Five questions to ask before your first AI agent

The first agent an organisation deploys matters more than its size suggests. It sets expectations, shapes how the board thinks about AI, and establishes habits that later agents inherit. A careless first agent creates scepticism that takes a long time to shift.

We use five questions to test whether a proposed agent is ready to fund. None of them is technical. All of them can be answered in a room with the right people, and a weak answer to any one is a reason to pause.

1. What decision or task, exactly?

"Improve customer service" is an ambition. "Answer questions about invoice status and send a copy of the invoice on request" is a task. Agents need the second kind of description.

A good definition names the trigger, the steps and the finish. What starts the work? What does the agent need to do? What does done look like? If the answer needs the word "etc", the scope is not yet clear.

It also helps to be explicit about what the agent will not do. An agent in Agentforce Service that handles invoice queries should have a clear rule for the moment a customer disputes the amount. Boundaries are part of the design, not a limitation of it.

2. What data does it need, and can it be trusted?

List the information a competent person would need to do the task well. Then ask three things about each item. Where does it live? Is it accurate and current? Is the agent permitted to use it for this purpose?

This exercise is often uncomfortable. Customer records turn out to be duplicated. Policy documents contradict each other. The definitive answer lives in one person's inbox. Better to find this out in a workshop than in front of a customer.

If the data is not ready, the choice is to fix the data first or to pick a task where the data is already sound. Both are respectable. Proceeding regardless is not.

3. Who is accountable?

An agent does not hold accountability. A person does. Before the agent goes live, one named individual should own its outcomes: the quality of what it does, the complaints it generates, and the decision to extend or withdraw it.

That person should sit in the business function the agent serves, not in technology. The technology team keeps the agent running. The business owner decides what it should be doing and whether it is doing it well enough.

A simple test: if the agent gave a customer a wrong answer this afternoon, who would be told, and who would decide what happens next? If two people could plausibly answer, or nobody, the accountability is not settled.

4. What happens when it is wrong?

Every agent will be wrong sometimes. So will every employee. The useful question is not whether errors occur but what they cost and how quickly they are caught.

Work through the realistic failure cases:

• The agent gives incorrect information. Who notices, and how?
• The agent takes an action it should not have. Can it be reversed?
• The agent cannot help. Does the customer reach a person quickly, with the context intact?
• The agent is confidently wrong. What signal would reveal it?

The answers determine how much human oversight the agent needs. Low-cost, reversible errors can be reviewed after the fact. High-cost or irreversible ones need a person to approve before the agent acts. Deciding this in advance is far easier than deciding it after an incident.

The useful question is not whether an agent will make mistakes, but what a mistake costs and how fast you will find it.

5. How will we know it worked?

Choose one primary measure before launch, and record where it stands today. Without a baseline, every later conversation about value becomes a matter of opinion.

The measure should be a business result, not an activity count. "Conversations handled" tells you the agent is busy. "Enquiries resolved without a handoff", "time to resolution" or "cost per resolved case" tells you whether it is useful. Pair the primary measure with one quality check, such as customer satisfaction or error rate, so that efficiency is not bought at the expense of experience.

Then set a review date. At that point the owner looks at the numbers and makes a decision: extend, adjust or stop.

Taken together, these questions work best as a gate, not a checklist. Put the proposed agent in front of the business owner, the data lead and the risk lead, and ask each question in turn. Strong answers mean you are ready to commission design work. Weak answers tell you precisely what to fix first.

It is the conversation we like to have with leadership teams before any commitment is made, because it costs an afternoon and can save a year.