Choosing the first process for an AI agent

The first agent sets the tone for every one that follows. If the first process goes well, the team gains confidence and the board approves the next one. If it goes badly, the project stalls, often for reasons that had little to do with the technology.

Six criteria

A good first process meets most of these conditions:

1. High, repetitive volume. Many similar cases arrive every week. If the process happens three times a month, the savings do not justify the effort of setting it up.
2. Clear rules. What can be done, what needs approval and what gets rejected is written down, or can be written in an afternoon with the people who run the process. What varies is how each case arrives: the message, the document, the question. That is where an agent helps; if nothing varies, a fixed rule may be enough.
3. Information that exists and is in order. The policies, forms and data the process needs exist and sit somewhere access can be granted. If the answer lives in one person’s head, there is nothing to work with yet.
4. Low risk if something goes wrong. A mistake can be caught and corrected without touching the customer’s money or the relationship with the supervisor.
5. An owner on the team. Someone who knows the process, answers for it and has time to review the first results once the agent is live.
6. A metric before and after. You can measure today, with real figures, how many cases arrive and how long they take. Without that baseline, there is no way to show the change.

Good first processes

In regulated companies, these processes tend to meet all six criteria:

• Certificates. Balance statements, contribution certificates or proof that an account is paid in full. They arrive every day, follow a template and draw on a system that already exists. The agent issues them and a person reviews only the cases with exceptions.
• Payment reminders. Notices before and after the due date, with the correct balance and the way to pay. The agent sends the reminder and answers questions, within the contact hours and channels the law allows. Payment arrangements outside the established terms go to a person.
• Frequent complaints. The reasons that repeat month after month (a charge the customer does not recognize, a certificate that never arrived, outdated details). The agent classifies the complaint, answers those with a known answer and passes the rest to the right team with the case documented.
• Document validation. Checking that a loan or onboarding application arrives complete: a legible ID, a current tax registration, proof of income. The agent checks and asks for what is missing; approval stays with the team. At Legalnova, a law firm, time spent on document review fell 70%.

Poor first processes

Other processes look attractive and are better left for later:

• Those that depend on case-by-case judgment. Unusual loans, special negotiations, legal opinions with no precedent. An agent can prepare the case, but the decision needs an experienced person, and the first project is the wrong moment to find out where that line sits.
• Those without data. If the information is spread across emails, personal spreadsheets and the team’s memory, put it in order first. The agent comes after.
• Those that move money without human approval. Disbursements, refunds, write-offs and changes to loan terms. The agent can review and prepare; a person approves. Starting here exposes the company exactly when the team is still learning to supervise the agent.

Measuring before and after

Before you start, measure the process as it runs today for a few weeks:

• How many cases arrive each week.
• How long each case takes, from request to answer.
• How many are sent back or corrected.
• How many team hours it consumes.

Once the agent is in production, measure the same things in the same way. Before starting, agree with the process owner on what result justifies extending the agent to a second process. That way the decision to continue rests on figures.

Checklist

Before approving the first process, confirm that:

• Enough similar cases arrive every week.
• The rules are written down, or can be written with the people who run the process.
• The information the agent needs exists, and access can be granted.
• A mistake can be caught and corrected without affecting the customer or the supervisor.
• Every operation involving money goes through a person’s approval.
• There is a named owner with time to review results.
• There is a measurement of the current process.

How Xplouse does it

In the diagnostic call, 30 minutes and free of charge, we review the candidate processes with you and choose the first one using these criteria. From diagnostic to production in 30 days. The agent does the work and your team approves the exceptions. Every step is recorded. See the operational layer.