When is an AI agent actually worth the trouble?

An AI agent is software that uses AI to choose and carry out steps in a task. With the necessary connections and permissions, it can retrieve information from one system and use it in another. That might mean reading meeting notes and preparing a follow-up in your customer system.

The label says little about what a solution can do. What matters is the information it can access and the actions it is allowed to take.

Following up after a sales meeting

In this hypothetical example, a salesperson needs to find the customer’s history, write an email and record the next task after every meeting.

An agent could retrieve the relevant customer notes and draft a follow-up. It could also prepare a task with a suggested date in the customer system. The salesperson checks the content before the email is sent and the changes are saved.

If the meeting notes do not say what the customer agreed to, the agent must flag that as unresolved. A well-written guess is still a guess.

The task requires access to the relevant information and a connection to your systems. If those connections are missing, the cost of setting them up belongs in the budget.

Fixed rules may be enough

When a task follows the same rule every time, ordinary automation is often enough. In the sales example, that applies if the salesperson simply needs a follow-up task created whenever a quote is sent.

AI is more relevant when the content needs interpreting, for instance when different meeting notes need turning into a suggested next step. That interpretation can be wrong. So the solution must also be able to leave a task unresolved and ask a member of staff to take over.

Agree where the agent should stop

A narrowly defined first task makes it easier to see whether the solution helps. The agent gets access to just the information the task requires and is initially limited to preparing drafts.

The agreement must be specific. Can the agent only read customer data, or also change it? Who approves an email? What happens when information is missing or contradictory?

A named member of staff is responsible for following up on errors. The solution must show which information the agent used and which changes it made, so errors can be found and corrected.

Human approval can be a permanent arrangement. It does not need to be a temporary step towards greater independence.

Include the checking

A trial compares the current way of working with the proposed one. The tasks must reflect everyday work, including those with unclear notes or missing information.

The calculation includes time spent before and after, including review and corrections. Errors and the number of tasks a member of staff has to take over count too. So do the costs of setup, ongoing operation and maintenance as your systems or processes change.

If checking takes as long as the previous work, the task has not become easier yet. The next step is to narrow the task, change the solution or stop the trial.

An agent is worth trying where staff repeatedly gather information from several systems. If the trial shows an overall time saving, the follow-up is ready for approval when the salesperson comes out of the meeting.