and within it the business owner named for that agent. That is why every agent needs an owner, a log and limits.
Can you build agents on our systems?
Yes. We connect AI to the software you already run, with scoped access, logging and a review step, and we train the team that will own it.
Primary sources to check against
Building effective agents (Anthropic engineering)
Model Context Protocol: connecting AI applications to business systems
Agents guide (OpenAI platform documentation)
Microsoft 365 Copilot overview (Microsoft Learn)
AI Risk Management Framework, AI RMF 1.0 (National Institute of Standards and Technology)
Frameworks, laws and product details change. Confirm the current text with each publisher and with your own counsel. Consultants in a Box is an independent consultancy; this page is not published or endorsed by any organisation named. Names and logos are trademarks of their owners.
Related guides
AI adoption strategy
A pilot proves a tool can do something. Adoption is when a team does its work a different way and a number moves. The gap between the two is ownership, data and habit, not technology.
Read the guideAI acceptable use policy
Staff are already using AI. A policy that fits on two pages, names the approved tools and says plainly what may not be pasted into them does more than a long one nobody reads.
Read the guideAI operating model
An operating model answers three questions: who chooses what gets built, who builds and runs it, and who answers for the result. Without it, every department buys its own tool and nobody owns the risk.
Read the guideAll AI for executives guides
The full leadership playbook, from readiness to agents.
See the hubWant to talk this through for your company?
Book a call and we will go through where you are, what to decide first and who should own it. Or start with the free business analysis.
